In August 2000, Intel was the most valuable company in the world, with a market capitalization of roughly $509 billion. It made that money by selling the chips inside almost every PC in the world. “Intel Inside” was both a marketing sticker and a description of the entire industry’s architecture: one company’s processor, running billions of machines, invisible to end users but indispensable to everyone who built for them. Intel’s dominance seemed self-reinforcing, because the more machines ran on Intel, the more software was optimized for Intel, and the harder it became for customers to leave.

But, gradually and then suddenly, the biggest customers left anyway. ARM captured mobile; Intel cancelled its Atom mobile programme after spending an estimated $10 billion in OEM subsidies to capture roughly 1% of the market, formally walking away in April 2016. Apple announced its move to custom ARM chips in June 2020 and completed the transition when it discontinued its last Intel Mac in June 2023. Amazon, Google, and Microsoft have all built their own ARM-based server chips; by 2025, more than half of all new CPU capacity added to AWS runs on Amazon’s own Graviton silicon. Intel’s market capitalization is less than half today what it was at peak.

In April 2026, the most valuable company in the world is NVIDIA. It has the same kind of dominance, which is the same kind of trap.

Chart courtesy of Claude

NVIDIA is so valuable because its core product, the GPU—originally designed to render video-game graphics—is exceptionally well-suited to training large AI models. That’s why NVIDIA, in our moment, is unavoidable: not just for AI, but also for AV.

To function, AVs need a robust and constantly-updating model of the world around them, which they gather from continuous streams of sensor data: cameras, lidar, radar, and more. They then need to fuse that data into a world model. Based on that model, they must make ongoing decisions, sometimes at a fraction of a second, to ensure the vehicle operates safely and appropriately. And they must do this reliably over millions of miles.

This is, or should be, an AI problem, and NVIDIA makes the hardware the AI industry runs on. It therefore seems reasonable to expect that NVIDIA is just as relevant to the future growth and profits of AV as it is to AI. Jensen Huang, NVIDIA’s CEO, certainly wants us to think that: he has described an AV as a robot that needs a brain, and claimed that NVIDIA is the natural supplier of that brain. And he’s got announcements to back his claims up. NVIDIA has partnered with Mercedes-Benz, Volvo, BYD, Jaguar Land Rover, and other manufacturers, and has disclosed automotive revenue of $2.3 billion in fiscal 2026.

It’s tempting to think that NVIDIA will be to the self-driving car what Intel was to the personal computer: the provider of the infrastructure inside everything, the platform the whole industry runs on.

I’m not so sure.

The analogy holds insofar as NVIDIA today doesn’t build its own AVs but instead provides the infrastructure for those of other firms. Relying on NVIDIA makes sense for companies that don’t own their full technology stack, but as and I have written, owning your own stack is a prerequisite for success, and the biggest companies know it.

The companies who have had the most success in this space are those who believe that their competitive advantage requires tight hardware-software co-design, and accordingly have all built their own silicon. This means that, despite its prominence in the AI sector generally, NVIDIA’s position in driving automation is specific: indispensable to the industry’s middle, unsuitable for companies that need to own the whole stack.

Subscribe now

NVIDIA Was King of ADAS

The most important product in NVIDIA’s automotive business is a chip called DRIVE Orin.

DRIVE Orin is a ‘system-on-chip’, i.e., a single piece of silicon that combines a GPU, a dedicated AI inference accelerator, and an image signal processor, designed to run the trained neural networks that parse camera and sensor feeds in real time. Unlike a data-centre GPU, which is built to train models at scale, Orin is optimized for inference: fast, efficient, and cool enough to operate inside a car. Reliability is important here because automotive functional safety is governed by stringent standards, specifically ISO 26262, an international specification that classifies components on a scale from ASIL-A (lowest) to ASIL-D (highest). NVIDIA had DRIVE Orin certified to meet ASIL-D systematic requirements, the level required for systems that could cause serious injury if they fail.

Earning that certification is not straightforward. The process requires years of design documentation, formal hazard analysis, and a third-party audit. In the most detailed public case, the certifying body reported that the process took a supplier more than four years. The engineering overhead alone is estimated to add 30% to standard chip-development costs. General-purpose chipmakers see no reason to undertake this safety certification because none of their other markets demand it. But NVIDIA took it on and dug itself a meaningful moat in the process. A rival chip that matches Orin’s raw compute but lacks ASIL-D certification cannot be legally deployed in a production vehicle in most major markets. The certification is part of what NVIDIA is selling.

NVIDIA didn’t stop at the chip. It built ‘DRIVE Sim’, a simulation environment for training and testing autonomous systems before they touch a real road, and ‘DriveOS’, the software layer that manages how Orin’s hardware resources are allocated. It also published a reference architecture, called Hyperion, that showed Tier 1 automotive suppliers exactly how to build a production system around Orin. The effect was to reduce the work an automaker had to do itself. If an automobile is indeed a robot that needs a brain, NVIDIA sells not only the brain, but the robot’s DNA, the instruction manual for the whole body.

The standard measure of chip throughput is ‘tera operations per second’, or TOPS; i.e., how many operations the chip can perform per second, measured in trillions. At launch, Orin boasted 254 TOPS, easily outpacing its closest competitors. Mobileye, the Israeli chip designer, offered its EyeQ5, which delivered only around 17.6 TOPS, while San-Diego-based Qualcomm’s Snapdragon Ride product, targeting the same applications, had a mid-range SKU at approximately 100 TOPS.1

NVIDIA’s investment paid off. Between roughly 2022 and 2025, DRIVE Orin became the dominant AI chip for advanced driver assistance (ADAS) across the Chinese electric vehicle market. More than ten Chinese OEMs shipped consumer vehicles running Orin during this period. BYD, the world’s largest EV manufacturer by volume, equipped more than one million vehicles with its NVIDIA-powered “God’s Eye” driver-assistance system by mid-2025. Other Chinese automakers like NIO, XPeng, Li Auto, Zeekr, and Xiaomi were scaling from basic cruise control to ADAS fast at this time, and NVIDIA had what they needed: a production-ready chip with the right performance, the right safety credentials, and a complete development package that allowed them to make the transition right away rather than build their own chips, or simulation environments, or vehicle-chip integration systems.

NVIDIA had competition in the Chinese market from Mobileye and Qualcomm, but not serious competition: only NVIDIA could deliver so much quality at such a scale, which meant it could charge a significant premium. NIO CEO William Li stated on the firm’s Q3 2025 earnings call that switching to its own chip delivered “approximately 10,000 yuan ($1,420) in cost optimization per vehicle,” a figure that illustrates how much profit NVIDIA gets from Orin.

Undoubtedly, some of NVIDIA’s customers would prefer to take their business elsewhere, but doing so is difficult. Finding a new supplier would mean re-earning safety certifications from scratch, establishing new sets of relationships with parts manufacturers, replacing software toolchains, and retraining engineering teams. Bringing chip-making in-house incurs all these difficulties and more.

But it might be worse than the alternative. NIO VP Wang Qiyan put the underlying logic plainly: “Leaving our fate in the hands of suppliers is terrifying.” That’s why, despite the cost and challenge, NIO ultimately switched to an in-house chip, a move that required over 600 engineers and more than 140 million dollars, taking four years from project start to deployment (2021 to March 2025).

Whether the implications of NIO’s switch were good or bad for NVIDIA depends on your perspective. It shows that reliance on NVIDIA’s product (what an economist might call its installed base) has become a moat, and that crossing it to make one’s own chips requires immense time, expense, and inconvenience. But it also shows that if the economic case for leaving becomes strong enough, NVIDIA’s customers that can do so will power through to switch, as NIO did.

The Frontier Built Its Own Silicon

Some notable names are absent from NVIDIA’s customer list. Waymo isn’t there, nor are Tesla and Zoox; none of the companies most associated with the frontier of genuine driverless capability. That’s because none of these companies adopted NVIDIA’s automotive platform— not Orin, not DriveOS, not the Hyperion reference architecture—even as a transitional step.

To understand why, it helps to look at how the most capable AV systems are built.

In self-driving cars built this decade, what separates the frontier companies is not a better chip, but rather their co-design loop, in which sensors, compute, and software are developed together, each constraining and enabling the others. The chip shapes the models; the models shape the chip; the sensors constrain both. Designing all three together, each informing the others, learning quickly how to improve and deploying those improvements is the key to success in this domain. Co-design permits a faster, more power-efficient, and ultimately cheaper unit than anything assembled from off-the-shelf parts could be.

This is why Waymo builds everything from scratch.2 Waymo operates more than 3,000 robotaxis across ten American cities, completes 500,000 paid rides every week, and has logged more than 200 million fully-automated miles. Its sixth-generation system, which began fully autonomous operations in February 2026, uses custom chips designed entirely in-house. Waymo didn’t adopt NVIDIA’s automotive platform at any stage of its development. From the start (as far as outsiders can tell), the company concluded that the hardware-software design loop was too important to hand off to an outside supplier.

The same pattern holds across every company that has reached genuine scale in full autonomy. Tesla has designed its own AI chips since 2019, and before its robotaxi programme shut down, Cruise had a 750-person hardware team and began developing four custom chips. Its head of hardware, Carl Jenkins, told Reuters that NVIDIA’s pricing was simply unsustainable: “There is no negotiation because we’re tiny volume”.3^[3]

There is one exception to the case, namely Aurora, the first company to operate a commercial driverless Class 8 trucking service. Aurora launched that service in April 2025. It operates on U.S. highways and had logged more than 250,000 driverless miles by January 2026. And it uses NVIDIA chips.

Aurora’s reliance on NVIDIA may seem to suggest that NVIDIA will be a major player in frontier AV after all, but the conclusion doesn’t follow. Waymo’s system is the product of a tightly-integrated sensor suite—13 cameras, four lidar units, six radar units, all custom-designed and built in California—combined with a custom compute stack and proprietary software developed together from the start; its performance depends on that integration at every layer, which is why controlling the hardware matters. Tesla and Zoox are going even further, building their own vehicles and integrating their sensor suites into them. Aurora’s strategy is the opposite: it aims to build a strong standalone driving-automation suite, the Aurora Driver, and deploy it across OEM platforms it doesn’t control and cannot customize. This means that for Aurora, NVIDIA’s off-the-shelf, safety-certified hardware is the cheapest and best solution to their needs.

That structural logic makes Aurora a different kind of test case: not whether NVIDIA’s platform survives frontier defection—Aurora does not, at present, have motive to defect—but whether NVIDIA can support a full-autonomy deployment at production scale without triggering the cost pressures that will drive Aurora toward custom silicon, as it did to NIO and Cruise. The 2027 deployment will begin to answer that.

At this point, one may wonder whether this is the AV equivalent of Apple vs. Android: a premium integrated tier and a larger “good enough” platform, both of which can coexist indefinitely. I think it’s unlikely to be the case. The mobile-phone sector sustained that bifurcation because each offered its own standardized platform on which others could sell software. In driving automation, the competitive advantage is the hardware-software stack itself. There is no equivalent of a ‘better app’ that runs on general-purpose compute and outperforms tight co-design.

So NVIDIA doesn’t have Waymo nor Tesla, but they do have Aurora, and more besides. It’s tempting to think that NVIDIA doesn’t need Waymo or Tesla, nor any American self-driving firm, given that NVIDIA’s strength in driving automation is in supplying the Chinese ADAS market, which is a strong foundation.

But that foundation is eroding.

Artist’s conception.

As I mentioned, NIO has migrated its main vehicle lineup to an in-house chip. Meanwhile, XPeng’s in-house Turing chip claims roughly three times the processing capacity of a single Orin unit. Horizon Robotics, a Chinese startup competing directly with NVIDIA in the automotive AI chip market, now commands approximately 50% of the Chinese domestic driver-assistance chip market by volume. The same pressures that push full-autonomy companies toward custom hardware—managing costs, controlling their own supply chains, and the painful inability to specify power consumption or performance—are now pushing on the Chinese manufacturers that have gotten big enough to go it alone, as NIO did. (It certainly doesn’t help that in September 2025, the Chinese government issued an advisory explicitly discouraging Chinese OEMs from purchasing NVIDIA chips, but that directive seems less like a sudden break and more like pushing on an open door.)

Outside China, NVIDIA’s position looks intact for now. Aurora has committed to the company’s next-generation chip, DRIVE Thor, which delivers roughly four times the processing capacity of a single Orin unit, with a 2027 deployment target. DRIVE Thor has more than 15 announced adopters across passenger vehicles, trucks, and robotaxis, and its product timeline appears to be proceeding normally. So NVIDIA seems safe, for so long as no Western OEM reaches the volume threshold that would make going in-house compelling.

Intel’s Present Is NVIDIA’s Future

NVIDIA is a supplier of silicon infrastructure to companies that don’t need to own the whole stack. Most of the companies building driver-assistance systems today are not Waymo or Tesla; they lack the engineering depth, the production volume, and the competitive logic that would make custom chips worth building. For them, NVIDIA’s complete package is the rational choice. And NVIDIA’s contribution is real: it enabled an entire category of companies to deploy sophisticated driver-assistance systems faster than they could have independently. That enables safer driving technology to reach ordinary drivers more quickly, and it deserves credit for it.

What NVIDIA hasn’t managed, and what the evidence suggests it structurally can’t, is to become the foundational platform for the companies at the frontier of full autonomy. The firms defining what genuinely driverless vehicles will look like have all concluded that owning their own compute is non-negotiable.

Aurora’s planned 2027 deployment on DRIVE Thor is the clearest near-term test of whether that conclusion has exceptions: perhaps Aurora will deliver commercial Level 4 freight autonomy at production scale on NVIDIA hardware and report that the cost and capabilities hold up as volumes grow. If we saw that, and/or a major Western OEM reversing course on a planned in-house chip programme and committing to Thor at volume (absent geopolitical pressure), I’d update my views. The first outcome would suggest that NVIDIA’s ecosystem can underpin full autonomy rather than just ADAS; the second would suggest the ecosystem moat is deeper than current evidence suggests.

But absent those developments, it’s fair to say that NVIDIA built the tools that let most of the industry move faster than it could have on its own. However, its fastest-growing customers are leaving it behind, and the firms that are furthest ahead never used it in the first place.

Share

Intel’s story didn’t end with its displacement. It remains a large and profitable business, selling chips into markets that don’t require custom silicon. NVIDIA’s story in driving automation may follow a similar arc: indispensable to the many, bypassed by the few who get large enough to need something built precisely for them. The parallel is to Intel-in-2018, not Intel-in-2024: the frontier defections have started, but the displacement is not yet irreversible.4 Whether NVIDIA escapes the trap is the question the next hardware generation will begin to answer.

The same dynamic that overtook Intel—frontier defection, followed by volume compounding among customers that reached the scale threshold—is already visible in NVIDIA’s automotive business. Whether it plays out over five years or fifteen depends on how quickly the Western ADAS market matures, but I think the direction is not in question.

Respect to , and for comments on earlier drafts.

For the first time ever, we couldn’t get a seat at the bar. The Long Branch Social House was packed, as full as we’d ever seen it. Instead of sitting in our usual place, we found an available table (one of only two) and settled in.

Our server T came over, and we asked him, in all innocence, what was happening tonight that had brought in a crowd.

He gave a forced smile. “You haven’t heard? We’re closing tomorrow. This is our last night in business.”1

I moved to Long Branch in southwestern Toronto four years ago and began exploring the neighbourhood. Back then, a different bar operated in that space, which I found unimpressive. Nothing was wrong with it; the food, the service, the atmosphere were all fair but not good. After my first visit I left with no intention of going back. When I found out later that it had closed, I was unsurprised.

Later my girlfriend, now my wife, moved in with me. She uses Instagram, where she follows local businesses, and learned that the site had new management. It was now operating as the Long Branch Social House, and she thought we should try it out. I don’t remember when exactly that was, but I suppose it was about two-and-a-half years ago, shortly after it opened. Coloured by my previous experience, I was skeptical, but willing to give it a try.

It’s funny, now, to realize that I don’t remember that first visit of ours, given how important it was. But that’s because all our visits ran together; for the last two and a half years, my wife and I went there at least once, and sometimes twice, pretty much every week.

The sociologist Ray Oldenburg spent most of his career thinking about places like the Social House. In The Great Good Place, published in 1989, he coined the term ‘third place’, meaning the informal gathering spaces that exist outside home and work: the café, the barbershop, the pub. Home is the first place. Work is the second. The third is where you go when you don’t want to be in the first two.

Oldenburg identified several qualities that define a genuine third place. It is neutral ground, with no one playing host. It’s socially levelled, with no hierarchies. It has a core of regulars who set the tone, who by example teach newcomers how to behave in the space, and who help those newcomers feel included. And it provides what Oldenburg called “a home away from home”: the same feelings of warmth and belonging, but without the obligations.

My wife has had locals before, and has lost them. She has learned to hold them loosely, knowing they may not last. I had no such experience: this was my first local, and my first time losing one.

And so I wasn’t prepared for how much it stung.

Subscribe now

My wife and I sat, and took in the crowd, and nursed our drinks. They were our usual order: T hadn’t needed to ask, only to confirm (a glass of Sauvignon Blanc for her, an Old Fashioned for me). And, to my own surprise, I grieved, just as I have when a loved one has died. I asked myself, what am I grieving? What have I lost here? I thought about third places. Oldenburg wrote that “nothing contributes to a sense of belonging in a community as much as membership in a third place”. And that was what it was. We had lost a small part of who we were.

C owned and ran the Social House with M. We saw her more often than him, as they also ran another establishment nearby where he spent the bulk of his time. T, a jolly hockey dad, and D with her dazzling smile, were the people we saw most often behind the bar. K was occasionally working a shift when were there, but we saw her more often as a fellow patron; she liked the place enough to spend time there off-shift. We saw less of A and the other wait staff, as they managed the dining area and we preferred the bar, but we always shared smiles and waves, and often a brief chat.

The principal attraction was the food, all of which was made in-house and was consistently great, much better than typical pub fare. When I was hungry, I often ordered the burger, with its side, a mountain of crisp-but-not-oily French fries. When I just wanted a bite, the fish tacos were consistently great. And I always checked the seasonal menu… its greatest hits included the shrimp scampi and the Scotch egg. I pleaded for those items to make the permanent menu, and was assured that they would come back (famous last words).

There was often live music, which ranged from ‘merely pleasant’ to ‘quite good’. Many of the patrons there on the last night were musicians who had graced the Social House’s stage, and had more to mourn than I did: the loss of a venue, of which there are fewer all the time, where they could get paid to perform. When the music wasn’t live, the sound system seemed to play only the greatest pop hits of the 1980s. I was aware that I was being pandered to, that this playlist—Prince, Madonna, Springsteen, Michael Jackson—had been calibrated for people of my age and demographic. I didn’t care. When you’re home, people cater to you, and the playlist meant I was where I belonged.

It was a working-class place; the patrons tended to jeans, baseball caps, and sports jerseys. There were multiple TVs. Some showed pub trivia, others showed compilations of equal parts ‘sinking a basket, at midcourt, while looking away from the net’ athletic feats, and ‘guy in the gym hits himself in the crotch with a free weight’ follies. The TV in the dining room wasn’t visible from where we usually sat but always seemed to be showing Gordon Ramsey’s Hell’s Kitchen on loop. No matter where one sat, there was always a visible TV showing a sports match.

But the Social House also had a huge disco ball in the open area in front of the stage. It hosted drag shows on Sunday afternoons in the summer; we attended a few, and had hoped to attend more. I have almost no interest in sports, but enjoy the company of people who have great interest and let it show. The place was a big tent, and contained multitudes. It never felt like a caricature of itself, like Cheers; or maybe it’s more accurate to say that, while I was there, I never felt like Frasier Crane.

Our visits began to run together. The Social House stopped being somewhere we decided to go and became somewhere we simply went, the way we go down to the lake on Saturday mornings, without discussing it. That’s how a good local works; you don’t notice it. It becomes part of the background, of the texture of your life.

On those Saturday visits to the lake shore, we would occasionally see C from a distance, exercising his dogs. He lives in the neighbourhood, as we do; the Social House was operated by local people for the benefit of same. I keep returning to the fact that some of the staff liked to socialize there when not on shift; it wasn’t just part of the community, it was a community.

I don't know, but I suspect, that the Social House closed because its landlord raised the rent, with an increase steep enough that running the place no longer pencilled out.

In Canada, there are no limits on commercial rent increases. A landlord may raise rent by any amount at lease renewal, with no obligation of reasonableness and no recourse for the tenant. Some urbanists think a truly open market like this one promotes better outcomes in the long run. Others think that commercial tenants, like residential ones, should be entitled to limits on rent increases, in order to permit long-term planning and encourage stability of occupancy. I haven’t studied the matter in sufficient depth to hold a position. But I do shake my head at the short-sightedness of a landlord who prefers an ultimatum to a negotiation.

I suppose that the landlord made a take-it-or-leave-it offer, and I suppose that the Social House chose to leave it. That means that the staff have lost their jobs, the musicians have lost their bookings, the party-goers (and there were a lot of them, retirement parties and 50^th birthdays and fundraisers) have lost their venue.

But the landlord has lost the rent. The building will now, in all likelihood, sit empty through the summer, which is precisely the season when a patio operation in Toronto generates revenue. It’s baffling to me that the owner would rather kill an existing business that was doing well, and better all the time, than take the chance that a new tenant, saddled with even higher rent, will take a gamble on setting something up there.

I don’t know the details. The landlord, whoever they may be, may be in the perverse situation that describes, where their financial position is better off with a high rent on paper, even if no one’s paying it. Or perhaps they are simply looking for an exit. Lakeshore Boulevard is under sustained development pressure: a few steps away from the Social House, just down the block, a retail strip is about to be replaced by a 43-storey condo tower. Perhaps such a redevelopment is in the Social House’s future; it does have an unusually deep lot, both for the area and for its particular block.

If that happened, it might be a good thing. The Social House is walking distance from a transit hub, home to both the Long Branch commuter-rail station and the terminal stop of the Lake Shore streetcar. I know that because on some evenings my wife would take transit from her office and walk over to the Social House to meet me there. A transit-rich area falling within Toronto’s boundaries, like this one, should be denser, and it seems that the market believes that too.

I agree with that we urbanists should honour the past we have lost without ceasing to build the future. But today I’m more in the mood for the former.

Robin Dunbar is an Oxford anthropologist who has studied what British pubs do for the people who spend time in them. He has found that people with a local pub have larger social networks, feel more socially engaged, and are more likely to trust their neighbours. Those without one have significantly smaller networks. That matters, because our social networks “provide us with the single most important buffer against mental and physical illness”.

A local pub is part of that buffer, which is why it is concerning that we’re losing so many of them.

Three minutes’ walk away from the Social House, just one block over, there used to be a beer-and-chicken-wings place, Sloppy Joe’s. It closed in March 2025 after 54 years of operation by two generations of the same family. The operators’ final statement read “Our father opened this business in April of 1971, driven by the dream of creating a neighbourhood gathering place that felt like home.” Sloppy Joe’s is one of the more than 5,000 bars and nightclubs that Canada has lost in the past twenty-five years.

In the year 2000, the country had almost 9,000 such establishments. Today there are fewer than 4,000.

Share

There are many causes for the change, and they stack. The pandemic ban on going out was more than many places could handle. Since it ended, we’ve had rampant cost inflation. Meanwhile, social drinking has been on the decline for decades, while the Internet has been on the rise. I mention the Internet because it’s the most obvious contributor to the broad and ongoing social trend to stay in rather than go out. A guest writer on Derek Thompson’s Substack described this recently as “toxic solitude”. In Canada, a 2024 survey found that 60% of Canadians feel disconnected from their community, and 36% report having no third place at all.

We don’t go out to third places, so they struggle and close, so it’s harder to find a third place, so we don’t go out, and the vicious cycle turns.

Friday night, we settled our bill. We said our goodbyes. The crowd was thinning now; the hard core would remain for a few more hours, but others had already begun to leave, and it was our turn. We walked out and went home.

Love songs sound saccharine, until you fall in love. Breakup songs seem overwrought, until you’ve been dumped. And if you’ve never had a local, this line seems trite: that sometimes you want to go where everybody knows your name.

But it turns out to be true.

So does this:

Every new beginning comes from some other beginning’s end.

Bets are a tax on bullshit.

I use that term deliberately, even though it’s coarse, because it has a specific meaning, one that nonsense or lies or malarkey doesn’t convey. As per the philosopher Harry Frankfurt, bullshit has a precise definition: it is speech produced without regard to its truth or falsity. It doesn’t intend to track reality, but to persuade, impress, or signal some quality in the speaker. Whereas liars aim to conceal the truth, bullshitters are indifferent to it. They are playing a different game.

Bullshit is everywhere, and it makes it hard to understand the world. This is a problem, and a bet is one solution to it, because a bet clearly aligns speech with belief. Someone might say that he’s sure Canada will win a gold medal in shot put at the next Olympics. Does this person really believe that? He might be bullshitting, saying words that don’t correspond to anything about what will happen at the Olympics, but instead correspond to how much he loves the country… or wants us to think he loves the country. Ask him to wager $100 on whether Canada wins a gold medal at shot put, and his incentives change: now, his claim must rest on what he actually thinks will happen, not what he wants us to think about him. If he declines the bet, you have learned something about what he really thinks.

Prediction markets are an attempt to scale up this insight.

Polymarket hosts well over a hundred open markets on innovative mobility, and their prices are telling. In the Tesla-adjacent markets—which dominate the platform, in our domain at least—a large portion of the crowd is systematically pricing technology risk when it should be pricing regulatory risk, and the result is a consistent overestimate of near-term Tesla delivery.

How Prediction Markets Work

As per Scott Alexander, who made the case at length in his Prediction Market FAQ, a price on a prediction market represents the collective judgment of everyone willing to put money behind their beliefs. If the market underestimates the probability of something, anyone with better information can profit by betting Yes, and in doing so, move the price toward accuracy. Any persistent mispricing is a standing invitation for someone who knows better to make money.

This self-correcting mechanism is what makes prediction markets, at their best, epistemically useful: they reward people for being right, and penalize them for using speech as performance.1 In that spirit, here is what I think the markets are getting wrong, and why.

Before I tell you that, though, let me say a few things as clearly and strongly as I can:

● I cannot trade on Polymarket; Ontario law prohibits me from doing so

● Even if it didn’t, I wouldn’t; I don’t hold mobility-related equities, because I don’t want my analytical work to be influenced by my portfolio

● There is no financial upside for me in the following analysis

● NONE OF THIS IS ADVICE TO YOU ON HOW TO GAMBLE, and if you do, I disclaim all responsibility

You may think that my own lack of skin in the game gives you good reason to discount everything I’m about to say. The only response I can make is that I do care about the truth for its own sake, and that perspective informs all my work at Changing Lanes.

Another thing: the markets never stop moving, but text is frozen in time. My analysis reflects the analysis at time of drafting this piece, which accounts for any difference in the price I discuss and the market you see.

Five Open Markets

Will Tesla launch driverless robotaxis in California by June 30? (17% chance)

Read more

Firstly, I’d like to direct your attention to two op/ed pieces of mine published last week, both having to do with the tragic crash of an Air Canada flight at LaGuardia.

• In the New York Post, I argue that the hidden contributing factor to the crash was lack of airport capacity in North America, and the corresponding imperative to start building airports again

• In the Globe and Mail, I argue that while it’s deplorable that the CEO of Air Canada spoke only two words of French in his message of condolence, it’s indefensible that only one air carrier is expected to abide by Canada’s Official Languages Act

I hope you find both to be of interest.

Secondly, this issue of Changing Lanes is co-written with , who writes the newsletter (which you should certainly subscribe to!). Jannik covers European innovation policy and technology while I cover innovative mobility. We divided the work accordingly; I took the driving-automation material, he took the broader German industrial story. We’ve wanted to work together on something for a while, and this felt like the right subject.

In September 2025, Mercedes-Benz CEO Ola Källenius said on a podcast that his company was “at the front” of automated driving. Listeners would be forgiven for finding the claim persuasive. After all, as he pointed out, Mercedes had 30 years of experience in developing and deploying driver-assistance systems. Not only that, Mercedes was the first manufacturer to certify an advanced driver-assistance system capable of handling all driving tasks under very specific conditions (what the SAE refers to as a Level 3 automated-driving system) in Germany. It was the first non-Chinese OEM to obtain an automated-driving permit in China as well. “Every new Mercedes has the technology and the onboard computer to drive autonomously,” Källenius said.

Four months later, in January 2026, Mercedes killed the system.

Mercedes confirmed that month that Drive Pilot would be dropped from the facelifted S-Class and EQS, the avowed reason being that Luminar, the LiDAR supplier Mercedes had depended on, had failed to meet its contractual requirements (Luminar subsequently filed for bankruptcy in December 2025). A few weeks later, BMW followed by scrapping its own “Personal Pilot” from the 7 Series facelift.

With no German manufacturer publicly selling a car with a Level 3 automated system (as per the SAE), German AVs seem, for all practical purposes, dead.

The choices that Mercedes has made are not unique, but part of a pattern, one that extends well beyond one model line or one company. Germany built the first automated cars. It holds every conceivable advantage, including engineering talent, manufacturing scale, regulatory influence, and established brands; and yet it is poised to license the automotive future from companies in California and Beijing, leaving the 716,000 workers in Germany’s automotive sector at the mercy of decisions taken elsewhere.

The reason for this surprising forfeit is that German OEMs lack the capability to train and scale an AI system that perceives an unpredictable environment and makes real-time driving decisions. This stack improves through more data, more compute, and more iteration cycles. German OEMs are increasingly absent from this stack as original developers.

Subscribe now

Understanding why this is happening requires looking not at what Germany failed to do, but at what kind of industrial economy it has, and the different kind that the twenty-first century requires.

A Huge Lead, Squandered

Germany beat everyone else to driving automation. As long ago as 1986, Ernst Dickmanns at the Bundeswehr University Munich began outfitting Mercedes vehicles with cameras and custom processors to teach them to read the road. By 1994, his twin automated sedans were driving themselves on European highways at 130 km/h in normal traffic. A year later, a modified S-Class completed a Munich-to-Denmark run at speeds reaching 175 km/h, with the system in control for 95% of the journey. The Eureka Prometheus program that funded this work included Daimler-Benz and every other major European automaker.

Germany presided over the birth of automated driving, but left rearing the child to others. There was no commercial pull, no institutional equivalent of DARPA’s Grand Challenges to force the technology out of the lab and onto the road, and Europe’s capability dispersed.

Källenius blamed regulation, not capability, for the slow rollout: “If you develop a product you’re not allowed to sell, no market emerges.” But that can’t be the whole story. China and the USA permitted automated fleets years ago. Waymo, Tesla, Baidu Apollo Go, and WeRide already offer Level-4 automated rides in cities like Beijing and San Francisco. According to Källenius himself, Mercedes was the first non-Chinese manufacturer to get a permit operating AVs in China.

The reason Mercedes and other German manufacturers fail in yet another transformative technology is not regulation.

German manufacturers remain world-class at what one might call the ‘carrier layer’ of automated vehicles: platform construction, crash safety, powertrain integration, supplier management, regulatory certification, and manufacturing quality at scale. Mercedes was among the first manufacturers to sell a car that could parallel-park itself without the driver touching the steering wheel, a feature that struck buyers in the mid-2000s as either impossible or magic; its remote-parking feature is equally impressive. Mercedes’s automated valet parking system, approved commercially in a Stuttgart parking garage in 2022, demonstrated genuine Level-4 capability in a structured environment. Drive Pilot itself is a legitimate engineering achievement: 35+ sensors including LiDAR, redundant steering and braking, centimetre-accurate positioning, and the world’s first series-production Level 3 certification. Germany knows how to build safe, well-integrated vehicles and how to navigate the most demanding regulatory regimes on earth.

But as strong as German OEM capability is, their weakness, namely their inability to build an AI sophisticated enough to automate the driving task, more than offsets it.

That weakness may not be obvious at first glance, given the course Mercedes is taking. The firm is pursuing a bifurcated approach. In the US and Europe, its new MB.Drive Assist Pro system runs on Nvidia’s computing architecture and AI software stack—a Level 2++ system that works in city traffic, on highways, and in parking scenarios. In China, Mercedes has partnered with Momenta, a Chinese AI startup, for the same capability. BMW and Audi have made the same choice: BMW has also partnered with Momenta for China-tailored driver assistance on its Neue Klasse vehicles, and with Qualcomm for broader markets. Audi uses Momenta for its new EV line developed with SAIC, and Huawei’s Qiankun system for its legacy gasoline models.

These partnerships are impressive, but point to a common frailty. All three German luxury brands have licensed core driving intelligence from the same Chinese startup for the same market for the same reason: they cannot match what domestic Chinese competitors are shipping.

The trainable driving stack—the system that perceives, predicts, and plans—is the component most likely to generate compounding returns through data accumulation and algorithmic improvement. One caveat worth noting is that robotaxi economics have not yet been proven at scale, and German OEMs still earn substantial margins selling vehicles. But the evidence that we have is suggestive: Waymo, Baidu Apollo Go, and Tesla are each building fleets or datasets that improve their systems with every kilometre driven. The value of the driving stack compounds, but the value of the vehicle platform does not. A manufacturer that licenses the former while owning only the latter is positioning itself as a supplier to whoever controls the learning system. And not only do suppliers earn lower margins than platform owners, they have no leverage: Nvidia or Momenta can put their driving stack into a different manufacturer’s cars more easily than Mercedes can build their own driving stack. The dependency runs one way.

By licensing the autonomy stack rather than building it themselves, German OEMs are surrendering the future.

An Engineering Economy, Not a Platform Economy

The reason German OEMs do not build their own full-stack autonomy systems lies less in technical incompetence than in a structural mismatch between two types of economy. Engineering economies reward precision, pre-validation, exhaustive testing, and incremental improvement. Platform economies reward aggressive data collection, rapid iteration, high risk tolerance, and network effects. The German industrial system is optimized for the first type, and produces excellent component suppliers and integrators.

It does not produce AI-native platform developers.

Volkswagen’s experience makes the structural point most bluntly. Its in-house software unit, CARIAD, spent years attempting to build an integrated software stack and largely failed, prompting what the Financial Times described as a complete reset, a pivot from indigenous development to integrating partner technology. VW had also co-funded Argo AI, a full-stack autonomy venture with Ford; when Argo shut down in 2022, VW’s response was to partner more, not invest more.

The pattern across German OEMs is consistent: when internal development falls behind, the response is to lease capability rather than build it.

The contrast with American technology companies is instructive (let us stipulate that Tesla is one of these; it has much more in common with them than with Ford or GM). General Motors shut down Cruise, and Ford killed Argo AI; one might cite either retreat as evidence that full-stack autonomy is difficult, and no one has an edge. But those were OEM-backed moonshots. Waymo continues to expand its commercial driverless service in multiple US cities, while Tesla’s effort to build a truly Unsupervised Full Self-Driving system accumulates training data from millions of vehicles on the road every day. It’s true that the American OEMs retreated, but American technology companies did not. Germany, not having the latter, needs to rely on the former.

None of this would matter much if automated driving remained a niche feature for early adopters, but the trajectory of automotive technology suggests it won’t. Power steering, anti-lock brakes, and power windows each arrived first as a luxury add-on and became a baseline expectation within a generation; backup cameras and Bluetooth links to a driver’s mobile phone are following that path now. We have every reason to expect that driving automation will do the same. The consumer paying six figures for a Mercedes in 2030 will expect the car to drive itself, at least in some circumstances; the one paying low-five figures for a compact will expect it by 2040. At that point, every car Mercedes sells will carry a software licensing cost to Nvidia or Momenta; a rent extracted from every unit, in perpetuity, by whoever owns the learning stack.

German manufacturers, among all players, should know this the best, and should be fighting the hardest to avoid losing the next battle over a market-defining technology. They have found no answer to Chinese dominance in electric vehicles. Instead, they are lobbying to delay the combustion phase-out, advocating for a slower transition. They are fighting to preserve a capability that their customers will abandon, while ceding to their Chinese competitors the capability everyone will want. Given those circumstances, the fixation on regulation is understandable: German OEMs can’t build competitive EVs, and they can’t make consumers not want to buy EVs, but they can push back against EU regulations that promote EVs. Faced with a difficult problem, they are pulling on the only lever they can reach, and that move is shaping their responses to other problems.

That over-reliance on what they know best is hampering them in other ways beyond technology development. American and Chinese automated-driving leaders are building business models around fleet operations: Waymo runs robotaxis, Baidu’s Apollo Go has reached per-vehicle profitability in Wuhan, and Chinese firms are expanding into the Middle East and Southeast Asia. The logic is to accumulate driving data, improve the system, and monetize through transportation-as-a-service. Meanwhile, German OEMs remain focused on selling vehicles and earning margins per unit.

One might reply that specialization is rational. If Nvidia and Momenta are better at automated-driving AI, why shouldn’t Mercedes license driving ‘software’ from them, while focusing on what they and their workers are best at, namely building the ‘hardware’? Carmakers haven’t been completely vertically integrated since the early days of the industry, and insisting they should become so integrated now seems foolish.

The rejoinder is that automakers have indeed always outsourced certain elements of their vehicles… but never the basis of their brand value. Mercedes commands a price premium because its brand makes a claim about engineering: these cars are built to exacting tolerances by engineers who know how to build them better than anyone else. That claim will remain true of the physical platform, but it will not be true of the software that increasingly defines what the car does. A Mercedes whose driving intelligence comes from Santa Clara and Beijing is not making the same brand claim as the one that justified its price premium in the past. Certainly, the marginal improvements that will make a 2035 Mercedes more desirable than its rivals—meaning the ones consumers will actually pay for—are almost certainly software improvements, not mechanical ones.

When the dominant monetization model shifts from unit sales to fleet-based data services, being an excellent vehicle manufacturer positions you first as a supplier to the platform, not the platform owner. Ultimately, an automaker’s brand value rests on the claim that ‘we know how to build these things and we can iterate the next generation.’

Increasingly, Germany can only assert the first part.

More Than an Automotive Problem

Källenius’s September rhetoric pointed to regulation as the binding constraint: “What we need in Europe is a considered, thoughtful discussion to bring more of this technology to the road.” Germany’s 2021 Autonomous Driving Act does impose strict requirements: Level 4/5 vehicles are permitted only in approved operational areas, require a technical supervisor, and must undergo extensive certification. The US system, by contrast, allows relatively easy testing permits across multiple states. China has issued over 16,000 test licences and opened 32,000 kilometres of roads for automated-vehicle testing.

But the regulatory argument has a fatal flaw: Mercedes itself chose to retreat from Level 3 not because regulation prevented it, but because the commercial case didn’t work. And the system it is replacing Drive Pilot with—MB.Drive Assist Pro—will launch first in China, then in the US, and only later in Europe. If European regulation were the primary bottleneck, one would expect firms to be pushing hardest for regulatory reform. Instead, the entire European public discourse revolves around stopping the EU’s combustion-engine phaseout mandated after 2035.

Germany has no equivalent of Waymo, no full-stack automated-driving developer, and no AI platform champion in any domain. The German Mittelstand produces excellent components, middleware, and specialized tools (sensor companies, simulation providers, remote-driving solutions).1 It’s telling that Momenta has opened a research centre in Stuttgart. The Chinese firm is recruiting German engineering talent to work in Germany, but on Chinese autonomy systems. The Fraunhofer system produces world-class applied research, and German universities train outstanding engineers, yet the value chain consistently terminates before it reaches the platform layer.

And it isn’t just automotive that’s struggling. For example, Germany played a major role in developing the software foundations of modern robotics, particularly through research institutions such as the German Research Centre for Artificial Intelligence and the Fraunhofer Society. Since the early 2000s, these institutes have pioneered robot perception, machine vision, sensor fusion, and human-robot collaboration. Their work enabled robots to interpret visual data, map environments, and safely interact with humans, so today’s warehouses can be automated, factories inspected, and robots do logistics and manufacturing for us. Only, not really in Germany.

While Germany advanced much of the underlying research, companies like NVIDIA built the core AI and simulation platforms used to train and deploy robots, while firms like Boston Dynamics and Amazon developed large-scale robotics systems and deployed them in real-world settings. Germany remains strong in industrial robotics hardware and automation engineering, but the scalable software stacks, AI ecosystems, and platform economics that now shape the robotics industry have largely been driven by US and Chinese technology companies.

A similar platform, pioneered in Germany and later lost to its industry, connected manufacturing to cloud services. The promise of ‘Industrie 4.0’ was that connecting factory machines to cloud platforms would allow manufacturers to pool operational data across entire production networks, use AI to predict equipment failures before they happen, optimize processes in real time, and continuously improve output. Whoever controlled that data layer would shape how the world’s factories run. German incumbents had every conceivable advantage in smart manufacturing: the world’s best machine tool builders, deep automation expertise, and anchor firms such as Siemens and Bosch. They blew it.

Siemens built MindSphere as a platform play, but refused to commoditize its own hardware. MindSphere was always biased toward Siemens equipment, which is exactly what prevented it from becoming the horizontal standard. AWS, Azure, and Google Cloud had no such conflict; they don’t sell factory machines, so they could be genuinely vendor-agnostic, which is what manufacturers actually wanted.

The broader German response, Gaia-X, was supposed to create a European cloud substrate; instead, it produced what one co-founder called “a crushing failure, a colossal waste of time” and what the CEO of Nextcloud described as “a paper monster” of standards documents with no working product. Today, MindSphere (quietly rebranded “Insights Hub”) runs on AWS and Azure infrastructure; the Mittelstand increasingly pipes its factory data through American cloud services; and only about four percent of global cloud capacity is European-owned.

Germany had the factory floors, saw the opportunity, defined the concept, built the machines, funded the research, created a national strategy with cabinet-level backing… and then watched American cloud companies capture the data platform.

The Carrier Layer Is Not Enough

Automotive is the hardest possible case for this thesis. In no other sector does Germany hold more of the traditional advantages: decades of engineering excellence, deep supplier networks, massive production capacity, global brand equity, and a domestic regulatory framework shaped around its own industry. If these advantages cannot produce an AI-native platform developer in automotive—the sector Germany has dominated for a century—there is little reason to expect one to emerge in domains where Germany has no comparable legacy position.

It would be good to be proven wrong. The argument would be falsified if a German OEM were to announce a full-stack automated-driving subsidiary—genuinely owned, not a majority-minority partnership with a foreign platform provider—backed with Waymo-scale capital and operating driverless miles on public roads. Such an announcement would be evidence of a structural shift. The Mercedes–Uber robotaxi collaboration, tentative as it is, is not that evidence.

The structural mismatch between Germany’s carrier economy and the platform economy goes deeper than capability. One major factor keeping German carmakers from investing in breakthrough R&D such as automated driving systems is the ruinous cost of failure. Restructuring is significantly costlier in Germany than in the US: estimates cited in a recent article for Works in Progress by Pieter Garicano put a German corporate restructuring at the equivalent of 31 months of salary per employee laid off, compared to just 7 months in the United States. High severance costs create a fundamental incentive for European businesses to “avoid innovative areas and concentrate on safe, unchanging ones.” This is not because Europeans are inherently more risk-averse, but because avoiding innovative bets is the rational response when the downside of failure is so much larger. Fixing it would require labour market reform; Garicano suggests, for example, allowing workers above the 90th percentile of income to opt out of employment protections entirely, which would make German services highly competitive with the American market.

The 716,000 German automotive workers are navigating a double transition—from combustion to electric, and from mechanical engineering to software-defined vehicles—with employers who are themselves becoming integrators of foreign capability rather than developers of their own. The carrier layer is not disappearing, but it is consolidating, automating, and generating less value per unit than the platform layer above it.

Share

Källenius told his podcast audience in September 2025 that Mercedes was “at the front” of driving automation. Four months later, that claim can’t be squared with the evidence. But of course, the problems didn’t emerge over those four months; they’ve been accumulating for thirty years. Germany’s genuine strengths in precision engineering, manufacturing quality, and systems integration are the skills that matter least in a competition decided by data accumulation and iteration speed. If the pattern established in automotive holds elsewhere, and there is evidence to suggest that it will, then Germany has a real problem.

It can rise to the occasion, or it can decide to keep doing what it did so well last century. If it chooses the latter, then Germany’s position in the next economy will be that of an excellent supplier, to platforms it does not own, in markets it does not control.

Respect to and for feedback on earlier drafts.

In 2020, Wendy’s and REEF Technology announced they would build 700 delivery kitchen locations across North America. These would not be restaurants, but just kitchens, sited in parking lots and operated entirely for delivery. It was a confident bet about a new way fast-food brands could make money. Not with dining rooms, drive-throughs, and a real-estate premium, but with just the food, dispatched to wherever the customer happened to be.

Less than two years later, Wendy’s CEO told investors the company no longer saw delivery kitchens as part of its growth trajectory. The target, which had already fallen from 700 to 150, fell to zero. Why the pivot? Because average annual revenues at REEF locations had come in below $500,000, or less than a third of what a typical Wendy’s restaurant generates. The REEF locations disappeared, and the parking lots went back to storing cars.

Ghost kitchens, commercial kitchens with no dine-in option that prepare food solely for delivery, were supposed to transform urban food service. A recent edition of Brian Potter’s Construction Physics newsletter, “Is the Future ‘AWS for Everything’?”, is worth reading for its main argument about economies of scale; right at the end, he uses ghost kitchens as a quick illustration of how those economies don’t always pan out, then moves on. Potter’s piece reminded me that five years ago, I was quite bullish on ghost kitchens as the future, before I set the matter aside to think about other things. Looking into it, I was surprised to learn that my analysis back then was dead wrong.

The evidence shows that ghost kitchens have indeed failed. But the emerging narrative about why—pandemic buzz around food delivery, venture capital poured in, post-pandemic consumers returned to restaurants, concept collapsed—is a poor explanation.

Here’s a better account. The ghost kitchen concept bundled two distinct, legitimate claims—a real estate thesis about urban land arbitrage and a mobility thesis about the structural shift to off-premises dining—and both were borne out. But there was a third, hidden claim about who would capture the value generated by those kitchens, and that assumption proved spectacularly false.

Subscribe now

Two Good Ideas…

Ghost kitchens were delivering on two claims at once.

The real estate thesis said: urban land near residential density has value that traditional restaurants aren’t capturing.

At minimum, a good foodservice operation of any kind requires quality ingredients and a capable kitchen, by which I mean not only the physical capital of stoves and ovens, but also the human capital to use it.

A traditional restaurant needs these, but also much more. It needs a ‘front of house’: hosts, servers, and others who deal with customers. It needs premises functional enough to serve patrons and attractive enough to make them comfortable: not only tables, chairs, cutlery, and so forth, but square footage to feature them all. It needs to be reachable, meaning it needs adequate parking, decent transit access, or both. It needs to be in a part of the city that patrons will find to be safe and appealing. And it needs to be close enough to where people live and work that the trip doesn’t feel onerous. Taken together, these requirements mean that the restaurant must be itself, and be in a place, that is desirable, accessible, and visible.

That combination costs money. The rent or mortgage a restaurant must be able to cover imposes a floor on the revenue it needs to generate.

The appeal of a ghost kitchen was that it could shed most of this. Forgoing front-of-house staff would cut operational costs, but more importantly, it would cut capital costs. It could do that in two ways: not only would the ghost kitchen not need to devote space to customer areas, it wouldn’t need to site itself on expensive land at all. It could locate itself in an unappealing part of the city, where land was cheap; even, at the smallest scale, in a retrofitted shipping container in a low-volume parking garage.

But that didn’t mean it could set itself up anywhere. Instead of needing to be somewhere customers would be willing to travel, it would need to be close enough to where enough customers lived that their meals could arrive in a timely fashion, and where delivery costs wouldn’t break the value proposition. Last-mile delivery cost and speed depend heavily on proximity to the people being served, a mechanism that readers of Changing Lanes will recognize also applies to transit and freight.

That was the real estate thesis: that there was enough commercial land unsuitable for a traditional restaurant, but near enough to dense clusters of people, to support this new business model… and that the market hadn’t yet priced this value in, meaning there would be arbitrage opportunities by moving quickly.

The mobility thesis said: restaurants built around sit-down service are poorly designed to also offer delivery service.

This claim follows from how restaurants are built. Their layouts are organized around the dining room: the kitchen is sized to feed the tables, not to maximize throughput. Their equipment and workflows are all calibrated for food that travels ten metres to a table, not ten minutes in an insulated bag. Their ingress points are not optimized for courier access; their front entrances are easy to reach, but courier vehicles would then be taking up space for patrons, and patrons don’t like it when delivery trade is happening in their space. Couriers might instead use back entrances and not disturb patrons, but those areas are deliberately hard to reach, and often can’t be accessed quickly, which delivery couriers require.

This awkwardness supposedly gave ghost kitchens an edge. For the delivery market, a purpose-built production kitchen, designed from the start around delivery—optimized for throughput, packaging, and dispatch rather than ambience and table turns—would in theory outperform a retrofitted dine-in operation on every dimension that mattered.

All of this is true a priori, but was meaningless until quite recently, because food delivery was a niche business. Younger people may not believe it, but well into living memory, the only food one could readily order and have delivered to one’s door was pizza (and, in some markets, Chinese food). The advent of the mobile Internet changed this by making delivery platforms feasible. Once it was possible to order food from one’s couch via a phone's interface, DoorDash, Uber Eats, and SkipTheDishes emerged to aggregate demand and handle logistics. Before the pandemic, meal delivery was already becoming a broadly accepted substitute for going out, but Covid accelerated the trend by making it the only option: dining rooms closed, people stayed home, and delivery volume spiked.

The arrival of food delivery as a business sector made the gap between an existing restaurant's ability to serve delivery markets and a kitchen designed for that purpose apparent for the first time. That was the mobility thesis: that custom-built kitchens for the delivery market could outperform traditional restaurants in that market, and that there was a first-mover advantage to siting and building those kitchens where they could offer the most value.

These were different bets, about real estate and about the food-service market, but proponents bundled them into a single proposition: a commercial kitchen without a dining room. What connected them was a shared assumption that whoever controlled the production infrastructure would capture the value it generated.

That assumption, it turned out, was the load-bearing one.

…and One Bad Mistake

Before turning to where the analysis broke down, it’s worth being precise about what didn’t. Both underlying theses have proven valid.

Off-premises dining is not a pandemic blip that reverted when lockdowns ended. The demand that ghost kitchens were built to serve is real, growing, and profitable.

Data visualization courtesy of Claude, sourced from NRA, Euromonitor, and Uber

The centralized production logic is also sound, in contexts that match its requirements. Reports say that of kitchens serving caterers, meal prep businesses, packaged goods producers, and institutional clients, two-fifths are making a profit and another two-fifths are breaking even. Meanwhile, hospital and institutional food service has adopted centralized production, with meaningful budget savings. These aren’t ghost kitchens as anyone in venture capital imagined them, but they vindicate the underlying logic: high-volume, repetitive, centralized production serving known customers through predictable channels works.

So ghost kitchens had every opportunity to generate business. Unfortunately, what they didn’t have was an opportunity to make a profit on that business. The shared assumption—that controlling the production infrastructure meant capturing the value it generated—failed, and it failed on both theses separately.

The mobility thesis broke on the customer relationship. A purpose-built delivery kitchen, however superior on every production dimension, owns the production but does not thereby own the customer. That’s true for two reasons.

The first is trust. The National Restaurant Association found that 70% of consumers say it matters to them that their food comes from a restaurant with a physically-accessible location. Customers who discovered they had ordered from a major chain operating under a made-up virtual brand name—who had been, as they put it, “catfished”—were upset because they had been deceived about something they cared about, and insisted that such places had to explicitly disclose that that was what they were.

One reading of this preference is that people want the dining experience: the room, the occasion, the presence of other people eating around them. On this reading, takeout is a substitute for dining in, and a delivery partner that has no such option is no substitute at all.

That reading isn’t wrong, but it’s incomplete. We might imagine customers don’t want to order from ‘No-Name Burger Bar’, which lacks brand value or associations with an in-house dining experience; but they should, on this view, be perfectly happy to order from a ghost McDonald’s or ghost Burger King. The whole point of big burger chains like these is that their food is a commodity, prepared the same everywhere, and they have deep and broad brand association and value, meaning that such a facility should not face any trust problem at all.

And yet Wendy’s, whose collapse from 700 REEF locations I wrote about at the beginning of this piece, is the clearest illustration that this reading isn’t enough. Less than two years after the original deal, the unit economics had simply never materialized.

A Wendy’s ghost kitchen, image courtesy of of this Wendy’s press release from 2021

Why? The answer appears to be that consumers are being practical. They want accountability, someone to hold responsible when things go wrong. And as far as customers are concerned, a dine-in restaurant that one can visit has (or seems to have) feedback loops that will make the operators care about hygiene and food safety in ways that an anonymous industrial kitchen will not.

Whether or not this is rational, this is how customers seem to think, and the customer is always right.

The second reason was lack of customer loyalty. A restaurant with a physical presence earns returning customers passively: a good experience encourages repeat business. A virtual brand is not so fortunate, because it doesn’t sell experiences, just food as a commodity. Every customer therefore is a new prize to be won, from the beginning, every time, through a platform that charges for the privilege and has no stake in whether any particular brand survives. 41 of 71 restaurants across five CloudKitchens locations had closed within one year; a 58% annual failure rate, because customer acquisition was a permanent operating cost, not a compounding investment.

Lack of trust or loyalty are two faces of the same coin, namely that ghost kitchens have no independent relationship with their customers. The platform owned that relationship entirely.

So the mobility thesis did not hold. How about the real estate thesis?

Unfortunately, it broke too, and on a single, decisive point. The real estate thesis, you’ll recall, was that operating kitchens rather than restaurants, and on low-value land, would create a significant margin relative to incumbent food producers, and the ghost kitchens would capture that margin.

They reckoned without the delivery platforms. The platforms wanted that margin, and the ghost kitchens couldn’t stop them from taking it.

A restaurant using a delivery platform pays a commission on every order, and often additional fees for marketing placement and payment processing as well. At the time the ghost-kitchen investment cycle was spinning up, the pandemic had made food-delivery the only game in town for meals outside the home, and so multiple U.S. cities moved to cap them at around 30% of the order price. It was against that backdrop that ghost-kitchen investments were made, assuming those rates as a baseline.

Unfortunately low rates were an artifact of the pandemic, and began to rise as Covid receded. While some cities, like San Francisco and New York, attempted to make their caps permanent in 2021, the platforms pushed back in court, and by 2023, they had largely won. New York held out longest, but as of April 2025, even that city had compromised. The new framework there allows 15% for core delivery, 5% for basic marketing visibility, and 3% for payment processing, plus an “optional” enhanced services tier of an additional 20%, covering expanded delivery radius and promotional placement, for a potential total of 43%.

(And 43% is not really optional; operators who decline the enhanced tier are deprioritized in platform search rankings, meaning that the real choice is between paying 43% or being buried algorithmically.)

Certainly a conventional restaurant using a delivery platform pays a commission too, but this was not an existential risk for those businesses, since they only pay it on the portion of revenue that comes from delivery. For most full-service operators, that’s a minority of income. Most comes through the dining room, which the platform never touches. Since a ghost kitchen has no dining room, the commission that is manageable for a restaurant with tables is lethal for an operator without them.

In any case, the failures of the mobility and real-estate theses reduce to the same point, which is that they both assumed if the food producer could generate value, they could keep it. But that was wrong: in the delivery economy, the platform can seize that value, and will.

The Platform Takes the Margin

What remains is an opportunity that hasn’t materialized, and seems unlikely to, even though the two theses about their value are still true. Parking lots are still underused, dead retail is still everywhere, and cities still contain spaces that sit between the density that creates demand and the production infrastructure that could serve it. And it will always be the case that a restaurant optimized for serving in-house customers will find it difficult to serve the delivery market.

Given that, what would need to change for ghost kitchens as originally conceived to succeed? In my view, one of two things.

Share

One would be a substantial fall in delivery platform commission rates. That seems unlikely, given the direction of the NYC settlement and the platforms’ demonstrated willingness to litigate caps into the ground.

The other would be operators building direct-to-consumer channels that bypass platforms entirely. That seems equally unlikely: it requires exactly the kind of marketing investment and customer-acquisition infrastructure that the ghost kitchen model was designed to shed.

I think the old analysis is still correct; urban land near residential density has logistics value that is currently underused. But it was an error to assume that the value created by unlocking that land would accrue to the one who did the unlocking. It accrued, instead, to the one who controls access to the customer. In the delivery economy, that’s the platform.

This is a pattern that Changing Lanes has traced before. In January, writing about Uber, I argued that what Uber’s subsidy years bought wasn’t network effects, but market power. By the time Uber started charging riders more and paying drivers less, there was no competitor left to undercut it and no regulator with the will to stop it.

Ghost kitchens have been losing the same race. The delivery platforms that ghost kitchens depended on had spent years, and billions, acquiring the same kind of structural dominance. When ghost-kitchen operators tried to capture the margin their model had created, the platforms were already there, and better armed: with commission structures, algorithmic levers, and litigation budgets against regulators.

The question of who captures the surplus generated by urban-logistics infrastructure is the key question of platform urbanism. Ridehail and ghost kitchens are a case study in the answer to that question. The answer, so far, seems to be that the platform always wins.

Respect to for feedback on earlier drafts.

In 1994, France opened Haute-Picardie, a high-speed rail (HSR) station roughly equidistant from Amiens and Saint-Quentin, about 40 km from either. Shuttle buses link it to both cities, but there is no connecting conventional rail service. Locals call it la gare des betteraves, the beetroot station, because the fields surrounding it are full of beets, not people. Annual ridership peaks at around 400,000. For comparison, Arras, the next station in line, has 4 million annually; Lille-Flandres, the next stop the other way, has 80,000 passengers daily.1

The chief value Haute-Picardie station gives the world is serving as a cautionary tale of why we should build transit stations not where it’s cheap or politically convenient to construct them, but instead where passengers want to go.

I’m thinking about Haute-Picardie because I’m thinking about Alto, the putative high-speed rail link that will connect Toronto to Montreal via Ottawa.2 I’ve written about Alto twice before; once arguing that the project is structurally flawed from the outset, and once that the project depends on a regional rail network that doesn’t exist, setting it up as an inferior substitute to air travel. I remain skeptical about whether Alto should be built at all, but if it is to be built, and increasingly it looks like it might, then I want it to succeed… and its success will depend on where it places its stations.

The emerging picture is not encouraging.

Station placement is not a mere detail; choosing where the stations are is the decision point that will most determine the success or failure of the line. And in Canada’s two most important cities, the structural conditions already look more likely to produce the gare des betteraves than the Gare de Lyon.

Subscribe now

Station Placement First, Everything Else Second

Despite the name, high-speed rail’s value proposition is not about speed, but about door-to-door travel time. That may seem to be a fussy distinction without a difference, but it’s not. It’s the crux of the matter.

On pure speed, trains can’t compete with aircraft. The fastest HSR in the world, parts of Japan’s Shinkansen network, reaches around 300 kph, while a typical commercial airliner cruises at roughly 800 kph. Air wins on speed by a factor of nearly three, and that ratio will not change in rail’s favour.

Nonetheless, HSR consistently wins versus air on many corridors: Paris–Marseille, Madrid–Barcelona, and Tokyo–Osaka, to name a few. Why? In part because HSR does away with most of the security theatre surrounding air travel, meaning that travellers can arrive minutes before departure rather than hours… but also, and more importantly, because they arrive somewhere they want to be. Not at a facility optimized for aircraft, typically far from the city centre, but at a station in the urban core, walkable to many destinations and connected by transit to the rest.

The arithmetic is revealing. A careful Toronto–Montreal traveller flying today faces roughly 45 minutes getting to Pearson, a two-hour arrival buffer recommended by airlines and CATSA for domestic flights, an hour in the air, baggage claim and deplaning, and then 45 minutes or so into downtown Montreal: something under five hours door-to-door. Conversely, Alto promises a three-hour train journey. Add 30 minutes at each end for first/last mile travel to and from downtown stations, and the total is four hours, meaning that Alto wins. HSR’s victory is partly because it foregoes security theatre, but as I have argued elsewhere, that’s neither a necessary nor permanent advantage. What can never be taken away from HSR is the placement of its stations: HSR can deliver passengers into the city rather than to a facility outside of it.

Erase that advantage by putting one or both stations somewhere inconvenient, and the arithmetic shifts; one bad choice can make HSR’s win on travel time into a tie, or worse. The TGV from Paris to Marseille captures roughly two-thirds of all journeys between those cities, not because it’s faster than flying—it isn’t—but because Gare de Lyon in Paris and Saint-Charles in Marseille are both centrally located. The competitive advantage lives in the stations.

Alto Is Poised to Compromise

Alto plans roughly 1,000 km of new track between Toronto and Quebec City. That choice makes sense: high-speed trains can’t share track with the freight-optimized network VIA Rail currently uses, and dedicated infrastructure is a prerequisite for the speeds that justify the project’s existence.

But dedicated infrastructure needs somewhere to go, and in Canada’s two largest cities, the right places for it to go already seem to be out of bounds.

In Toronto, the obvious candidate is Union Station. It already serves as the intermodal hub linking the subway (TTC), the regional rail network (GO), and the intercity rail network (VIA), and anchors the city’s central business district, which is immediately accessible on foot, via two distinct networks: the surface-level streets and the underground PATH tunnels. This dense web of connections and accessibility makes Union Station Canada’s busiest transport hub, hosting 65 million passenger trips per year. It was no surprise when Alto’s CEO Martin Imbleau told the Senate transport committee that, with respect to Toronto, “the objective would be to have a station in the vicinity of Union Station”.

It was surprising, to me at least, when he immediately qualified that with “if it’s feasible and we can make it affordable… it needs to be economical”.

I can understand not arriving in Union itself. Despite recent, lengthy, and remarkably-expensive upgrades, the facility may be at or even over capacity. But an Alto stop being merely close by would be fine. One might imagine that Alto being in the vicinity of Union Station would mean something like how Toronto’s intercity bus terminal or air-rail link station are nearby.

One has to imagine, because Alto has not been more specific. The project’s consultation documents identify three candidate zones: one near Union, an eastern site, and a western site, but without details. The eastern site would be “close to public transit and the financial district”, and the western site “closer to attractions like the CN Tower and Rogers Centre”, but no more detail is available. Alto has stated it’s studying approaches “from the north or the east, using existing corridors or solutions such as tunnels or elevated tracks,” but where they imagine siting the station has not yet been revealed. Alto has also not released any information on tunnel lengths, engineering assessments, or cost estimates that would allow anyone to judge what it means by vicinity.

The most specific candidate to emerge from independent analysis is East Harbour, the mixed-use development planned for the former Unilever site near the mouth of the Don River. East Harbour has genuine transit credentials: the Ontario Line will serve it, and Metrolinx has long planned a major interchange there. In that sense, it satisfies Alto’s stated criterion of being “close to public transit.” But East Harbour sits roughly three kilometres east of Union Station; three kilometres away from everything that matters, but close to nothing important. A passenger arriving there and wanting to reach the central business district, or government precincts, or any tourist attraction still faces multiple transfers through local transit. The western candidates are worse: they draw tourists, not commuters, and the surrounding streets offer nothing approaching Union’s transit density.

The implicit benchmark Alto should be aiming for already exists, in the form of the two facilities directly across the street from Union. The Bay Street bus terminal is a distinct building, connected by a pedestrian bridge. The UP Express station at John Street is a covered walk away. Both are technically separate from Union Station, but neither feels that way, because passengers move between them on foot. That is the standard that we should hold vicinity to: not a twenty-minute transit trip, but a ten-minute indoor walk. Reaching that standard, in a corridor approaching downtown from the north or east, will almost certainly require a dedicated tunnel or elevated structure into the core of the city. That would incur an enormous cost, but the alternative is worse: building a terminal station in Canada’s largest city that no one wants to use.

Bad as the Toronto situation is, Montreal’s is even worse.

Alto’s current routing hypothesis is that HSR will reach downtown Montreal from the north. The obvious way into Montreal from the north is the pre-existing Mount Royal Tunnel and its access to the Gare Centrale. Unfortunately, to serve Montreal’s REM light metro, the tunnel was converted from heavy rail to automated light metro standards, installing REM-specific signalling and building two new underground stations within it. The REM runs automated driverless trains at headways as brief as 90 seconds, using technology that is incompatible with conventional or high-speed rail in every way: gauge assumptions, signalling, vehicle dimensions, and power systems. In other words, interoperability with the REM through the Mount Royal Tunnel “will not be possible“, meaning that Alto’s access to the Gare Centrale is unlikely.

Since the pre-existing downtown access route is unavailable, Alto proposes to build its own. It is studying a tunnel, perhaps as long as 10 km, under Mount Royal. Recent Canadian tunnel construction costs suggest a price of more than $1 billion (CAD) per kilometre, and if the project ends up paying that much, the Montreal access infrastructure would cost somewhere between $7 billion and $16 billion, so between 12% and 18% of the total project budget.

Nor is that all. Coming in from the north means that the Montreal tunnel, being built at great expense, will be a stub line; trains will arrive from the north, stop, then reverse back out to continue east toward Quebec City.

This may sound like a minor inconvenience, but it is not.

A commuter train terminating at a downtown station reverses as a matter of course; the train arrives and empties; sometimes the crew walks to the other-end cab, sometimes they stay where they are; and the train departs as a new outbound service. There is no time penalty, since the switchover is accomplished while passengers alight and new ones board. GO trains at Toronto’s Union Station work exactly this way, and they work fine.

But Alto’s Montreal station is not a terminus; it’s an intermediate stop on a Toronto–Quebec City corridor. Every train that serves Montreal must arrive, stop, and then continue in the opposite direction. That means every through-service is doing what GO does at Union, but in the middle of its journey, not at the end. The difficulties will compound quickly. HSR can’t be run from the back, with the locomotive pushing rather than pulling, as GO does; a mid-journey reversal requires the crew to walk the full length of the train and reconfigure safety systems for reverse-direction operation. A realistic estimate for a well-run reversal at an intermediate HSR station is ten-to-twenty minutes added to every schedule.

For commuter service, fifteen minutes is noise. For HSR service whose value proposition is compressing the Toronto–Quebec City journey to something competitive with flying, fifteen minutes is significant: a permanent tax on every through-journey for the life of the line. Worse, reverse running is a capacity constraint on the entire corridor: Alto will only run as many trains through the Montreal station per hour as it can complete reversals, which is far fewer than a through-station allows.

Italy has been living with this problem since the 1990s. The Florence station, Santa Maria Novella, sits in the middle of the Milan–Rome corridor, and every HSR service that serves the station must reverse there. The result is a systemic throughput ceiling that has constrained Italy’s primary HSR route for three decades. After twenty years of planning, Italy is currently spending €2.73 billion to build a 7 km underground bypass beneath Florence precisely to eliminate that reversal (construction began in 2023 and completion is still years away.)

The Florence high-speed rail bypass (Passante AV di Firenze) project location. This satellite map shows the relationship between Florence’s historic centre (Centro storico), the existing Firenze Santa Maria Novella terminal station, and the planned high-speed bypass infrastructure:

• The orange lines and box show the existing rail lines and station

• The red area marks the excavation site for the new underground Firenze Belfiore high-speed station

• The dotted yellow lines mark the alignment of new HSR-only tunnels that will allow high-speed trains to serve Florence without having to reverse3

The lesson is unambiguous: designing a reversal into your system from the outset is a foundational error and patching it will require extraordinary cost later. Alto’s current hypothesis for Montreal is making this error now.

The patterns of concern we see in Montreal and Toronto exist elsewhere. In Ottawa, the politically-preferred station option is the former Union Station at 2 Rideau Street. It’s convenient to downtown, but reaching it would require a new tunnel of approximately 4 km, and the original building is a stub-end terminal, meaning yet another reverse-run on the line. The operationally sensible option is the existing VIA Rail station on Tremblay Road, which has through-tracks, room for expansion, an on-site LRT station, and highway access… but is four LRT stops from downtown. Neither option is ideal.

The intervening stop between Ottawa and Toronto, Peterborough (population less than 100,000) is a gare des betteraves in the making. Trois-Rivières and Laval are not so obviously inappropriate, but the pattern persists: vague references to transit-connected sites without specifics, meaning all the intermediate cities are drifting toward the greenfield model that produces beetroot stations.

Not a real HSR station. Let’s keep it that way

What Alto Needs to Do

No one building HSR in a dense, expensive, mature city finds it to be easy. To pick only one example, HS2 in Britain is currently spending billions tunnelling into London’s Euston specifically because the alternative, terminating at Old Oak Common, was described by a government source as “pretty much the definition of a railway to nowhere“. I don’t fault Alto for taking its time to determine what its options are and how much they will cost.

But they do need to make their thinking public at some point. When they do, what would change my assessment of Alto’s prospects is a committed, publicly-announced station location in downtown Toronto with a specific site, a credible access plan, and an honest cost estimate; a solution to the reversal problems in Montreal and Ottawa; and a demonstrated understanding, at the project’s leadership level, that the station question is prior to all other questions. That is, station placement is the constraint around which everything else must be organized, not a consideration to be traded against others.

We live in hope of these things, but that hope is fading in the face of discouraging language: “the intent is to go downtown,” “it needs to be economical and affordable,” and it is “too early to speculate on locations”. These sound reasonable in isolation but alarming in accumulation.

They’re alarming because getting this right is fundamentally important.

Share

For Alto to succeed, downtown stations—walkable to many major destinations, connected by higher-order transit to the rest, and configured to permit through-running—are prerequisites. Without them, Canada will spend tens of billions of dollars (possibly more than a hundred billion by the end) on a project that cannot compete with the air links that already connect these cities. This would mean that upon completion, we would then need to spend even more to get something that actually works.

Constant readers know that, in my view, this project isn’t worth doing, and that the money would be better spent making air travel work better: more sustainably, and with fewer delays. But disagreeing with me to insist that Canada build HSR is defensible.

But it’s only defensible if you bite the bullet that we must spend whatever it takes to get the stations we need.

Believing we can build high-speed rail on the cheap is building half a bridge: spending too much money for something that won’t succeed. The obvious example of that failure mode is California HSR, which promised to connect Los Angeles and San Francisco, and may end up connecting only Bakersfield and Merced. The political logic that produced that outcome, namely putting off hard decisions and doing the easy things first, is precisely the logic that Alto’s official communications are beginning to replicate.

Alto will deliver the speed in high-speed rail; of that I have no doubt, because it’s the easy part of the project. The hard part, which will determine whether Alto only duplicates existing air service at vastly greater expense, is where the trains will stop.

On current evidence, we should all be concerned.

Respect to for feedback on earlier drafts.

In the early twentieth century, American streets were shared spaces. Pedestrians walked where they wanted, when they wanted to. Vendors plied their wares from the roadbed, next to children playing in the street. Scenes like this, still common in many parts of the world today, were the natural default, extensions of what streets had always been: mixed-use public space.

Then automobiles arrived, and the industry and its allies spent the better part of a decade working to redefine who owned them. At the time, jay was slang for a rube too dim to understand how to behave in a city, so the industry coined the term jaywalker for anyone who walked in the street without benefit of a signal. They went so far as to lobby newspapers to print the word without scare quotes, to show the term’s implicit claims made up the consensus view.

In 1925, responding to this effort, Los Angeles passed an ordinance that declared pedestrians had to cross at intersections, and only at intersections, and only if the light was green. The rapid adoption of similar laws across the country is evidence that the automakers’ campaign had triumphed: whereas only a few years before the question had been whether cars belonged in the streets, it was now whether pedestrians did.

"Don't Jay Walk — Watch Your Step" (1936 or 1937), Isadore Posoff. Library of Congress. Public domain.

This is the story of how ‘motordom’ became dominant in urban transport planning. Among urbanists, it is widely known, thanks to Peter Norton’s documentation of it in Fighting Traffic (2008), where he coined that term. Norton’s work has received wide currency because it fits naturally within what has become the dominant lens in progressive transportation discourse. In this framework, the first question one asks regarding a change is always whom it privileges, and whom it constrains. Who is assumed to own the street? Who must adapt? Who must yield?1

Let me be clear: motordom has done real damage to the city, and the pedestrian-priority framework is right to push back. Car-centric planning has caused genuine, lasting harm, in ways I won’t bother rehearsing here: clearance of urban neighbourhoods for expressways, Robert Moses, oversupply of street parking, subordination of pedestrians and cyclists to vehicle throughput, and so on. Motordom has made our cities uglier, more dangerous, and less equitable, and the people who bear the worst of it are consistently those with the least political power. Every urbanist knows this sorry tale.

But this itself has become a problem: every urbanist knows this sorry tale.

Subscribe now

As I will show below, they know it so well that it has calcified from a historical explanation into a reflex. As a result, whenever a new claim about vehicles, about pedestrians, about who is endangering whom appears, it is not examined on its own terms, but is instead pattern-matched to the motordom template. New things are classified as instances of the old Bad Thing and dismissed rather than engaged; the conclusion precedes the analysis.

Call this the Jaywalking Analogy. It is familiar and legible, which makes it persuasive. It assigns good guys and bad guys, which makes it seductive. And it is applied indiscriminately, which makes it an obstacle to thinking clearly about automated vehicles.

Analogies Aren’t Arguments

In August 2025, here in Changing Lanes, I argued that Robotaxis Have a Bullying Problem. The problem in question was that pedestrians in San Francisco had discovered that Waymo robotaxis would reliably stop if a person stepped in front of them, and were exploiting this behaviour by crossing in front of them when they had the right-of-way, knowing the Waymo would stop. I argued this was a genuine problem: that the behaviour would spread as it became known, that it created real inefficiencies in urban transport networks as a whole, and that, left unaddressed, it would make city-dwellers as a whole worse off, especially the most vulnerable.

Lots of people liked the piece. Lots of others didn’t.

Pushback came from many directions, mostly on social media and mostly in the form of short, dismissive, even abusive remarks. I won’t reward that sort of behaviour with attention, so I will focus here on ’s response at Carbon Upfront! It pains me to do so, because Alter, a sharp thinker and writer, actually put in the work by writing a full essay in response rather than merely making a snarky comment. I regret singling him out, since he took the highest and best approach toward writers with whom we disagree: not performative dunking for the crowds, but fully written-out rejoinders. But since he did the work, I want to highlight it here, because his response demonstrates the mechanism I’m describing clearly.

Alter’s essay is titled “Will self-driving cars and robotaxis bring us Jaywalking 2.0?” Why did he call it that? Because in his view, I was merely running the jaywalking playbook of a century ago, this time on behalf of automated vehicles. And in my view, he reached this conclusion not by examining the argument’s evidence, but by pattern-matching, as follows:

1. Here is a piece raising concerns about pedestrian behaviour near vehicles, and

2. It maps onto the motordom campaign of the 1920s, which means

3. It is a version of that campaign.

Alter goes on to predict that any policy response to robotaxi exploitation would inevitably turn authoritarian—pedestrians will be fenced off, criminalized, or surveilled—because “laws everywhere are designed to favour drivers and penalize people who walk or cycle”.2 He quotes Rebecca Solnit approvingly: “We don’t need new ways to use cars; we need new ways not to use them.”

Notice what the Jaywalking Analogy did here. Confronted with a specific, novel situation, namely pedestrians discovering they can exploit an automated vehicle’s safety programming, the essay classifies it as an instance of a known bad pattern before asking whether it actually fits that pattern. Consequently, none of the evidence in the newsletter needed to be engaged with, because the conclusion had already been reached.

The specific weaknesses in Alter’s analysis are worth pointing out, because they show how much gets skipped when the pattern fires early.

Read more

This might seem like textbook protectionism. Ford leads a jurisdiction whose economy depends heavily on auto manufacturing.1 His comments came just as Prime Minister Carney announced a trade deal allowing up to 49,000 Chinese EVs per year into Canada, vehicles that will compete directly with the products of Ontario’s auto sector. Retired CSIS officer Neil Bisson piled on, warning of “another portal into our infrastructure, both communication-wise and energy-wise.” Cybersecurity CEO David Shipley maintained his earlier characterization of Chinese EVs as “rolling spy vans.”

It’s easy to imagine this is just fear-mongering: dragging out hackneyed fears of The Other to justify protecting a favoured domestic industry. Canada’s federal government, which negotiated the deal, certainly seems to think so; Public Safety Minister Gary Anandasangaree remarked blithely that “From a Canadian safety and security, public safety perspective, we don’t have concerns.”

Unfortunately, Anandasangaree is wrong, and Ford is right. Chinese electric vehicles absolutely can function as espionage platforms. The technical and legal capabilities exist, as do the political incentives. This isn’t fear-mongering, but an accurate description of how connected vehicles work, what they can do, and what Chinese law requires of Chinese companies.

Having said that, Canada should let the Chinese EVs in anyway.

Why Doug Ford Is Right

China’s 2017 National Intelligence Law requires companies to “support and cooperate” with national intelligence efforts. This is a binding law that Chinese automakers must obey, and there are a variety of ways they could assist with the work.

We don’t often describe them in these terms, but modern connected vehicles are mobile computers that can record their entire journeys and everything that happens inside the car while they do so. To put it more precisely: modern vehicles are sensor platforms that can capture external video from their surround cameras (used for parking and driver assistance), as well as a continuous record of where the vehicle went and the precise route it took to get there (using GPS and inertial-location traces). In that regard, they also log speed, braking, and steering inputs, which they collect as frequently as every three seconds.

That would be bad enough, that a bad actor might be able to know all the details of our trips, i.e., where we went and how we got there. But connected vehicles also collect data on the car’s occupants. They take information from internal video from driver-monitoring systems and microphone audio from always-on voice assistants. The avowed purpose of this is to assemble datasets on seatbelt use, occupancy, and biometric proxies like driver attention and fatigue, but it takes little imagination to conceive of other purposes these capabilities could fulfill.

It seems like belabouring the point to add that if you pair your phone to the car’s audio system, the car also captures all of your contacts.

Morio, MG iGS top camera, Auto China 2016. 2 May 2016. Licensed under CC BY-SA 4.0.

Once collected by a Chinese-built vehicle, this data will, per the National Intelligence Law, be available to Chinese intelligence services. And clearly the PRC thinks it will be valuable to them: when Tesla began selling cars in China, it had to build a Shanghai data center, because Chinese law also stipulates that vehicle data captured in China must not be stored abroad. Beijing takes its ability to access vehicle data in China seriously; it’s reasonable to it would find vehicle data from other countries to be equally important.

So, for the record, Ford, Bisson, and Shipley aren’t crying wolf, nor exaggerating a threat to back a protectionist play. Chinese EVs entering Canada could indeed serve as espionage platforms, feeding data back to Beijing through legal compulsion that Chinese manufacturers cannot refuse.

Before you decide how concerned to be, though, please remember that everything you fear a Chinese car might do, American cars are already doing.

Here in Your Car, I Can Listen to You

Between 2018 and 2024, GM collected driving behaviour data from vehicles participating in their OnStar Smart Driver program: acceleration, speed, hard braking, late-night driving, location. They collected this from an estimated eight million vehicles, as often as every three seconds.

Having collected it, GM sold this data to LexisNexis Risk Solutions and Verisk Analytics for the sum of fifty cents per vehicle. LexisNexis and Verisk used it to create “risk scores” for the associated drivers, which the firms sold on to insurance companies. The insurers used them to raise premiums: some drivers saw theirs increase by as much as 80%, and one driver was rejected by seven consecutive insurance companies. To its credit, the USA’s Federal Trade Commission (FTC) moved in January 2025 against this practice. Their proposed order, finalized last month, bans GM from sharing drivers’ geolocation and behavior data with consumer reporting agencies… for five years.

(Why not indefinitely? I wish I knew.)

Nor was GM alone. Hyundai shared data from 1.7 million vehicles, which they captured from any customer who activated their car’s onboard internet connection. Honda also played the game, though they didn’t play it as well: they only sold data from 97,000 vehicles at 26 cents per car.

Over 25 class action lawsuits against GM alone have been consolidated into a single proceeding, which is underway, but the damage has been done and the point established: American and Japanese carmakers capture lots of data from their loyal customers, and they don’t feel any need to protect it or their customers’ best interests. Instead, they sell it to willing buyers for trifling sums: given a choice between protecting your entire driving history or being given a dollar (or less!), the carmakers choose the dollar.

Sometimes they don’t even need a dollar; they just need to be asked nicely.

According to a 2024 US Senate investigation, Toyota, Nissan, Subaru, Volkswagen, BMW, Mazda, Mercedes-Benz, and Kia share vehicle location data with law enforcement upon receiving a subpoena.2 A subpoena is a very low bar to clear: unlike a warrant, it doesn’t require judicial approval, nor a finding of probable cause, just that the material is relevant to a proceeding already underway. Canadian law sets a nominally higher bar: production orders compelling data disclosure from third parties generally require a judge’s sign-off. But that formality matters less than it seems, because nothing prevents automakers from sharing data voluntarily before any legal process begins, and no Canadian investigation has documented whether they do.

And the police know exactly what to ask for. In 2025, WIRED made public-records requests to the California Highway Patrol (CHP) and obtained training documents that showed that police agencies are very well aware of what data cars collect. The CHP trains its officers to know exactly what data is available from each automaker. Presumably, when they ask for that data, they get it.

So insurance companies and the cops have your data. Anyone else?

Yes: foreign spy agencies.

One U.S. defense contractor marketed vehicle telematics data explicitly for intelligence purposes, claiming it could “remotely geolocate vehicles in nearly every country except North Korea and Cuba on a near real-time basis“ with access to “over 15 billion vehicle locations around the world every month.” Sample maps in their pitch materials showed vehicle locations across Russia, Ukraine, and Turkey; one presumes cars in Canada, Europe, and the USA itself would be equally easy to track. U.S. intelligence agencies themselves have extensively purchased commercial location data from brokers.

Which means that if they want it, Chinese intelligence agencies can buy it commercially also.

True, the Protecting Americans’ Data from Foreign Adversaries Act and Executive Order 14117, both from 2024, ban sales of personally-identifiable sensitive data to China, Russia, North Korea, and Iran. But it’s not clear that this law is enforced, nor has anyone been prosecuted under it. Given how cavalier the automakers are with this data, and how skilled the PRC is at hacking, it’s reasonable to suppose they have this data anyway, laws or no. (Meanwhile, Canada lacks any such prohibition at all.)

I have said before that, when it comes to public surveillance, we live in the worst-case world. We have simultaneously too much surveillance for a society committed to civil liberties, and yet too little for one committed to public safety. This perversity applies especially to how we treat vehicle data.

If we cared about privacy, we’d insist that carmakers never sell on our commercial data, and never disclose it without a warrant, and certainly never disclose it without letting us know. And if we cared about public safety, we’d require that insurers and the police be given this data proactively, or take some other, systematic approach.

But we do neither.

Our Four Bad Options

So yes, Chinese EVs are spy cars, or can be. But so are American and Japanese and South Korean and European cars, at far larger scale, with their data available for commercial purchase. Given this, what are Canadians to do? We have four possible policy responses, none of which are attractive.

We can ban Chinese EVs, as protectionists across North America would like us to do. Will this make us safer? No, since China can buy vehicle data from the manufacturers whom we do permit to sell us cars. In return for being no safer, we lose access to cheap, state-of-the-art vehicles that help us make progress on our climate goals, while making the Canada–Chinese trading relationship harder, at a time when we need all the trading partners we can get. So this option is just security theatre… that reduces our standard of living… while accelerating climate change.

We can crack down on all vehicle surveillance. At first glance, this sounds promising. We could ban commercial data sales, require warrants for law-enforcement access, mandate transparency reports, establish independent oversight bodies, and require cybersecurity certification. All of which will help us prevent exploitation of our data, but only by non-Chinese manufacturers. Chinese carmakers are still subject to Beijing’s National Intelligence Law and can be compelled to share data, irrespective of what other countries’ law requires. So this option means that our data is still given away, but only to the PRC, which doesn’t seem like a win.

We could do both: ban Chinese vehicles while simultaneously fixing the Western vehicle data exploitation problem. This would actually work… but it has a sequencing problem. Banning Chinese EVs, or tariffing them so hard so as to implement a de facto ban, is easy; it took Canada only four months to implement the 100% tariff on Chinese EVs back in 2024. Comprehensive surveillance reform is a multi-year regulatory project that will take much longer: it would require new regulations with comment periods, large grants of time for industry to enter into compliance, and the creation of new enforcement mechanisms. And I am no expert on law-making; doing all this requires new legislation as well, which is notoriously time-consuming.

So ‘do both’ collapses into ‘ban now, reform later’. It delivers all the costs—reduced consumer choice, slower climate progress, and trade friction—immediately, while the security benefits arrive years hence, if ever. And even then, the arithmetic is unflattering: the Carney deal permits 49,000 Chinese EVs annually into a market of 1.5 to 2.0 million vehicles, virtually all of which already feed data to commercial brokers accessible to any buyer, Beijing included. We’d be paying real costs to close a small window while the front door remains ajar.

That leaves the fourth, least-bad option: don’t ban Chinese EVs.

Why is this an option? Not because Chinese surveillance concerns are imaginary; to the contrary, they’re quite real. It’s an option because, right now, the ban is pointless. In an environment where Beijing can purchase Canadian vehicle data commercially, from brokers fed by GM and Toyota and Volkswagen, the marginal risk from Chinese-manufactured vehicles is vanishingly small.

This is an argument about today rather than always. The case I’ve made depends entirely on the data-governance vacuum that allows commercial brokers to sell vehicle location and behaviour data to any buyer. Fill that vacuum by banning commercial-data sales, requiring warrants for law-enforcement access, and seriously enforcing those rules, and the calculation changes. In a well-governed environment, Chinese EVs genuinely do become a distinct risk, a tunnel through the PRC could access Canadian data and that no Canadian law can close. At that point, absent any plausible and verifiable commitment from the PRC to give up access, excluding the EVs would be sound policy.

So the honest position is this: let them in now, and get serious about governance, understanding that a properly-reformed data regime may well lead us back to the question of Chinese EVs on different terms. The goal isn’t to make peace with surveillance; it’s to address the large problem before the small one, without pretending the small one doesn’t exist.

Doug Ford is right that Chinese vehicles have the potential to spy on us. He is wrong that banning them, today, makes anyone safer. Build the governance framework first, and then we can determine what the risks are and how to address them.

“Why Are American Passenger Trains Slow?” starts with a puzzle: passenger trains in the USA were faster in the 1940s than they are today. The piece explores why this is the case. The answer isn’t that passenger rail went wrong; it’s that freight rail went spectacularly right.

Read it here.

If you take a plane in New York City, you must first perform a series of rituals. You set aside any liquids you possess, then you remove your shoes. You place your laptop and sometimes your phone into a plastic bin. You take off your belt and sometimes your shoes and place them into a plastic bin as well. You enter a machine that sees through your clothes. Only then may you board your plane.

If you take the subway in New York City, you swipe a card at a faregate and walk onto the train.

Put another way, airports have ‘checkpoint screening’: systematic inspection of every passenger and their belongings before boarding. Subways and rail stations do not.

Once we start thinking about this asymmetry, the stranger it seems. LaGuardia Airport hosted 32.8 million passengers in 2025, which averages roughly to 90,000 per day. Meanwhile, Penn Station processes more than 600,000 riders per day. Despite the fact that Penn Station has more than six times the number of passengers, no one verifies their identity, checks their bags, asks what liquids they are carrying, nor inspects their belts and footwear.

It’s not as if terrorist attacks on railways are unheard of. Madrid’s commuter trains were bombed in 2004, the London Underground in 2005, and Mumbai’s suburban railway in 2006, causing hundreds of deaths. And yet none of these now feature checkpoint screening. Indeed, the absence of checkpoints is regarded as a merit of rail and a demerit of air; there is no debate over just how many hours before one’s trip one should arrive at a rail station. Meanwhile, the USA’s Transportation Security Administration (TSA) employs over 56,000 people and spends more than $11 billion per year ensuring that no one boards an airplane with an unexamined shampoo bottle.

The asymmetry is so familiar that it barely registers as a choice. It feels like a law of nature: air travellers are screened but rail travellers are not. But it is a choice we’ve made, and the fact of that choice permits only two conclusions: either rail security is unconscionably negligent, or aviation security is irrationally excessive.

Our behaviour reveals which we actually believe.

A History of Violence

In the early days of commercial aviation there was no security at all. For a taste, watch Bullitt (1968) or Airport (1970), where it’s taken for granted that one can carry guns and bombs through terminals, onto the tarmac, or into aircraft without any mechanism for authorities to stop it, or even notice.

Those portrayals fall squarely within the so-called Golden Age of Hijacking, which began in 1961 when Antulio Ortiz, a passenger on a flight from Miami to Key West, threatened the pilot with a gun and demanded to be flown to Cuba. His was the first of 159 hijackings over the next ten years. After a 1972 incident where hijackers threatened to crash a plane into a nuclear reactor, in January 1973 the Federal Aviation Administration finally mandated that every passenger and their carry-on bag be inspected for weapons. Metal detectors appeared at airports that year.1

The 1973 system had a clear purpose: prevent hijackers from bringing weapons aboard. Metal detectors caught guns and knives and, in principle, explosives carried by passengers, while X-ray machines did the same for carry-on luggage. This physical system to deal with hijacking complemented the social system, which was to cooperate. Acting on the theory that hijackers wanted hostages, not corpses, the doctrine for crew and passengers alike was to comply rather than resist. Going along with demands bought time for negotiation, which generally ended with surrender, or with the hijackers escaping the plane and being apprehended elsewhere without loss of life to those aboard the airplane.

This system, of metal detectors, X-ray machines, and cooperative passengers, persisted largely unchanged for nearly three decades. It was imperfect, but it addressed a real problem, and it worked reasonably well.

This was the model that the 9/11 attackers exploited. They carried box cutters on board the plane, which were seen as tools rather than weapons and as such were permitted. Once aboard, they relied on passengers and crew behaving passively. Lack of resistance meant they were able to carry out their attacks on the World Trade Center and the Pentagon. Notably, Flight 93 did not carry out such an attack, because the passengers did resist. Having learned, via Airphone, that other captured flights were being deliberately crashed, the passengers on that plane understood their only chance of survival was to fight back. They attempted to overpower their captors, who destroyed the plane rather than lose control of it.

That shift on Flight 93, from compliance to resistance, has turned out to be a permanent psychological change. Richard Reid, the ‘shoe’ bomber, was subdued by passengers and crew in 2001. Umar Farouk Abdulmutallab, the ‘underwear’ bomber, was subdued by passengers and crew in 2009. In both cases, everyone else on the plane understood that the right move was to restrain the hijacker rather than submit to his demands. This means that one of the two vulnerabilities the 9/11 attackers used is now closed.

The other is closed as well. By April 2003, all commercial aircraft were required to feature hardened cockpit doors. The flight deck is now mechanically isolated from the main cabin, and will remain that way irrespective of what might happen there. To commandeer the aircraft, as the 9/11 terrorists did, now requires breaching that barrier. At a cost of $12,000 to $17,000 per door, plus annual extra fuel costs of $3,000, these doors make it more-or-less impossible for the cockpit to be captured, meaning that, in the future, any attacker’s bad acts will be confined to the cabin.

This means that the specific attack vector that made 9/11 catastrophic, using aircraft as guided missiles against ground targets, is now defended against by layers that don’t depend on checkpoint performance. Cockpit doors provide physical protection. Passengers provide active resistance. The weaponization-of-aircraft scenario requires defeating both.

Despite these changes, checkpoint screening has become ever more elaborate in the post-9/11 era. After Reid’s failed shoe bombing in 2001, passengers were required to remove their shoes for X-ray inspection. After a foiled liquid-explosives plot in 2006, liquids were restricted to containers of 100 millilitres or less. After Abdulmutallab’s failed underwear bombing in 2009, full-body scanners were deployed. Each measure was a reaction to a specific plot. Each remains in place decades later, despite none of these measures having ever demonstrably prevented a subsequent attack.

Alist, Denver Airport Security Lines, 2008, Flickr, licensed under CC BY-NC 2.0

Indeed the evidence that checkpoint screening catches any threat is weak. In 2015, the Department of Homeland Security red-teamed its own screening and found that screeners failed to detect threat items in 67 of 70 tests: a failure rate of 95%. We’re told things are better now, but I’m not aware of any subsequent published test, so there’s no public evidence to support the claim.

So if the 9/11-style vulnerability has been addressed by hardened cockpit doors and changed passenger psychology, what is the marginal security value of the vast post-9/11 checkpoint expansion? The 1973 system screened for guns and knives; perhaps that still serves a purpose. But the layers added since—shoe scanning, liquid restrictions, body scanners—what are they for?

Asymmetries Everywhere

One way to answer that question is to look at how we treat other modes of mass passenger transport.

We don’t screen rail passengers. The London Underground handles 150,000 entries per hour without checkpoint screening; such checkpoints are also absent at Penn Station in New York, Gare du Nord in Paris, and Union Station in Toronto. All of these are open systems. Passengers merely buy a ticket and board.

This isn’t because rail is safe from terrorism. There were bombings of the London Underground in 2005 that killed 52. There were bombings of Mumbai’s suburban railway in 2006 that killed 209. And there were bombings of Madrid’s commuter train in 2004 that killed 193 and injured over 2,000.

These were not minor incidents, but we didn’t retrofit rail stations with aviation-style checkpoints. There certainly were security responses: the UK introduced surge policing and expanded CCTV coverage, India introduced frisking on some lines, and Spain implemented some access control, though only for high-speed intercity services, not conventional rail. The pattern is consistent across every nation that has suffered a major rail attack: increased vigilance and policing, but no universal checkpoint screening.

Why not?

The answer seems to be ‘it would be difficult and expensive’. For the Underground to screen 150,000 passengers per hour would require an army of screeners and would transform stations into permanent queues. This would be unbearable for the Underground or any urban transit system, which depend on rapid throughput to be useful. The geometry is different too: airports have natural choke points, while rail stations often have dozens of entrances, again for maximization of throughput.

That may apply to urban rail, like subways, but less so to stations for intercity or commuter rail, which have less throughput and more choke points. Security stations there would be feasible, at least in regard to infrastructure; they would still have imposed immense delay. But the Madrid bombings killed nearly 200 people; if aviation logic applied, that should have been enough to make us insist on screening, no matter the cost.

But we didn’t insist, and that choice reveals a preference: we value convenience and throughput over marginal security gains, even after catastrophic attacks. When we weigh the trade-off directly—when the costs of checkpoint screening are immediate and visible—we conclude that checkpoint screening isn’t worth it.

But in aviation, we act as if the opposite is true. We maintain a regime whose costs are staggering (over $11 billion annually in direct federal spending in the USA, plus equivalent per-capita amounts in other nations, plus hundreds of millions of passenger-hours in queues globally) and whose marginal benefits are undemonstrated.

Let me pause to acknowledge a counter-argument: perhaps aviation checkpoint screening deters terrorists, who shift their attacks to softer targets like rail. The Madrid, London, and Mumbai bombings might be evidence of successful deterrence, with subsequent displacement. But if checkpoint screening merely displaces attacks from aviation to rail, the net security benefit is zero; we’ve spent billions and wasted millions of hours to move the threat from one set of passengers to another… implicitly, a set of passengers we think less deserving of our protection.

And the fundamental point remains: whether those rail attacks were sui generis or displaced from harder targets, they killed hundreds, yet we didn’t impose checkpoints. We revealed our preference.

The Security Ratchet

If that revealed preference is for the rail model of security, why doesn’t aviation security move in that direction?

The reason for the air vs. rail distinction is a separate asymmetry among political incentives. An official who maintains excessive security incurs no blame for doing things the way they have always been done. Passengers may grumble, but passengers always grumble. Conversely, an official who loosens security would incur heavy blame in the event of an attack, regardless of whether the loosened measures would have prevented it.

Put another way, any official who changes the system must first incur costs of time, attention, and effort. If things go well, they receive no benefit in return, because no one notices; but if things go poorly, the disbenefit they receive would be massive.2

This incentive structure produces a ratchet. Security measures accumulate, but almost never recede. After the shoe-bomb plot, we started removing shoes. After the liquid explosives plot, we restricted liquids.3 After the underwear-bomb plot, we deployed full-body scanners. Each measure responds to a specific plot, but none is ever removed, at least not without a technological excuse. The only significant rollback in two decades came in July 2025, when the TSA eliminated the shoe-removal requirement… but only because new scanning technology could inspect footwear while still on people’s feet, not because anyone concluded the requirement was unnecessary.4

There’s a strong tell that the system understands that checkpoint screening is theatre, namely TSA PreCheck.

PreCheck allows enrolled passengers—currently over 20 million—to skip shoe removal, keep laptops in bags, and pass through expedited lanes. The price is $85 and a background check. The implicit logic is that pre-vetted travellers present lower risk and can therefore face reduced screening burdens. (Outside the USA, other nations have their own versions; Canada has NEXUS, for instance.)

The concession here is significant. If expedited screening is safe enough for 20 million PreCheck passengers, what is the marginal security value of the full screening regime for everyone else? The programme’s eligibility rules would have structurally excluded the 9/11 hijackers, who were neither U.S. citizens nor permanent residents. This might seem to justify PreCheck as a different security model—identity verification plus light screening—rather than simply “less screening.” But the identity verification that PreCheck provides (criminal background checks, watchlist screening) would not have flagged them: they had no disqualifying records and were not on any watchlists. The citizenship bar does real work, but it’s a blunt instrument that excludes millions of innocuous foreign travellers while doing nothing to catch citizens who radicalize.

What PreCheck actually demonstrates is that the TSA treats post-9/11 additions to checkpoint screening—the shoe removal, the liquid restrictions, the laptop divestiture—as dispensable for a large population. If those rituals were essential to security, waiving them for anyone would be unconscionable.5

Confusions about security distorts how we think about transport investment.

The case for high-speed rail rests partly on convenience: city-centre departures, walk-on boarding, no two-hour pre-flight buffer. But that convenience is a function of what we don’t require. If we screened rail passengers like air passengers, rail’s travel-time advantage would shrink. The comparison is invidious in a way that favours rail, but rail’s advantage in this case depends on a policy choice.

The implication runs both directions. One underexplored response to the question ‘how do we improve domestic travel?’ is not ‘build high-speed rail’ but ‘make flying less miserable by relaxing aviation security to evidence-based levels’. This is politically challenging, but the fact that it would be difficult to implement doesn’t make it wrong.

To be clear: I think that the rail model—layered security based on intelligence, visible policing, and passenger awareness rather than universal checkpoint screening—is the right baseline for mass transit in free societies. Aviation should move toward this model, not away from it. I am not arguing that we should be able to walk onto planes the way we walk onto trains, though that position is more defensible than it might first appear. But I am arguing that the post-9/11 checkpoint expansion has not demonstrated its value, and that aviation security should return to something like the 1973 baseline.

The costs of the current regime are enormous: fiscal (billions of dollars), temporal (hundreds of millions of passenger-hours), and philosophical (routine searches that we don’t permit in other contexts). The benefits are undemonstrated. While many people intuitively feel that checkpoint screening offers important protection, that intuition rests on unawareness that hardened doors and a shift in passenger psychology means airplanes can’t be weaponized now. Once you understand that the 9/11 attack vector is blocked by those layers, the case for the post-9/11 checkpoint expansion collapses.6

To return to where we began, LaGuardia’s 90,000 daily passengers submit to rituals but Penn Station’s 600,000 do not. We tolerate open rail because we’ve made a judgment—implicit, unspoken, but revealed through action—that checkpoint screening isn’t worth the cost. We maintain fortress aviation because we’re trapped in an equilibrium we cannot escape, not because we’ve judged differently.

This asymmetry is a policy choice we’ve made, and one we could unmake, if we could find the courage to do so.

Respect to and for feedback on earlier drafts.

So what was the admission?

Waymo’s CSO, Mauricio Peña, under questioning by Senator Ed Markey, stated that some of Waymo’s fleet-response operators, i.e., humans who provide guidance to the firm’s automated vehicles (AVs) when they encounter ambiguous situations… are located in the Philippines. The senator called this “completely unacceptable” as both a “safety issue” and a “national security risk”, arguing that having people thousands of miles away influencing vehicles on American roads introduces unacceptable latency and cybersecurity vulnerabilities. The implied narrative here is juicy: Waymo admits their cars aren’t actually self-driving! Jobs are being shipped overseas! Foreign workers are controlling American cars!

I understand why, if one is not paying attention, this might feel like a scandal. But for those paying attention, the fact of remote assistance is not new, and isn’t particularly interesting; Waymo wrote a whole press release about this back in 2024. Focusing on the fact of remote assistance misses the important things that Peña said.

It is the view of Changing Lanes that geographic location matters: cross-border teleoperation poses difficulty in accountability and oversight. But to attack remote assistance is an error. It’s not a scandal that Waymo uses it, but it is a problem that Waymo, and the industry as a whole, prefer not to foreground their use of it, and that our regulators have been slow to treat it as public infrastructure that requires coordinated planning.

The Basics of Remote Assistance

Let’s begin by clarifying what teleoperation or ‘fleet response’ actually is.

Modern automated driving systems (ADS) construct a model of the world around them through sensor and mapping data, and then apply decision-making algorithms to determine how to proceed. For example, when approaching an intersection for a left turn, the ADS identifies the traffic signal, tracks oncoming vehicles, detects pedestrians in the crosswalk, and executes the turn when conditions permit. At a stop sign, it recognizes the sign, brings the AV to a complete stop, yields to cross traffic and pedestrians, and then proceeds when safe. When detecting an obstacle ahead—a stopped vehicle, a pedestrian stepping into the road, debris—it slows or stops as appropriate.

These are straightforward road situations, but an ADS will inevitably encounter situations with which it is unfamiliar: scenarios the engineers didn’t explicitly program for, or edge cases absent from (or insufficiently represented in) their training data. These might include unusual blockades around construction sites, emergency vehicles with non-standard light patterns, ambiguous signage, or an environment featuring complex interactions between other vehicles, pedestrians, and cyclists.

In these cases, a well-designed system will (or rather should) recognize the limits of its own competence. Rather than guessing or freezing indefinitely, it seeks human judgment. The AV enters a safe state—typically slowing down and moving to the side of the travel lane if possible—and transmits its sensor data to a remote operator. The operator reviews the situation and provides guidance, which the ADS then evaluates against its own perceptions before deciding whether to accept it. If it does, then it acts appropriately.

Some wags refer to this as ‘phoning a friend’, one of the lifelines on the game show Who Wants to be a Millionaire? I like this analogy, because the same dynamic applies. Faced with uncertainty, the ADS gains more information from a source it deems trustworthy. Given that information, it is better equipped to make a decision, but everyone, including the ADS, understands that it is ultimately responsible for its actions. The ADS therefore assesses the information it’s given from that perspective.

Based on this, one can see why Waymo prefers the term ‘fleet response’; the remote human doesn’t actually ‘operate’, i.e., drive, the AV. These humans instead provide high-level guidance or authorization to the ADS, which executes the driving task itself.

Consider a construction zone where traffic cones narrow two lanes into one, but the cone placement is irregular: some cones are offset and others have been knocked over, making it ambiguous which lane or lanes are closed. A human driver would slow down and assess the situation, taking note of the overall cone pattern and the behaviour of vehicles ahead, and then proceed. Given the same situation, though, an ADS might recognize that its confidence in its assessment of which lane is safely passable was low.

The ADS would therefore contact fleet response. The remote operator would review the data from the vehicle’s feeds and provide guidance: “The right lane is closed but the left lane is open and available to use”. The ADS compares this guidance to its own perception of the environment: can it see the left lane clearly? Are there obstacles in that lane? Does the guidance align with what its sensors detect? If the guidance matches the ADS’ understanding and falls within safety parameters, the ADS accepts it and proceeds. If there’s a mismatch, the ADS can request additional clarification or refuse the guidance and remain in its safe state.

This doesn’t resemble remote driving as much as air traffic control. The responders help Waymo ADS navigate ambiguous situations, but they don’t steer the cars themselves.

None of this is new. Waymo published a detailed blog post about their fleet response system in May 2024, describing how it works. “Fleet response specialists” provide guidance on edge cases, with the ADS in control throughout. Operators view camera feeds and sensor data remotely, then provide context that the ADS evaluates before deciding whether to accept and act on. Waymo emphasized this isn’t “remote driving” but rather “high-level assistance”. Zoox uses a similar model.

Conversely, Chinese AV companies like Baidu Apollo Go and WeRide use actual teleoperation, i.e., remote operators who can drive the vehicles, and Tesla’s Austin robotaxi pilot uses teleoperators to remotely monitor the fleet and intervene if vehicles get stuck. The details are opaque, though it’s been noted that Tesla posted job listings for roles to “access and control” vehicles remotely.

So we shouldn’t be surprised. We’ve known for years that remote operators help Waymo ADS navigate edge cases and provide oversight in ambiguous situations, and that this is actually tame compared to the approaches taken by the firm’s rivals. The breathless coverage suggests Waymo got caught admitting something they’d hidden, but they didn’t; people just weren’t paying attention.

More importantly, remote assistance shouldn’t be controversial. It’s not a shortcoming or evidence that the technology “doesn’t work”. It’s necessary infrastructure, analogous to support systems we already accept for human-driven vehicles. Taxi drivers radio their dispatch when they encounter ambiguous situations, delivery drivers call their companies when they can’t determine whether they’re permitted to use a loading zone, and commercial truckers communicate with their fleet managers about severe weather that might require route changes.

We don’t view these communications as evidence that human drivers are inadequate. We instead recognize them as sensible coordination that improves outcomes, and we should want AVs to do the same, and request assistance when they encounter uncertainty. The alternative is worse: AVs that either freeze completely when facing ambiguous scenarios, creating traffic disruptions and stranding passengers, or AVs that proceed with unwarranted confidence, making dangerous decisions based on insufficient information. An ADS that recognizes the limits of its understanding and seeks human judgment is demonstrating good system design, not bad.

Refusing to deploy automated driving until we have a technology that can handle every edge case would mean waiting forever. The space of possible driving scenarios is effectively infinite. No amount of pre-deployment testing can anticipate every scenario, so it would be foolish to wait, especially as real-world deployment is precisely how the technology improves. Each edge case an ADS encounters and successfully resolves becomes training data that will mean future interventions aren’t necessary. And as I will never tire of saying, delay here is not costless. Waiting means accepting roughly 40,000 traffic deaths annually in the U.S. from human drivers that ADS at scale could prevent.

So if remote assistance is both normal and desirable, does Waymo’s Congressional testimony matter?

I think it does, because it revealed genuine policy gaps that need repair.

Read more

It was the morning, and many parents were dropping off their children, meaning that the situation was as rich with incident as one might expect: children on the sidewalks, a crossing guard, and lots of double-parked vehicles. As the Waymo proceeded through the area, a child ran into the street from behind a parked SUV toward the school. The Waymo slowed its speed and while it did hit the child, they sustained only minor injuries and were able to walk away.

In the wake of this incident, I saw lots of people on X come to the company’s defense, making the argument that had a human driver been operating the vehicle, the accident still would have happened, but lacking the Waymo’s digital vision and ‘reflexes’, the collision would have taken place at a higher speed. In that case, the child would be gravely injured or likely dead, as opposed to being able to walk away from the incident.

Waymo made a similar claim, arguing that in fact this incident demonstrated “the material safety benefit of the Waymo Driver”.

Let’s stipulate two things immediately: the claim is almost certainly true; and it’s almost certainly a bad communications strategy. Given that Waymo’s most-important selling point to the public is that driving automation will reduce the frequency of road incidents, arguing that the Waymo Driver hurts people less is actually an argument for the firm’s opponents.

But leave it aside. As an advocate for driving automation generally rather than any particular firm, I have no stake in Waymo’s fortunes. But I do have an interest in this technology spreading widely, which is why, to me, the interesting question here is why Waymo and the firm’s supporters expected this argument to carry the day.

The answer is a disagreement about benchmarks. The driving-automation sector, its critics, and its regulators are measuring automated-vehicle (AV) performance against different standards.

I think there is a right benchmark, but it’s harder to defend than its proponents admit, and harder to dismiss than its opponents would like.

Choosing a Benchmark

Intuitively, when we think about an AV company’s safety record, we compare their automated-driving systems (ADS) to the general driving population. The denominator includes every driver on the road: not only the attentive and calm, but also the sober and drunk, attentive and texting, the seventeen-year-old with a learner’s permit and the eighty-year-old who should have stopped driving years ago. Against that population as a whole, a well-engineered ADS looks very good. It doesn’t drink, nor text, nor doze off, nor even fiddle with the car’s audio system. That means that an ADS is indeed better than the average, which is a valuable improvement we should want to see implemented everywhere. Impaired and distracted driving kills tens of thousands of people a year in North America, and eliminating those deaths would be a historic achievement.

But is this the right standard?

An ADS that outperforms the average is outperforming a benchmark that includes the worst. And indeed Waymo takes this into account by also employing what it calls the NIEON standard: Non-Impaired, Eyes ON. This standard refers to a hypothetical human driver who is always attentive and never tired. If the Waymo Driver (its ADS) can outperform this standard, the safety value of the service is established even more strongly.

Even stronger standards are possible. , a researcher who has worked on self-driving car safety for decades, argues that the right comparison is to a “careful and competent” human driver who adjusts behaviour based on context. Under this benchmark, the relevant question about the Santa Monica crash isn’t reaction time. It’s whether the vehicle was driving with a level of caution appropriate for a chaotic school drop-off involving double-parked vehicles, a crossing guard, and visible children. A careful human driver approaching that scene slows down before any child runs into the road; they slow down because they expect that just such a thing might happen. If the Waymo didn’t, the crash isn’t “unavoidable.” It’s a predictable consequence of insufficient contextual awareness.

Koopman is pointing to something very important: we can judge an ADS’ performance against several kinds of human driver, and our choice of comparison class, applied to the same incident, produces different conclusions. On one view, the Waymo vehicle performed well: it detected the child and braked faster than a typical human would have managed given the same perception-response constraints. Call this the ‘Average’ standard. On another view, the same crash is a failure: a prudent driver would have been going slower before the child appeared. Call this the ‘Professional’ standard.

The latter is what we already apply to people who drive for a living: bus operators, truckers, or taxi drivers. AVs are, or aim to be, commercial vehicles carrying paying passengers on public roads. In such cases, we have appropriately higher expectations. A bus driver who hits a student near a school is judged much more severely than the average motorist, precisely because we expect them to know and perform better. We don’t expect perfection, but competence exercised with care: evidence that the operator understood the environment, acted appropriately within it, exercising reasonable judgment about potential risks and trade-offs. Indeed, this capability is precisely what the term ‘professional’ means.

This argument, that we should hold ADS to the Professional standard, is one I have a good deal of sympathy for.

But it’s also riddled with problems.

Against the Professional-Driver Standard

The first problem is that the Professional standard relies on the concept of reasonableness without explaining how that standard would work for a machine. The critique of the Average standard is that it’s not enough for an ADS to simply follow the rules-as-written without paying attention to larger context and behaving ‘reasonably’ given that context. That critique leaves a basic question unanswered: how is that kind of reasonableness supposed to be defined and applied before something goes wrong?

Professional drivers are not governed by a single, clear rule for how cautious they must be in different situations. In practice, two different professional drivers might behave quite differently around schools, parked cars, or places where pedestrians might be hidden, and we don’t regard that difference as evidence that one or the other is in the wrong. In such cases, we only look for wrongdoing after the fact, by reconstructing the scene, appealing to shared norms, and judging which risks a driver should have anticipated. This is the approach we use in tort law for humans.

For machines, though, reasonableness must be turned into guidance that shapes behavior in advance. If we skip that step, the standard is supplied only after an incident, through judgments about what the ADS should have done… meaning that accountability shifts from clear expectations to after-the-fact storytelling. We would regard that approach as unfair if applied to human conduct, and it’s equally unfair if applied to a machine.

It’s unfair because there is an obvious failure mode here. If reasonable-behaviour-in-context is defined retrospectively, if every crash becomes evidence that the system should have been more careful because ‘look what happened’, then no ADS can ever be vindicated in the case of an incident. Every incident will always be the ADS’ fault, because it will always be possible to say that the incident wouldn’t have happened had the ADS been better.

This problem has some merit. For a machine, reasonableness does need to be translated into guidance that shapes behavior before something goes wrong, and proponents of the Professional standard weaken their case by treating that translation as easy or obvious. There is real definitional work to be done, and it is not obvious that has been.

Applied to the Santa Monica crash specifically, however, any concern with after-the-fact storytelling is hard to entertain. This was a school zone during drop-off hours, with double-parked vehicles, a crossing guard, and children visible on both sides of the street. These are not contextual cues that become legible only in retrospect. They are cues that every experienced driver recognizes. Saying the ADS should have been more cautious here is not merely hindsight bias.

It’s also something we have no difficulty saying about the exercise of judgment in other safety-critical fields, like aviation or medicine. Pilots and doctors both are granted discretion, but are also held to a standard developed through professional consensus (and in some instances, case law). These fields have not collapsed incident review into arbitrary hindsight. That in turn means that automated driving can benefit from an equivalent framework, once we develop one.

The second problem is the downstream effects of the Professional standard. Suppose regulators require an ADS operator to explain, after every serious incident, why the system chose to drive at a given speed in a given context. Those explanations will rapidly harden into informal expectations, which will become de facto rules.

And those rules will become liability triggers.

Anyone who has watched how environmental impact assessments or accessibility standards evolve in practice knows this pattern, and knows where it leads. The final stage of the process will be ADS that avoids complex environments entirely because, though they are safe to enter in aggregate, any individual incident within them will be narratively indefensible. (Not to mention that opportunistic humans will note that abundance of caution and exploit it.) The result is a fleet of robotaxis that only operates on wide, empty boulevards after midnight, which is both legally safe and utterly useless. That’s an extreme I don’t think we would reach—the slope isn’t that slippery—but I’d prefer not to slide any distance in that direction.

This problem is harder to dismiss. We know from our experience of the housing, energy, and environmental-protection sectors over the past five decades that regulatory regimes do calcify, incident explanations do become expectations, and those expectations do become liability triggers. To avoid this risk, we need regulators to state plainly that not every incident in a complex environment implies fault.

The third problem is the unfair asymmetry here. We don’t require human professional drivers to produce a machine-legible account of their risk posture after every incident, because they can’t provide one. Instead, we judge them by plausibility and norms, not by internal-state inspection. Requiring AVs to demonstrate, algorithmically, that they recognized and responded to every relevant contextual cue is a distinct and more demanding obligation, imposed on systems that can be inspected precisely because they can be. The ability to look inside the machine creates the expectation that we must, which creates a standard no human driver, even a professional one, has ever been held to.

This problem, however, has much less force. The only reason that we don’t, after a road incident, demand detailed cognitive-state accounting from human drivers is because we cannot. If we could, we would, and we’d be right to do so. The ‘inspectability’ of ADS does not create an unfair burden on ADS, but reveals an unfair burden that regulators have always had to carry, for lack of any way to shed it. The important distinction is between post-incident inspection and a standing requirement to justify every decision in advance. Regulators should be explicit that they insist on the former, not the latter.

The fourth problem is a concrete human cost to raising the benchmark, and it is the problem I find hardest to dismiss. Every kilometer not driven by an AV that is demonstrably safer than the average human is a kilometer driven by a human who might be impaired, distracted, or exhausted. Elevating the standard from ‘better than average’ to ‘comparable to a professional driver’ is not a neutral move. It is a choice to delay deployment in exchange for a higher standard of care around vulnerable road users. That trade-off may be worth making, but it’s one that critics of driving automation treat as costless, and it isn’t.

“Person Holding White Smartphone Inside Vehicle”, Roman Pohorecki

This argument risks proving too much. It is true that every kilometer not driven by an ADS that is safer than the average human is a kilometer driven by a human who may be impaired, distracted, or exhausted. But an ADS that meets the Professional standard also displaces drunk and distracted drivers; the only difference is that it further exercises heightened caution in places where that is necessary. The foregone-safety argument would justify deploying any system that clears the Average bar, regardless of how it behaves around children (or other vulnerable road users). Few people would accept that trade-off once stated plainly; no regulator that I know of would.

This leaves us in an uncomfortable place. The Average benchmark is inadequate, but the Professional benchmark is a work in progress. We must resist the temptation to resolve the tension by choosing one and ignoring the other.

What Survives

I can see a workable path forward.

Firstly, Waymo and others must meet bright-line requirements for the most legible high-risk contexts: these might include school zones during operating hours, active construction sites, or areas with foreseeable heavy pedestrian activity, for example streets outside a sports arena right before and after a match. In environments like these, the cues are obvious and can be specified in advance. Requiring heightened caution in these settings is no different than requiring an ADS to obey speed limits.

Secondly, for everything else, we should insist our regulators not only impose a full-disclosure regime, but commit explicitly in advance that bad outcomes do not automatically imply fault. The point of disclosure is not to justify punishment, but to make it possible to determine whether an ADS acted appropriately despite an incident. The goal is to ensure that the post-incident conversation includes what the ADS perceived and how it responded, not merely how fast it braked.

A child was struck by a Waymo near a school on a Friday morning during drop-off, and the company’s response was that it hit her less hard than someone else might have. So was this a success, or a failure? The answer is: it depends on the standard we apply.

That answer in turn points to a problem that will not resolve itself. I hope the industry and regulators work together to build a defensible Professional standard, which specifies where heightened caution is required and how it should be demonstrated. I fear that, if they do not, regulators will determine that standard themselves after a more serious crash, in worse circumstances and in a less-forgiving mood.

Respect to for feedback on earlier drafts.

Air travel today is very cheap. Back in 1980, a round trip ticket between Chicago and Dallas cost around $185, so $725 in 2026 dollars.1 A similar ticket today would be about $480, meaning the flight is 33% cheaper. Sure, your seat is smaller, and the overhead bins are all full, and the plane is more crowded… but if you want a better experience, you can still pay extra for it. What most people want is low base prices, and we’ve got them, so stop complaining already.

Another story goes like this: in the last several decades, airlines have merged into a cozy oligopoly (‘because capitalism’) and killed competition. As a result, they now extract rents from captive passengers. Four carriers control 74% of the domestic market, and use their market power to shrink seats, charge for bags, and oversell flights because consumers have nowhere else to go. The problem is consolidation and monopoly, and the solution is antitrust.

At first glance, both stories make sense. Consumer price sensitivity is certainly real, and consolidation has given airlines market power they’d never had before. Where they both fall short is the question of timing.

Airlines in the USA have had the legal freedom to charge for bags, unbundle fares, and cram seats tighter since deregulation back in 1978. Given that, why did bag fees appear in 2008 rather than 1988? Why did the ‘Basic Economy’ fare class emerge after 2012? Why did plane crowding (the percentage of occupied seats aboard a flight, which industry analysts call load factor) jump from 72% to 84% this century, rather than before?

‘Consumers want cheap fares’ and ‘airlines are greedy’ are both timeless explanations for time-specific changes. That means that something else is going on.

We know what that something else is. The features that make modern air travel so frustrating are scar tissue from three near-death experiences, all post-dating the year 2000: the business-travel shock after the terrorist attacks of 9/11, the fuel shock of the 2008 global financial crisis, and the travel-demand shock of the COVID pandemic. Each of these shocks left scar tissue on the sector. A market dynamic that I call the Iron Law of Air Travel ensures this scar tissue never heals; once a cost-cutting measure exists, consumer behaviour locks it in, because the airline that tries to compete on quality rather than price loses the fare comparison and bleeds customers.

A Tale of Three Crises

Once upon a time, namely in the USA before 1978, the Civil Aeronautics Board set fares, controlled routes, and decided which airlines could fly where. Airlines couldn’t compete on price, so they competed on service: meals, legroom, lounges, and more. This is the Golden Age of air travel; you can catch sight of it in films of the era. (This portion of the trailer for Airport ’75 gives a brief glimpse.) Fares were high enough that flying was mostly for the affluent and for business travellers whose companies covered their costs.

Then the Carter Administration deregulated the sector. Airlines could now set their own fares and routes, and new entrants flooded the market. The result was fierce price competition that drove fares down, changing air travel from an elite activity to a plebeian one. (Airplane!, a parody of earlier films like Airport ’75 but made in 1980, firmly in the deregulated era, depends in part on the audience remembering a world which had definitively ceased to exist.)

Legacy carriers, faced with competition from upstarts, responded with exquisite price discrimination, extracting enormous premiums from business travellers through mechanisms like Saturday-night-stay requirements and last-minute booking penalties. Leisure travellers might be able to delay a trip but business travellers often couldn’t, and so had no choice but to pay exorbitant fees for sudden bookings. Similarly, business travellers had no reason nor desire to stay in another city over a weekend, so business travellers also had no choice but to pay a hefty premium to forego the (entirely arbitrary) Saturday-night stay requirement. Economy passengers thus enjoyed a subsidy: they paid reasonable fares because business travellers paid five-to-ten times more for the same trip.

The shock of airplane-based terrorism on September 11, 2001 collapsed demand for air travel for several years. By 2004 demand was back… but not among business travellers. By 2004, the Internet had gone mainstream, meaning that it was easier to book flights, and to compare prices, than it had been in the preceding era of travel agencies.

This technological change played out in several ways. Corporate booking platforms gave companies visibility and control over employee travel. Airlines responded by negotiating contracts with large firms, but that squeezed rates. Most importantly, the Saturday-night-stay requirement became less effective as low-cost airlines like Southwest defected from that consensus, forcing traditional airlines to defect too. The subsidy that had made economy-fare pricing viable didn’t slowly erode, but collapsed suddenly.

With business travellers no longer cross-subsidizing economy travellers, airlines needed another way to make cheap seats profitable. Spirit Airlines pioneered the answer in 2006–07, which was ‘unbundling’. Separate the bag fee, seat selection, and other services from the base fare. Sell the bare minimum at a rock-bottom price, then charge extra for anything beyond it. This is the origin of the ‘you see $89, you pay $200’ phenomenon.

CRJ-700 interior, photo by Cory W. Watts, via Flickr (CC BY).

After the business-travel shock came the fuel shock. Fuel prices rose a massive 338% between 2003 and 2008 as an industrializing China absorbed the world’s spare oil capacity even as big institutional investors moved into commodities futures. It’s true that prices went down later in 2008 courtesy of the global financial crisis… but that crisis also depressed demand for air travel. The combination of the fuel shock and the crisis was devastating for commercial airlines. Thirteen went bankrupt in 2008 alone; more than fifty U.S. airlines filed bankruptcy between 2001 and 2013. Northwest, Continental, and US Airways all went away, leaving four major carriers standing: Delta, American, United, and Southwest.

The surviving carriers learned to match supply to demand. Empty seats are money lost, since the plane must fly even if that seat is empty, and that trip can never be sold again. So airlines strove to ensure that as few seats as possible went empty; load factors rose from 72% in 2000 to 84% by 2019.

The travel-demand shock came in 2020 with the Covid-19 pandemic. U.S. corporate travel budgets declined by more than 90% as travel everywhere ceased and companies shifted en masse to virtual meetings. With the pandemic’s end, business travel returned, but it has plateaued at 70%-to-80% of 2019 levels, which airline executives describe as permanent. Remote work taught companies they could operate successfully even if their employees took fewer business trips.

Each crisis forced airlines to cut costs and squeeze more revenue from fewer passengers. But why did these adaptations persist even after the crises passed? Why didn’t fares re-bundle, seats widen, and planes empty out once fuel prices fell and demand returned?

The answer lies in what I call the Iron Law of Air Travel.

The Iron Law

The Iron Law of Air Travel says that consumers always choose the lowest price. Given an array of options—leg room, lateral room, early boarding, more baggage allowance, more—customers will ignore all of them and take instead whatever is cheapest. There is lots of evidence to confirm this rule, but my favourite piece is that Expedia, the airfare marketplace, knows their customers and by default sorts flight results by price starting with the lowest.

There are certainly counter-examples. Just last year, my wife and I flew to Europe for our belated honeymoon, and paid extra to enjoy ‘premium economy’ in each direction, which offered a few amenities, the most important of which to us was extra legroom. We are not alone: premium economy sells on long-haul routes. It’s also the case that travelers will reliably pay more to escape a stopover or transfer; the pull of the one-seat ride is felt in air travel as much as in other modes. And of course some travellers are wealthy, and as such are price-insensitive and will reliably upgrade.

But these groups are a minority, insufficient to shift the equilibrium. The Iron Law describes the centre of mass, not every particle. And the centre of mass is that consumers expect airlines to compete on price, not on other metrics. What that means is that even as demand for air travel rises, the old world won’t come back.

The Iron Law means that airlines and consumers plan around the market they have, even if everyone involved finds it suboptimal, and no one can unilaterally defect.

Airlines might prefer to compete on quality rather than racing to the bottom on price, and passengers might prefer better service. Unfortunately, the airline that tries to offer ‘better but more expensive’ loses the price comparison and bleeds customers. Airlines that do compete on price buy aircraft accordingly, meaning their infrastructure locks in the number of passengers who can get better service, and the market assigns those prices at a high point, too high for most consumers to justify paying for what they will get. No individual actor can break the pattern without being punished.

Imagine you wanted to start an airline that broke the mould: ‘Honest Air’. Honest Air would offer bundled fares: checked bag, seat selection, Wi-fi included, the works. Further, Honest Air would build its business plans to aim at a 75% load factor, so the middle seat would usually be empty. In return, Honest Air would charge double what its competitors do, at $400 for a route where Basic Economy goes for $200. Your business plan looks reasonable at first glance. Passengers say they hate the current experience, and you’re offering something better, so they’ll pay for it. Free markets to the rescue!

But the Iron Law reigns. Your customers will search Expedia, see your $400 fare next to a $200 fare, and book the $200 fare, and crack wise about how “no matter how much each class pays, we all arrive at the destination at the same time”. Then they will complain on X about bag fees and cramped seats without changing their behaviour. Honest Air’s planes fly half-empty, it burns through its capital, and within eighteen months, it’s out of business (or pivoting to the same model as everyone else).

I’m not just spinning a tale: in 2007–08, three airlines—Eos, MAXjet, and Silverjet—tried exactly this model, and suffered this fate. The three airlines ran only premium service; bundled their fares; and offered business-class-only or premium-heavy service. And all three were dead by late 2008.

The fuel spike finished them, but their underlying problem was the Iron Law: as much as their market surveys suggested people would pay more for better service, people booked cheaper seats with competitors instead.2

Changing the Game

What would change my mind about the Iron Law? Evidence that a significant segment of consumers reliably pays premium for better economy products when cheaper alternatives exist on the same route. Not premium economy on long-haul; that’s a different market. I mean domestic short-haul where premium economy competes with regular economy. If such a segment exists at scale, the Iron Law is weaker than I think, and competitive pressure could overcome the status quo, given time. I’d also update if a new entrant succeeded with a bundled model, though the history of such attempts is not encouraging.

So if my diagnosis of the problem is correct, what can be done about it?

Changing Lanes readers will be able to guess that my preferred solution to this problem is policy change.3 The most promising intervention I can see would be to target the interface, and require all airline tickets to be advertised at a price that includes one checked bag and a selected seat. Perhaps later customers could choose to forego amenities for a price break, but the default display would be the bundled version, which would force comparison shopping to a like-to-like basis.

Airlines would game this, of course, defining the standard experience downward (excluding inflight Wi-fi, for instance, and then selling it separately to customers who want it). They’d still compete. But that’s fine by me; my goal isn’t to eliminate price competition, but to ensure consumers can compare what they’re buying. Right now, the Iron Law operates partly through interface design: consumers click the lowest number because the interface shows them the lowest number.4 Mandatory bundled-price display would change what the interface shows. It redirects consumer psychology toward total value rather than base fare alone.

The USA DOT has moved in this direction with transparency rules requiring disclosure of baggage and change fees alongside fares. But that’s a half-measure; disclosure isn’t the same as bundling. Passengers still see the low base fare first and the fees in fine print. Full-price display would go further, and I think it’s worth trying. The downside is modest, and the upside could be significant: competition shifts toward schedule, reliability, and service quality rather than who can hide the most fees. None of this would be easy, and it doesn’t guarantee a better equilibrium. But understanding how we got here at least clarifies which interventions might work.

Air travel is unpleasant, and fixing it will require changing the game, not just wishing the players would make different moves.

Respect to for feedback on earlier drafts.

“Seeing Like a Sedan” examines the great philosophical divide in driving automation: Waymo’s sensor-fusion approach versus Tesla’s commitment to cameras-only. I trace the technical lineage from the DARPA Grand Challenges through to today’s Austin streetscape, where both companies now operate robotaxis perceiving the world very differently. The piece digs into the safety data, the limitations that keep surfacing in real-world conditions, and the quiet convergence now underway.

What’s at stake? Whether we want robotaxis that are as safe as human drivers or those that are better.

Read it here.

2023 was the worst year in a generation for the bike industry. There hadn’t been a boom and bust like this since the ‘70s.

2024 was worse.

Kevin McLaughlin

The bicycle industry is in crisis, but you probably haven’t heard about it. But ask the insiders, and you’ll hear about bike shops closing or downsizing, retailers absorbing losses they cannot sustain, and a glut of inventory that won’t move. The underlying condition—a COVID-induced boom-and-bust cycle now stretching into its fourth year—remains largely invisible to consumers and policymakers alike.

This matters beyond the fortunes of bike retailers. E-bikes represent one of the most promising tools for urban mobility: a mode of transport that can replace car trips, reduce emissions, and make cycling accessible to people who would never consider a traditional bike. But the industry that sells and services these machines is struggling to survive, and the policy environment that could support their adoption barely exists.

To understand what’s happening, and what it means for the future of cycling in cities, I spoke with Kevin McLaughlin. Kevin is a Canadian pioneer and leader in healthy-city ventures, spanning shared mobility, sustainable transport, and micromobility, for several decades. He is currently the founder and CEO of Zygg, a Canadian e-bike subscription and leasing service, and has previously founded or co-founded Toronto’s AutoShare car-sharing service, the Modo car coop in Vancouver, and Evergreen Canada. He has commuted by bicycle in downtown Toronto for more than 30 years.

Kevin McLaughlin

We spoke on 16 December 2025. This transcript has been edited for length and clarity.

Year Four of the Crisis

Andrew Miller: What is the crisis in the cycling industry?

Kevin McLaughlin: We’re in year three, at least, of the crisis, [which is] overproduction and oversupply created from the COVID boom. Capacity that had been serving Asian markets suddenly got turned on for North America. All of a sudden there were too many bikes.

The bike business has these intrinsic parent-child relationships between manufacturers and retailers. When we started in 2019, you had to place all your orders in October. The manufacturers push the bikes out to the stores. And then during spring and summer retail stores have their chance to sell, with end-of-season sales in the fall.

2022… that boat stuck in the Suez Canal is a good metaphor. At that point, demand was still COVID-high, massive. All these bikes were getting ready, but a lot of them showed up late that year. Instead of arriving in April, bikes were coming in August, September, October… you missed your season. You’d made these orders and were stuck with inventory. That rolled into 2023, oversupplied, and then more kept coming. It just never stopped.

2023 was the ‘worst year in a generation’ for the bike industry. There hadn’t been a boom and bust like this since the ‘70s. 2024 was worse. And 2025… there’s been no great recovery. Supply—from local bike stores, big-box retailers and now Direct-to-Consumer (DTC) online sales—continues to vastly outweigh demand and has for years. And now the sales start in the spring, not the fall.

AM: Is the pain from oversupply hitting manufacturers and retailers equally, or is one group getting hurt more?

KML: It is asymmetrical, and worse for the bike stores. Once trust is broken between manufacturers and retailers, if somebody pushes inventory down on you, you’re not going to accept that again.

One of the oldest and best bike stores in the city, Urbane, has moved from College and Spadina—where they were for maybe 15, 20 years—further west to a much smaller location. One retailer in 2023 put money back into his company for the first time, a healthy equivalent of what he would normally take out. In 2024, he put in way more still. I know retailers that started COVID with three locations and are down to one. That’s not uncommon. But big brands are suffering too, with one of the biggest e-bike companies—Rad Power—going under late last year. And every bike brand, DTC or local retailer that goes under means another batch of stock at bargain prices.

Why E-Bikes Matter

AM: For most people who don’t own an e-bike, they pattern-match to two things: food couriers cutting them off, or battery fires. What’s the case for e-bikes that they’re missing?

KML: I would say there’s another pattern—maybe not big enough—which is: my neighbour, my best friend, my bridge partner got an e-bike, and she won’t stop talking about how much she loves it.

I’ve spent essentially my whole life trying to help build businesses that offer people environmental products that they want to use, that impact the environment less than what they’re using now. E-bikes are a fucking game changer. They really are.

A car is a victim of its own success. They solve problems for us. Then we built our cities around them, and now there’s too many cars, and until now no good alternatives. But an e-bike really provides an opportunity for lots of people to make lots of trips in a way they’re happy with that they otherwise would have had to drive a car for. Too long to walk, inconvenient for transit.

I’m in my 50s. My most recent bike shop was 8km away, about a 30-minute daily bike ride. It was a straight shot on two streetcars that could take anywhere between 20 and 45 minutes, on any particular day with no advance notice. Whereas an e-bike, I could do it in 20 minutes every day. And I loved that ride. There was never a day I wished I hadn’t. With an e-bike, it could be snow on the ground, it could be raining—those are fair things to opt out on certain days. But I was never thinking ‘that hill is too big’ or ‘I don’t feel like it’ because the assist basically makes it fun.

I’ve got a lot of dad jokes from the e-bike business. One of my favourites, especially if they’re clearly in their 40s, 50s, 60s or 70s: e-bikes come in three speeds—5, 10, and 20 years younger. That’s really how it makes you feel when you get on an e-bike. All of a sudden you’re like, “Holy cats, I’m not tired.”

What E-Bikes Need to Succeed

AM: What are the infrastructure requirements to allow people to cycle? What problems do we need to solve beyond the bike itself?

KML: Thirty-plus years ago, when I was helping launch Evergreen, I’d end up in church basements talking about planting native species. There’d be five people, three of whom came for the free cookies. And there’d be somebody from the biking community trying to encourage people to commute by bike. The number one thing was ‘showers at work’. That’s what was going to help people ride to work. Today, most companies’ HR departments wouldn’t want to deal with huge showers for their employee base. But thanks to e-bikes we don’t need them. Pedal-assist bikes mean you can arrive at work sweat-free, so you don’t need showers. You need something else.

What people want now is safe bike lanes and safe roads.

Already in COVID, especially as I watched the growth of food delivery and was riding an e-bike myself… I’ve commuted by bicycle 30-plus years downtown. We built more bike lanes in COVID than the last 25 years before that in a city like Toronto. We built them so fast, but they’re full, and they’re more than full. Not only are they full, they’re full of multiple kinds of people we didn’t have 10 years ago. You not only have cycling commuters, but you’ve also got food delivery people on e-bikes. Other commuters on e-bikes. More tourists on Bixi [i.e., public bikeshare] bikes. All these different users are trying to squeeze into lanes that sometimes really aren’t meant for more than one abreast.

Artist’s conception

We’re already essentially at 100% capacity on bike lanes. In some places, we’re definitely getting to the point where, yeah, you might need to drive to your front door, but you’re going to have to do it at 20 or 30 kilometres an hour. This road is a shared road, and it’s got to be made safe for a variety of users.

AM: So we need better bike lanes. We need secure bike storage—

KML: Bike lanes, or we need to deal with traffic management in a way where we’re willing to say, “Yeah, driving 70 or 80 or 90 on Parkside Drive is not okay.” You can drive somewhere, but you shouldn’t be able to drive at those speeds. Look at speed governors, speed enforcement with cameras. The bottom line is, in our cities, we need to prioritize transit. Transit needs priority. We need to slow down and make things safer.

The Safety Gap

AM: Let’s move from making the city safe for cyclists to making everyone else safe from cyclists. There are two problems people talk about: e-bikes are dangerous because people ride like maniacs, and e-bikes explode and cause fires.

KML: For both problems, the solutions start with the federal government dealing with what comes through our borders and what gets manufactured here.

Part of the problem is it’s not hard to make an e-bike travel faster than is legal, or to purchase a bike designed from the outset to go faster than legal limits.

AM: My e-bike motor cuts out at 32 kilometres per hour. I can go faster, but it’s only my physical effort that maintains that speed. Is that where the law says it has to be?

KML: Essentially across Canada, the law is 32 kilometres max assist speed and 500-watt motor maximum. In Europe, it’s 250 watts—they have better motors—and their max is 25 kilometres. The US is higher: 25 miles an hour (40 kilometres) and 750 watts. Lots of bikes have 750 watts. Essentially all motors can produce more with software. Some have no lock, which means you have a button that you can click on for ‘more power please’.

There should be limits on power. There simply isn’t any enforcement on the rules that exist.

AM: Will that also help with the battery problem? Is that a problem with all batteries or some batteries?

KML: Any lithium-ion battery thrown on a campfire creates a big problem. When you think about the fire in that Thorncliffe Park building, or in Hong Kong… we all have lithium batteries in our computers, phones, gadgets. E-bike batteries are much bigger than those for most consumer items, an EV car excepted. The problems stem from badly made things, badly treated devices, people fiddling with them or trying to charge them faster than they should.

There are certifications. The UL standard in North America is the top standard for safety and certification. But I read something… Health Canada was doing a survey about battery safety, but their survey explicitly doesn’t deal with e-bikes because they sort of fall under some other vehicle designation. There’s a line in it that said there’s essentially zero requirements for lithium-ion batteries to be certified in Canada. It’s freaking shocking. It’s not like there haven’t been fires. That problem starts with the Feds… has to start with Feds saying what can come into the country and creating rules. Even today’s rules aren’t enforced.

When we started renting bikes six years ago, the first bikes had seven-amp batteries: 36 volts, seven amps. Today, typical food delivery people have 35 amps and 48 volts. That’s almost ten times as large, almost an order of magnitude bigger. Are they better and certified? Some are, not all. There are bikes out there with 60-amp batteries. It’s crazy the amount of energy on that bike that can get put into a subway car or an apartment building.

AM: So it makes sense—the issue isn’t that the technology is bad, it’s that there are manufacturers cutting corners. Just like with any unsafe product, we need regulation to prevent misuse.

KML: We need a bar. There is essentially no bar for making a lithium-ion battery, and we need one. We need it higher than now. Then we need enforcement.

What irked me with GO Transit’s rules: they talked to me, then nine months later said, “In two weeks, this new rule is coming in”. Imagine! When any level of government says, “We’re moving towards some percentage of your cars have to be EVs,” it’s “Oh, we can’t do it in 10 years. That’s too fast.” But while they were planning this, two years ago, the bike show was on. There was no GO Transit booth at the bike show talking to retailers about these new standards coming in. New York essentially said everyone is going to have to abide by this… And the date’s going to be here, 12 to 18 months out, so you can prepare for that. And that’s what everyone should do, signal standards and socialize them with retailers well before enforcement.

AM: There’s an interesting idea here that the body most affected is actually the transit providers. They’re going to have to exercise leadership, even if they’d prefer not to.

KML: They’ve already done it in Toronto. TTC just bans all e-bikes for five months of the year: December, January, February, March, April. And GO requires batteries to meet these two certifications, one’s European and one is this UL certified. One or the other. Everyone is all for safety. It’s just… why are we just doing it there? Why are we not bringing this elsewhere and encouraging the whole industry to move forward?

The crisis Kevin describes is real, and largely invisible. Consumers see bikes on sale and assume it’s a good thing. They don’t see the retailers bleeding money or the broken trust between manufacturers and the shops that sell their products. The industry that services and supports e-bikes—the machines that could meaningfully shift urban transportation—is hollowing out.

What struck me most in our conversation was the policy vacuum. To me, Kevin’s story suggests two policy problems operating on different scales of urgency.

Firstly, we have a retail sector trying to survive an inventory hangover that has lasted years, stemming directly out of the pandemic and still unresolved. Secondly, a regulatory vacuum around batteries and power limits that is now being partially filled, imperfectly, by transit agencies with the most to lose.

Regarding the first, I don’t think this requires inventing a new policy regime. What it requires is acknowledging that bike retail absorbed a pandemic shock in service of public health goals, and that the hangover has not cleared on its own, and there is an appropriate public role in helping. That could mean temporary credit facilities, inventory-financing support, or targeted programs that help viable retailers survive long enough for demand to normalize. These are policy choices, and they require us to acknowledge that there is, in fact, a problem.

Regarding the second, Canada has essentially no certification requirements for lithium-ion batteries in e-bikes. Health Canada’s battery-safety survey explicitly excludes them, while Transport Canada is not prioritizing the enforcement of the rules on motor power and speed. Meanwhile, the infrastructure that would make e-bikes safe and practical—protected lanes, secure parking—is already at capacity in the cities that have built it, and nonexistent in most others… and that is where it isn’t being removed, a misguided rollback of the improvements we made during the pandemic.

I shake my head at the fact that it is transit agencies like GO Transit and the TTC that have become de facto regulators because they have skin in the game; they’re the ones whose subway cars and buses catch fire when batteries fail. So they’ve imposed certification requirements and seasonal bans that other orders of government would be better equipped to promulgate and enforce. It’s regulatory leadership by accident, driven by necessity rather than strategy.

For all of that, e-bikes remain, as Kevin puts it, a “fucking game changer.” They solve problems that have bedevilled cycling advocates for decades: hills, sweat, distance, age. They make trips possible that would otherwise require a car. But realizing that potential requires policy attention that hasn’t materialized on battery safety, on infrastructure, and on industry support. The technology is here. The policy isn’t. And that means that the industry that could most easily improve urban transport, by making e-bikes ubiquitous, is fighting for survival.

E-bikes won’t fix our city streets on their own. But they do make the genuinely better option—better for the user, better for the environment, better for the city—easier, not harder. If you’re curious, check out Kevin’s company, Zygg, which would be happy to sell you an e-bike, or offer you an e-bike subscription service; you can try this out for yourself without committing to purchase.

We’ve reached lots of peaks before. Take home video:

• Production of VHS cassettes peaked in 2001, as DVDs displaced them1

• DVDs in turn peaked in 2007, having been displaced by streaming video

Music followed a similar trajectory:

• Vinyl records peaked in 1978, having been displaced by audiocassettes2

• Audiocassettes peaked in 1989, having been displaced by compact discs (CDs)

• CDs peaked in 2000, having been displaced by digital downloads (which were then displaced by streaming)

Books are harder to assess. Unlike films or music, there is no single, continuous, globally-accepted series tracking how many books are produced each year. So the question “when did book production peak?” turns out to be surprisingly difficult to answer.

So let’s instead answer a different question, and look for the closest defensible proxy. For books as physical objects, that proxy is copies printed; these track print-unit sales over time relatively closely, since publishers print roughly and only what they expect to sell. In the best-documented large market, the United States, sales of print books peaked in 2021, when roughly 840 million print copies were sold.3 Every year since has been lower.

That fact alone doesn’t mean much. We hit peak DVD, but watch more video than ever before; same with CDs and music. The format may have changed, but the activity remains, and in fact is doing better than ever. So perhaps the real question is not whether books as objects have peaked, but whether books as a medium have.

Have they? To answer that, we need to be precise about what “peak book” could mean.

Production Up, Attention Down

If ‘peak book’ includes e-books, then we are nowhere near it.

In the United States (again, the market where publishing statistics are particularly detailed), we can see that e-book sales rose rapidly from the mid-2000s through the early 2010s, as dedicated e-readers and smartphones made long-form digital reading convenient.

We had no idea what was coming. Back then, when e-books were introduced, the annual number of new titles was in the low hundreds of thousands; in 2005, for instance, roughly 280,000 new titles were released. By the early 2020s, once self-published books are included, the figure was closer to three million.

That’s right: production of books, whether physical or electronic, has increased by more than an order of magnitude in less than two decades.

That may seem impressive, but hold on: industry revenue has barely moved. Both unit sales and revenues from e-books peaked around 2013 and since then have declined modestly. In nominal terms, it is roughly where it was twenty-five years ago; in real terms, it has fallen sharply.

Why are e-books exploding despite not finding any readers? It’s because electronic publishing removes the key bottleneck, which is production. Once, getting a book into the world required a publisher with access to a press and to distribution networks. The limits on access to those meant one also needed to gain institutional approval. Those constraints limited supply. But they also meant that works that did make it into the world had been professionally edited—an important contribution to quality—as well as at least some chance of being encountered.

None of this is true any longer: today the bottleneck is discovery. When millions of books, most poorly written and unedited, compete for a fixed pool of reader attention, most are never found.4 The result is a statistic that any would-be author should sit down and reflect upon: the average book now sells only a few hundred copies.

What this means is that, on one side, physical books are in decline… but on the other, despite their vast numbers, e-books aren’t taking their place in the culture. So, yes, if ‘peak book’ means production of e-books, it hasn’t happened… but if it means production of physical books, it was 2021.

Rome Didn’t Fall in a Day

Leaving physical objects aside, “when was peak book” can be answered in at least three ways, each harder to pin down than the last, but all arguably more consequential.

The first is the most straightforward: when did sales of books, physical and electronic, peak relative to population?5 There are several ways to approach this, none fully satisfactory, since national publishing statistics vary, international comparisons are messy, and population growth complicates aggregation, etc. Still, plausible reconstructions across different countries find a cluster around the early 2000s, roughly 2000 to 2005. If that estimate is wrong by a few years in either direction, it would not change much.

So peak book-sales was roughly twenty to twenty-five years ago.

The second way in is this: when did books cease to be a widespread form of popular entertainment? This is harder to measure, but it matters far more. In my lifetime, reading was close to universal, not just among the educated but across the middle of society. Cheap paperbacks were everywhere, from Penguin Classics to the ones Gordon Lightfoot described in 1970 as “the kind that drugstores sell”. These were mass-market products because reading was a default way to spend idle time.

If you’ll permit me the indulgence, I’ll say that my own lived experience suggests that this world was already fading by the late twentieth century. Growing up in the 1980s, to me the world felt saturated with books: I still have vivid memories of visiting the small bookshops of Canada—Coles, W.H. Smith—as well as of buying cheap books at newsstands and pharmacies. But by the mid-1990s, that saturation was thinning. By the 2000s, reading—especially long-form reading—was no longer a universal habit. That shift has since accelerated; long-term surveys now show that teenagers who read daily for pleasure have become a minority.

Here, the answer seems to be the same as for the last question, around the end of the 1990s and beginning of the 2000s. In this case, the Harry Potter books, so celebrated at the time as a gateway for young people to the joys of reading for pleasure, turn out to have been the end of the old world, the last time the bell rang so loudly.

The third way to ask the question, and of particular interest to me as a non-fiction writer, is when did writing a book stop being the default way to change the world? Not the only way, not merely one way among many, but the obvious way, the default strategy: the thing one did if one wanted to intervene seriously in public life.

For much of the twentieth century, it was a strong strategy indeed.6 Just off the top of my head: Ken Kesey changed the way his and subsequent generations thought about mental-health provision in One Flew Over the Cuckoo’s Nest (in retrospect, not obviously for the better). Betty Friedan’s The Feminine Mystique catalyzed second-wave feminism. Ralph Nader’s Unsafe at Any Speed and Rachel Carson’s Silent Spring created constituencies who argued successfully for regulation of business and the environment.

Books could, sometimes, be engines of cultural change, and people who had arguments that mattered wrote books to get them across. But that strategy, on the whole, no longer works.

It’s hard to say when that happened, because people are still writing non-fiction books about public policy that have strong impact on the world: just last year, for instance, we had Klein and Thompson’s Abundance, and Yudkowsky and Soares’ If Anyone Builds It, Everyone Dies. Other notable candidates from this century might be Piketty’s Capital in the Twenty-First Century or Kendi’s How to Be an Anti-Racist. These examples are striking precisely because they are rare; non-fiction public policy books that have a big impact seem limited to one, or perhaps two, a year. If I told you the name of the fifth-best-selling book of public policy of last year, would you recognize it? Suave and intelligent readers of Changing Lanes probably would, but I bet most people have never heard of it.7

Al Gore, c’est moi

Everything Ends

So, if we look at all our answers, when was peak book?

The data points cluster. In absolute terms, print sales peaked in 2021, but that was a pandemic-driven anomaly in a longer decline. E-book revenue peaked in 2013. Books sold as a percentage of the population peaked back in the first decade of this century. Reading rates have been falling for decades, with the sharpest drops among the young. It’s still the case that books can shape elite discourse, but fewer and fewer have an opportunity: perhaps one or two a year.

By any of these measures, peak book is behind us, falling somewhere in the early twenty-first century, with the exact year depending on which metric you privilege.

And what that means is this:

Books are a dying art form.

“Books bookshelf read” from Pixabay (2016), public domain under CC0 1.0 Universal

That’s a pungent statement, so let me hasten to explain what it means, and the best way to do that is to think about a different dead art form: opera.

Opera is still performed. The Metropolitan Opera still stages productions; regional companies still exist; new operas are still commissioned, with roughly thirty premieres per year in North America alone. OPERA America counts more than 200 professional companies in the United States and Canada.

But none of that means opera is alive. Attendance has collapsed: in 2017, only 2.2 percent of American adults had attended, even once, a live opera performance. And of those who do consume opera, they aren’t interested in new works. As per Sam Hughes at Works in Progress:

This interesting website has a database of all operas performed globally today. Of the top fifty operas by number of performances, only one was written after 1914 (Turandot, 1924)… The vast number of modernist operas written over the last century have virtually no place in the performance canon.

So some few people still enjoy opera, and those few people are, I imagine, much wealthier and more educated than the median person. But that doesn’t matter to our evaluation of whether opera is dead. Opera is certainly dead… because opera no longer participates meaningfully in the broader cultural conversation. No one expects a new opera to define a generation’s concerns, or to be the place where important cultural work happens. In fact new opera, though it exists, is barely attended to, even by the art form’s few fans.

Books are not there yet, but they are closer than many readers of books would like to admit.

To be clear, in the intervening years people have never stopped encountering text: articles, newsletters, social-media posts, and on and on. But the book is a particular kind of text: a long-form collection of words, typically no fewer than 80,000 words, expected to be consumed over long, uninterrupted stretches of linear time. And I think uninterrupted is key here. Reading a book requires not only a commitment of time, but of attention as well. Books are cognitively demanding; one cannot read a serious book while doing something else. Conversely, one can, and many routinely do, watch television, listen to podcasts, or scroll video feeds while cooking, commuting, exercising, or half-working. These other kinds of media tolerate distraction but books don’t.

That makes them unsuited to our time, and even less so for the world that is coming.

The death of the book has been a long, uneven process, still going on. Books lost cultural power first, then leisure dominance, and only later did this show up clearly in sales figures. Books were outcompeted, slowly and unevenly, by media better suited to the modern attention economy, which no longer rewards sustained, exclusive focus. They will continue to exist, and to matter to some. But the era in which they served as the central organizing unit of intellectual and cultural life is firmly behind us.

Which is a roundabout way of saying that while I’m proud of what my co-authors and I accomplished with our book, The End of Driving (buy it here!), I think writing a second book would be a bad use of my time.

The book is dead; long live the Substack.

Critics argued that the company was a ‘bezzle’. It looked like a company, but it was actually a long con. Its putative core business, selling rides, was dressed up with regulatory arbitrage and accounting gimmicks to suggest future profitability, in order to separate credulous investors from their money now. According to the critics, at some point soon the investors would realize they’d been taken, dump their shares, and the whole enterprise would collapse.

Four years later, Uber is still here… and not ‘burning investor cash while promising future profits’ still here, but ‘making real money’ still here. As of last month, late 2025, Uber was generating billions of dollars in trailing twelve-month free cash flow. It didn’t have a fluke quarter, and it didn’t pull an accounting trick, like taking gains on paper from asset revaluations. Uber has achieved sustained operating profitability across its core businesses for ten consecutive quarters.

The consensus was wrong. Why?

The 2021 critics were not cranks (well, the best of them weren’t). Their arguments were detailed, numerate, and grounded in Uber’s own filings. They had a strong and careful argument that ride-hailing was structurally incapable of producing durable profits in competitive markets. And yet the company did not collapse.

This is a puzzle. Let’s solve it.

The Analysis

The serious work began with Hubert Horan, a transport consultant with decades of experience analyzing airline and transit economics. Starting in 2016, Horan published a multi-part series titled Can Uber Ever Deliver? on Naked Capitalism, later expanded into a comprehensive essay for American Affairs in 2019.

Horan’s thesis was that Uber possessed no structural cost advantage over traditional taxis. The technology that Uber deployed—GPS dispatch, smartphone apps, cashless payment—was genuinely useful, but it was also replicable. Any taxi company could adopt the same tools, and many had. The app wasn’t a moat; it was a commodity. What Uber offered was lower prices, but those prices were kept artificially low through investor subsidies. Passengers were paying less than the true cost of their rides, with venture capital covering the difference.

And Horan claimed that what was true on the customer side was also true on the driver side. Horan documented how driver incentives further inflated costs beyond what the fare structure could support. Uber paid bonuses, guaranteed minimum hourly earnings, and offered surge pricing that flowed disproportionately to drivers rather than to the platform. These were necessary to attract and retain drivers in a competitive labour market, but they meant the company was hemorrhaging cash on both sides of the transaction.

So why were investors giving so much money for Uber to in turn give away to riders and drivers, without any hope of return? According to Horan, Uber was telling everyone it was playing a “winner-take-all” game. Once Uber got big enough, it would have network effects; a rider in Toronto would have the app on their phone and would naturally use it whenever they took a business trip to New York or Los Angeles, rather than researching and using a local competitor. The more riders used it, the more drivers would want to be on the system, and the more drivers there were on the system, the more riders would want to use it (because wait times would be lower than with a smaller competitor). By growing fast, Uber would build out a network, discouraging any rival from emerging. And once that had happened—rivals squeezed out and no new rivals able to appear—Uber would be able to raise fares on riders while cutting driver compensation, becoming very profitable forever.

Horan’s argument was that network effects didn’t apply here. In classic network-effect businesses, the unacceptably high cost of switching locks users in. But riders had both the Uber and Lyft apps on their phones, and drivers were registered on both platforms, and in the heady days of the 2010s, both were switching freely to whoever offered the better deal. There was no lock-in and no accumulating advantage. The subsidies weren’t buying a monopoly but only market share; and that would evaporate the moment the subsidies stopped.

Horan also emphasized that regulatory arbitrage was central to Uber’s economics, not incidental to them. The company’s cost structure depended on avoiding taxi medallion costs, dodging commercial insurance requirements, and classifying drivers as independent contractors rather than employees. Strip away the regulatory advantages and Uber’s costs would rise to meet or exceed those of traditional taxis, without any offsetting efficiency gains to compensate.

The numbers supported this reading. Uber lost $8.5 billion in 2019, the year of its IPO. It lost another $6.8 billion in 2020. By 2023, the company’s accumulated deficit exceeded $30 billion. When Uber reported nominal profits in 2021 and 2022, Horan and others pointed out that these were driven by one-off revaluations of minority equity stakes in other companies (like Didi, Aurora, and Grab); these were paper gains that had nothing to do with whether the core business of moving passengers from point A to point B could generate positive cash flow.

Let’s pause here to make the argument clear. Horan (often amplified by Yves Smith, editor of Naked Capitalism) was making several distinct claims:

1. Subsidy illusion: Uber’s value proposition to consumers, and its growth as a company, were propped up by investor cash, and absent subsidy, the company would only be able to cover its costs by charging fares no one would be willing to pay

2. Regulatory dependency: Uber’s unit economics depended on temporary regulatory arbitrage (classification, insurance, licensing) that would only last until regulators ceased to permit it

3. The impossibility of profit: Uber was operating in competitive markets, and as such competition would reduce their profit to zero

4. Inevitable collapse: Once subsidies ended and regulation tightened, Uber would fail rather than adapt into a durable business

Through 2022, Claims 1 and 2 fit the observable facts well: Uber’s initial growth had indeed been subsidy-driven, and regulatory arbitrage really was central to its operating model. And given what the numbers showed, Claim 3 was reasonable in the context of then-Uber, because the company seemed structurally unable to quit buying market share, and Claim 4 was a conclusion deduced from the preceding three.

Cory Doctorow—science‑fiction author, activist, and incurious take-slinger—took this argument further. Doctorow credited Horan’s work, calling him “the only person you need to read” on Uber’s economics, added one more step. Given Claims 1 through 4, Uber’s ultimate failure is assured, so why is the firm proceeding? Enter Claim 5:

5. Uber is a bezzle: Uber isn’t a firm, but a long con: it takes investor money (as well as a high stock-market valuation) now in return for a promise of future profits that are never coming, and which the firm’s leaders know are never coming

Doctorow’s language is explicit and blunt: pungent examples include “Uber was never going to be profitable. Never”, or Uber “is about to die.” And it included specific predictions.

• August 2021: “Uber’s time is up… Uber is going broke”

• August 2022: “Every bezzle ends. Uber’s days are, therefore, numbered”

• August 2023, after Uber had begun reporting operating profits: “Uber was, is, and always will be a bezzle”

All five claims, taken together, congealed into a meme: Uber is a bezzle.1

What Changed

The simplest way to say what happened between 2022 and 2025 is not that Uber proved the critics wrong, but that Uber stopped being the same object the critique was describing. The company that achieved profitability in 2023 operates under different constraints, charges different prices, and extracts value differently than the company Horan analyzed in 2016 or Doctorow declared dead in 2021.

Several factors spurred the company to change. The obvious one was the pandemic, which delivered a massive shock to ride-hailing demand, temporarily collapsing trip volumes and forcing the company to confront how expensive its growth-at-any-price strategy was. Less obviously, when the pandemic went away, so too did the cheap capital that had been a feature of the markets from the 2008 global financial crisis onward. But those macro-environment shifts only provided an impetus; they weren’t the changes themselves.

So what were the changes that made Uber profitable?

It did exactly what its investors expected it to do: it charged riders more and paid drivers less.

Starting in late 2020 and accelerating through 2022, Uber enacted substantial fare increases across its markets. Riders in many cities saw typical fares jump by 20 to 50 percent from 2019 levels. By 2022, Uber rides were often more expensive than the traditional taxis they had disrupted. Simultaneously, Uber increased its ‘take rate’—the share of each fare it keeps after paying drivers—from roughly 22 percent before the pandemic to nearly 30 percent by 2024. This was a direct transfer of value from drivers to Uber’s bottom line.

Uber was able to make these changes simultaneously because it shifted from a transparent commission model to algorithmic allocation, where what a passenger pays and what a driver earns are determined separately (and can diverge dramatically). Uber now uses data on rider behaviour to charge each customer something closer to their maximum willingness to pay, while offering drivers the minimum they’ll accept for each trip. This is price discrimination on both sides of the market, enabled by information asymmetry and executed at scale.

The critics had predicted this move, or something like it. What they didn’t predict was that Uber could pull it off without triggering the expected consequences. Horan’s model assumed that raising prices or cutting driver pay would invite one of three responses: riders would defect to cheaper alternatives, drivers would quit for better opportunities, or competitors would undercut Uber’s new margins.

None of these materialized at the scale required to discipline Uber’s pricing.

Riders didn’t flee because they had nowhere to go. Traditional taxi systems, lazy monopolists hollowed out by a decade of stiff competition, couldn’t absorb returning demand even if riders wanted them to. Lyft, Uber’s only significant North American competitor, was in worse financial shape and couldn’t afford to restart the subsidy wars; instead, it mirrored Uber’s price increases. New entrants faced not only the difficulty of building a rival to Uber, hard in a capital-constrained environment, but also the credible threat of Uber retaliation; the market had watched Uber burn billions to crush competitors, and no investors were willing to test if the company was willing to do it again. The result was something closer to a duopoly than a competitive market.

Drivers didn’t quit en masse because they lacked alternatives. The gig model ensures a constant supply of new entrants: people between jobs, those needing supplemental income, those valuing schedule flexibility. Much as in long-haul trucking, driver turnover is high, but aggregate supply seems always to be sufficient to meet demand, as friend of Changing Lanes has discussed recently. There were moments of strain—early 2021 saw acute driver shortages as demand rebounded faster than supply—but Uber navigated these with targeted, temporary incentives rather than structural pay increases. By 2022, the driver pool had stabilized at the new, lower compensation level.

The painful truth is that Uber’s early subsidies bought something durable after all: not the network effects the bulls had promised, but market power. By proving it would crush any rival, Uber created an effective barrier to entry. By the time it raised prices, there was no one left to undercut it.

The critics’ second major bet—that regulators would eventually force Uber to internalize its true costs—also failed to materialize as predicted, though not for lack of trying.

The crucial battle was California’s Proposition 22 in November 2020. California had passed AB5, a law that would likely require Uber to classify drivers as employees rather than independent contractors, with minimum wage obligations, benefits, and tax contributions. Uber, joined by Lyft and DoorDash, spent over $200 million on a ballot initiative to carve out an exemption, and won. The result preserved Uber’s contractor model in its largest market and sent a clear signal about the company’s political capacity.

That strategy didn’t succeed everywhere. The UK Supreme Court ruled in 2021 that drivers were ‘workers’ entitled to minimum wage and holiday pay. Uber absorbed the costs and raised fares to compensate. These were real expenses, but manageable rather than existential.

The pattern across jurisdictions was similar: regulators extracted concessions, Uber absorbed them, and the business continued. The scenario where labour reclassification would expose Uber’s fundamental unsoundness never arrived.

On the expense side, Uber made cuts that would have been politically difficult before the pandemic forced the company’s hand. In 2020, it laid off thousands of employees and shed over $1 billion in fixed costs. More importantly, it abandoned the moonshot projects that had consumed capital without generating revenue.

The autonomous vehicle unit, Uber ATG, had spent over $2.5 billion developing self-driving technology that never reached commercial viability. Uber sold it to Aurora in late 2020. The flying taxi project went similarly. E-bike and scooter operations were offloaded to Lime. Unprofitable international markets were exited or sold. What remained was a leaner company focused on two core businesses: rides and delivery.

Uber also quietly contracted the scope of its ride service. The pre-pandemic Uber had prided itself on ubiquity: coverage across suburbs and exurbs, service at all hours, fast pickup times everywhere. Much of this was subsidized; rides in low-density areas and off-peak times hemorrhaged money. Post-2020 Uber concentrated on dense urban cores and peak demand, where drivers could achieve higher utilization. The service became more like a traditional taxi fleet: profitable routes and times first, everything else second.

Evaluating the Claims

With the mechanisms traced, let’s return to the five claims that comprised the bezzle thesis and assess them against the evidence.

Claim 1: Subsidy Illusion. The claim was that Uber’s growth and value proposition were propped up by investor cash, and that absent subsidy, the company could only cover costs by charging fares no one would pay.

The subsidy was real: over $30 billion in accumulated losses funded below-cost rides for a decade. That investment did what it was supposed to do, namely digging a moat. Uber used the subsidy period to destroy the competitive alternatives that would have disciplined its pricing, and riders proved willing to pay the higher fares the company then decided to charge.

Let’s call this one correct in its prediction, but false in its diagnosis.

Claim 2: Regulatory Dependency. The claim was that Uber’s unit economics depended on temporary regulatory arbitrage—contractor classification, insurance avoidance, licensing circumvention—that would only last until regulators closed the loopholes.

Uber’s model does depend on the contractor classification; internalizing full employment costs would substantially erode the firm’s margins. But that classification does not appear to be temporary. Uber proved capable of reshaping the regulatory environment through lobbying, ballot sponsorship, and when all else fails just absorbing the costs.

Let’s call this one false as well. The arbitrage turned out to be more durable than critics expected, in part because Uber actively defended it.

Claim 3: Impossibility of Profit. The claim was that Uber, operating in competitive markets, would see profits competed away to zero because it possessed no structural cost advantage.

The key word is competitive. Uber secured market power sufficient to escape discipline. The critics were right that Uber had no cost advantage, but wrong that this precluded profit; in an uncompetitive market, pricing power is all you need, and Uber acquired it. So this claim was conditionally correct, but was predicated on an assumption that proved incorrect.

Claim 4: Inevitable Collapse. The claim was that once subsidies ended and regulation tightened, Uber would fail rather than adapt.

To the contrary, Uber adapted. The company that exists in 2025 is not the company the critics analyzed: it charges different prices, operates in a different market, and runs a different cost structure. The prediction of collapse treated the firm as static when it was dynamic. This was the core analytical error: assuming that because early Uber couldn’t be profitable, no future version of Uber could be either. This claim has proven to be false.

Claim 5: Uber Is a Bezzle. The claim was that given the foregoing, no honest person could believe that Uber would ever succeed, so the firm’s leaders must be confidence tricksters, extracting money now in the hopes of getting out before the inevitable collapse. And yet the collapse has proven evitable. So this claim is false as well.

That means that all five claims in the thesis are incorrect or false.

I’ll give them their due: Uber’s critics correctly identified that early Uber ran on subsidized growth disconnected from operational fundamentals, and correctly predicted that reaching profitability would require raising prices and squeezing drivers. They incorrectly predicted that Uber would prove unable to do this.

If Uber’s operating profits prove unsustainable over the next two to three years—if they depend on one-time cost cuts or a favourable competitive moment that passes—the bezzle thesis regains force. A recession that collapses demand, or a resurgent competitor willing to restart the subsidy wars, would test whether Uber’s profitability is structural or circumstantial.

Conversely, if Uber maintains profitability through an economic downturn or competitive challenge, the case for genuine transformation solidifies. The longer the profits persist, the harder it becomes to argue that they’re illusory.

Image: “Ponzi scheme” by Stock Catalog (CC BY 2.0, via Flickr)

[Yes, that’s really the image title…]

I highlighted Doctorow’s 2021 post because it represents the moment the bezzle meme reached peak influence: confident, widely shared, and, in retrospect, timed almost precisely as the underlying facts were beginning to shift.

The failure mode here is not bad analysis, but static thinking. We are good at analyzing snapshots and bad at tracking drift. We are good at conditional models and bad at noticing when their premises expire.

So what’s the lesson? Not that the critics were foolish. Horan’s analysis of Uber’s unit economics was rigorous and, on its own terms, correct. Uber really did have no cost advantage over incumbent taxis. It really was burning investor cash to subsidize below-cost rides. The critics predicted that profitability would require raising fares and squeezing drivers, and that’s exactly what happened.

The error was treating the competitive and regulatory environment as fixed, and thus that market discipline and legal constraints would eventually bite. They didn’t, because Uber spent the subsidy years making sure they wouldn’t.

I’m sanguine about this outcome. The taxi cartels with their medallions were extractive; they enriched speculators while limiting ride supply and delivering indifferent service. Uber modernized the sector, expanded access, improved reliability, and created flexible work that millions of drivers have chosen over alternatives. Against this, Uber and other ridehail firms have increased urban congestion and depressed transit ridership; but the solution to this isn’t demonizing Uber, it’s implementing a congestion charge, which we largely have not done. We’ve got no one to blame for that but ourselves.

Meanwhile, the sector is changing again. Waymo alone now provides over 450,000 robotaxi rides per week, and Zoox and Tesla are ramping up their service. Uber, having sold its own self-driving unit in 2020, has pivoted to partnerships: Waymo vehicles are now available through the Uber app in Phoenix, Austin, and Atlanta, with more cities coming. The economics aren’t there yet—Uber’s CFO conceded in August 2025 that “AVs today are not profitable”—but the trajectory is unmistakable. It may soon be the case that we think Uber walked so Waymo could run.

The Navier N30; image courtesy of Navier

Understanding what the N30 represents made me believe that urban water transit, a category I’d previously dismissed as impractical, might finally become real.

To be clear, hydrofoils like the N30 are not a replacement for buses and trains; Navier isn’t building mass transit.1 A six-passenger hydrofoil will never compete with a subway car. But deploy enough hydrofoils, running frequently between multiple points, and you create something valuable: premium ferries that offer speed and comfort impossible with conventional boats, serving routes where water creates geometric shortcuts that roads can’t match. They could improve on the water taxi so much as to create something new, adding a genuinely novel category of trip to a city’s transport network.

This is not a new idea, but one that some people have been proposing in vain for more than a century. Hydrofoiling boats are an old technology, first proposed in the late 19th century and first successfully demonstrated by Enrico Forlanini in 1906. But for the most part they’ve only ever been toys, or fads, never reaching full-scale deployment in the West. Has anything changed?

Navier, a Bay-Area startup, thinks it has. During the Progress Conference in October 2025, founder and CEO Sampriti Bhattacharyya and her engineering team invited me and other attendees to visit their production facility in Alameda, where I got to inspect the N30 up close.2 It’s an impressive machine, and Navier is working to commercialize it at scale. What makes them believe they can succeed where Boeing and others failed? To find out, in November 2025 I interviewed Bhattacharyya.

From her, I learned that three technological convergences—in control systems, batteries, and materials—have shifted hydrofoil economics from insupportable to viable. If Navier’s view of the situation is correct, and the company succeeds in making hydrofoils readily available, its success will have implications for the world’s navies, its pleasure craft, and more… but especially for what interests us particularly at Changing Lanes, namely improving urban transport.

So: under what real-world conditions will modern hydrofoil technology move the needle on urban transport? And where will it never work, no matter how good the technology becomes?

Once, a Toy or a Sideshow

First: what is a hydrofoil, anyway?

Put simply, hydrofoils are boats that generate lift to get the hull out of the water. Like an aircraft wing, a hydrofoil uses a curved upper surface and a flatter lower surface. As the boat accelerates, the boat deploys these foils on posts beneath the hull. As water flows over the foil, the curve forces water traveling over the top to move faster than water flowing underneath. This creates a pressure differential—lower pressure above, higher pressure below—that generates upward force. It’s Bernoulli’s principle, the same mechanism that keeps airplanes aloft, but operating in water, a medium 800 times denser than air.

Lift matters because it decreases resistance. A conventional boat pushes through water, and the resistance it encounters increases exponentially with speed. (Specifically: drag scales roughly with the cube of velocity, meaning that to double your speed you need eight times the power to overcome resistance.) This is why fast powerboats are so fuel-inefficient: they’re fighting enormous drag forces as they plow through water at high speed. But lift the hull out of the water and the equation changes dramatically. The Navier N30, once on foils, keeps only the thin wings and drive pods submerged. In motion, as little as five percent of the hull’s surface area is in contact with water.

That’s the key to hydrofoil efficiency: at speed, you’re no longer dragging a hull through water, but instead supporting the boat’s weight with minimal submerged surface. The hull flies above the water, slicing through rather than plowing. This drag reduction translates directly into energy efficiency and range. A conventional electric boat traveling at 30 knots might use 200 kilowatt-hours to cover 40 nautical miles; the N30, foiling at the same speed, uses about 40 kilowatt-hours for the same distance.

Lift has other gifts to offer besides efficient operation. A traditional boat’s stability comes from buoyancy; the boat remains upright because of the shape of its hull and the distribution of its weight. Conversely, a hydrofoil’s stability comes from its control system. Stipulating that the struts are longer than the waves are tall, the hull rides above the surface and doesn’t feel any chop (i.e., the short or ‘choppy’ waves kicked up by wind). The N30’s six-foot struts keep the hull four feet above waves; in three-foot chop, passengers feel almost nothing.

The principles were grasped early. Just over a decade after the first hydrofoil was built, Alexander Graham Bell—yes, the inventor of the telephone—set water-speed records with hydrofoils in 1919; those records would stand for decades. That sounds impressive, a testament to Bell’s genius, but a moment’s thought reveals the opposite. The fact the records lasted so long is a bad thing, Perversely, it tells us hydrofoil was a dead-end technology. (We’ll get to why that was the case presently.)

Through the mid-20th century, hydrofoils were little more than a curiosity: the Soviet Union operated large fleets of riverine hydrofoils, and the U.S. Navy experimented with patrol craft, but none achieved sustained commercial use. The most serious commercial attempt came from Boeing in the 1970s and 1980s. The aerospace giant built its Jetfoil series of passenger hydrofoils that could carry up to 400 passengers at 45 knots. These were actual commercial vessels operating routes in Japan and Hong Kong. Boeing built about 30 of them, before killing the program entirely.

Why mothball the Jetfoils? Because they were expensive to operate and even more expensive to maintain. The control systems required to keep a hydrofoil stable, constantly adjusting foil angles to compensate for waves, wind, and weight distribution, were built with cutting-edge technology for the time: vacuum tubes and early computers. It was a bravura effort, but doomed:

• The hulls were aluminum (lighter than the fiberglass common in smaller boats, but still heavy)

• The weight of the hull and the gasoline engine meant immense fuel costs, too expensive with post-Oil-Shock gasoline prices

• The control system was complex and prone to failure

• Foils and actuators required constant calibration

• Saltwater exposure corroded the foils and other moving parts

• The analogue electronics of the era drifted, requiring intensive, expensive dockside engineering hours for every operating day

It didn’t take long for Boeing to accept that the economics didn’t work. By the 1990s, the firm had exited the hydrofoil business entirely.

That’s the pattern: hydrofoils are discovered or rediscovered as an exciting technology, someone builds a few, and then the project quietly shuts down when operational costs became clear. The physics works well but the economics don’t.

The 2020s Convergence

What’s changed is the convergence of three technological shifts. As Bhattacharyya explained when I asked what’s different from Boeing’s era: “Three things really change the dynamics from the technology perspective: cheaper, faster computing and sensing, batteries, and scalable manufacturing.” We can see all three in Navier’s N30.

The first is cheap, fast, reliable computing and sensing. Boeing’s Jetfoils failed partly because, absent these capabilities, active control was prohibitively expensive. By active control, I mean the complicated series of frequent, precise moves necessary to keep the hydrofoil steady and on course in moving water. Today, active control is cheap, thanks to the smartphone revolution. The sensors that stabilize your smartphone’s camera—accelerometers, gyroscopes, magnetometers—can stabilize a hydrofoil. It’s the same tech stack that enables drones, robotaxis, and other automated vehicles: commodity parts manufactured by the millions, no longer exotic but off-the-shelf.3 They’re available at low cost while being more accurate and reliable than anything available in the 1970s.

The second is battery technology, another gift of the smartphone revolution. Electric motors equipped with high-capacity batteries are lighter, cheaper, and cost less to ‘fill up’ than a gasoline engine. Electric drivetrains are also mechanically simpler, meaning less maintenance, no oil changes, no heavy transmission to support; but the real edge is in cost of fuel. Given that hydrofoiling itself is dramatically energy-efficient compared to traditional hull-in-water boat travel, an electric hydrofoil offers immense savings on operating costs. Navier claims their vessels use 90 percent less energy to run than a comparable conventional boat (30 cents per mile for electricity versus several dollars for gasoline). And while today’s batteries do impose range limits, urban water-transit distances are short: San Francisco to Oakland is six miles, for instance, while Manhattan to Brooklyn is merely two.

The third is scalable carbon-fiber manufacturing. Hydrofoils only work if the boat is light enough for the wings to generate sufficient lift. Traditional hydrofoils used heavy fiberglass hulls due to lack of anything better. But today we do have better: carbon-fiber hulls weigh one-third as much as fiberglass for the same strength. There was a time when carbon fiber was exotic aerospace material, but thanks to modern manufacturing, it’s now affordable for boat-sized structures. As Bhattacharyya notes, the weight difference is dramatic: “If you think about a 30-foot fibreglass boat, it is like three times the weight of what this boat weighs, which makes it easier to lift with the foils.” To be clear: the lighter hull does not reduce drag directly, but it does change the system’s operating point: a lighter vessel can lift off at lower speeds and with smaller foils, reducing wetted surface area and the power required to reach and sustain flight. In practice, this translates into higher efficiency and longer range.

Each of these factors alone would have helped hydrofoils improve, but not enough to be viable, one imagines. But all three converging at roughly the same time means that hydrofoils, previously a toy or curiosity or white elephant, are now competitive with traditional watercraft. The N30 is a capable machine: 30 knots top speed, up to 75 nautical miles of range on a 114 kWh battery, and requires less than an hour to recharge with a fast charger, and all for a base price of $850,000 USD (comparable to a high-end luxury boat).

Nor is that all. Navier is also poised to introduce full vehicle automation, meaning that their watercraft could be the Waymo of the water.

Navier’s boats already feature automated docking, waypoint following, collision detection via radar/cameras/AIS, 360-degree camera coverage, and cloud diagnostics. These are valuable enough on their own: they keep the vehicle even more stable in the water than older hydrofoils, reduce the pilot’s workload, and, critically, allow the boat to dock itself, the trickiest, slowest, and most complex maneuver a watercraft must undertake. As technology proves itself and regulations evolve, full automated navigation will become possible, offering the same potential that robotaxis do: cheap, fast, reliable, safe travel, available 24/7.

Bhattacharyya frames the value proposition around what she calls “the 3 Cs”, namely cost, convenience, and comfort.

• Cost (savings) come from hydrofoil efficiency

• Convenience comes from integration with existing infrastructure. Unlike ferries, which require specialized terminals costing tens of millions of dollars and decades for permitting approval, the N30 can use any existing marina or dock. “You can take your car to the marina, jump onto a hydrofoiling vessel, and all you’re looking at is: how much time from point A to B”, and what you’ll have to pay for the trip

• Comfort may sound like a mere nice-to-have, but it is arguably the most important item on this list; it is what makes small-boat transit possible at all

That last claim needs explication. In two-foot waves, a conventional boat (with what mariners call a ‘planing hull’) subjects passengers to repeated impacts exceeding 1.5g; that's brutal enough to cause seasickness on a single trip and, for regular commuters, risks genuine injury, while guaranteeing an unpleasant time. No amount of cost reduction would make a trip like that viable as transit. Conversely, on foils, vertical acceleration drops below 0.2g. This is a step-change: the ride becomes smooth enough to work on a laptop, hold a conversation, or simply arrive without needing to recover. As Bhattacharya put it, “There's no loud engine noise. You can actually have a conversation.” The increase in comfort that hydrofoils offer crosses the threshold from non-viable to viable.

That’s the future that Navier and its competitors are aiming for. Let’s stipulate that they can make it happen, and can build automated hydrofoils at scale. Where should these craft be deployed?

Where Hydrofoils Will Work

Navier, like all startups, has big dreams, and I applaud their verve.

But as an outside observer, I’m less sanguine than they are that this is a universal technology. In my view, most coastal cities are poor candidates for hydrofoil transit, and no amount of technological improvement changes that, because the geography is fundamentally inappropriate. Viable hydrofoil markets need three conditions simultaneously, and most places fail on at least one.

Put another way: the hydrofoil equation is potential for shortcuts × amount of sheltered waters × complexity of network topology. If any of these factors is zero, the use case is zero.

Let’s take those factors in turn.

The first is the potential for shortcuts. A city’s waterfront needs to offer shortcuts that roads can’t match. If your coastline is linear, with water running parallel to roads and rail, boats don’t save meaningful time. Chicago’s lakefront is an obvious example: yes, Lake Michigan is right there, but the lakeshore follows the same north-south line as Lake Shore Drive and the Red Line. The water doesn’t create new opportunities, but only duplicates an axis that already has good transport options.

What you want is non-linear water geography: bays, harbours, straits, or complex coastlines where water creates triangulation opportunities. The obvious example here is San Francisco Bay, which is indeed where Navier has begun to offer water-shuttle service. Going from Oakland to San Francisco by car means driving across the bay via the highly-congested Bay Bridge, a trip of 18 kilometers and who knows how much delay. But traveling directly across the water, one can make the trip in a 13-kilometer, straight-line, uncongested and predictable trip. The bay’s contours give water travel a geometric advantage that roads can’t replicate.

The second is sheltered waters. Hydrofoils handle chop better than traditional small boats—their active stabilization keeps the hull above waves rather than slamming through them—but they still need protection from open-ocean swells. Bays, harbours, and coastal waters work; the open Caribbean does not.4 San Francisco Bay, classified by the Coast Guard as “partially protected,” sees regular three-to-five-foot chop, conditions that would make passengers on a conventional small boat miserable but that the N30 handles comfortably.

The third is multiple origin-destination pairs. A single point-to-point route—mainland to island—is better served by a conventional large ferry. Big ferries have terrible economics for on-demand service and multiple routes, but for a single fixed route with predictable demand, they can’t be beat, because with limited trips and a simple schedule they can always clear the market for trips. Conversely, ferries work poorly, and hydrofoils win, when you have complex network topology: multiple origins and destinations, such that travel demand is a kaleidoscope of combinations. In cases like that, on-demand small boats with short headways provide more value than a big ferry on a fixed schedule.

These three conditions must occur together; fail on any one and hydrofoils probably don’t make economic sense.

There’s a fourth factor worth acknowledging: time certainty. Water doesn’t congest the way roads do. A hydrofoil trip takes the same twenty minutes at rush hour as at midnight; a car trip through bridge traffic might take thirty minutes or ninety, depending on the day. That predictability has real value, particularly for high-income commuters whose time costs are significant.5

But I’m skeptical that time certainty alone—without geometric shortcuts—creates genuine urban transit. Transit serves masses of people across many origins and destinations; it requires accessible infrastructure and price points that attract riders beyond the wealthy. A hydrofoil service competing on comfort and predictability with a car trip of similar duration is a premium product, not mass transit. It’s a better limousine, not a better bus. In corridors where water also saves distance, the economics shift: you get time certainty plus time savings, a combination strong enough to attract ridership beyond the luxury segment. That’s why I weight geometric shortcuts heavily in my framework; not because time certainty doesn’t matter, but because without distance savings, you’re building a niche service rather than changing how a city moves.

Taking all of these factors together, we can immediately see that San Francisco Bay is the gold standard. Non-linear geography? Absolutely; crossing the bay saves enormous distances. Sheltered waters? Yes; it’s a bay, not open ocean. Wave conditions get rough during storms, but most days the Bay is navigable for small boats. Multiple origin-destination pairs? Oakland, San Francisco, Alameda, Berkeley, Sausalito, Tiburon, and more make for dozens of potential routes creating a genuine network. If hydrofoils work anywhere, they work here.

Call San Francisco Bay a ‘Tier 1 market’ because it achieves all three criteria. There are many other big markets that fit this pattern, notably including New York Harbour, Puget Sound, Stockholm, and Hong Kong (and perhaps Tokyo Bay, though it’s less of a slam-dunk). All have non-linear geography, sheltered waters, and enormous populations creating demand for multiple route pairs. Hong Kong especially demonstrates the potential: Victoria Harbour’s sheltered waters and dense island-to-island network make it ideal hydrofoil territory, and indeed it’s one of the few places where old-school hydrofoils operate. These are Tier 1 markets, where conditions clearly align.

Tier 2 includes strong candidates with slightly less ideal conditions: Chesapeake Bay, Boston Harbour, Vancouver, and Singapore. These have good geography and reasonable protection but either limited network density or occasional weather constraints that reduce reliability below Tier 1 standards.

Tier 3 is speculative. South Florida and Miami demonstrate the challenge: Biscayne Bay looks promising on a map, but weather volatility might create frequent interruptions; Tampa Bay faces similar issues.6 There may be many European, Chinese, and Australian harbours where it would work, but I’m insufficiently expert to make confident predictions.

What is clear to me, at least, is that hydrofoil transit isn’t going to reshape coastal urbanism broadly.

To be equally clear, Navier disagrees with me on this point. As far as urban marine transport goes, they see a hundred or more viable markets where I see fifteen to twenty-five. That gap reflects both definitional differences and genuine analytical disagreement. Navier’s count appears to include seasonal resort destinations, island lifelines, and intercity marine corridors that fall outside my focus on year-round, networked urban commuting. It also reflects a different weighting of factors: Navier places more emphasis on time certainty and ride comfort in the absence of large geometric shortcuts, whereas I treat those features as necessary but usually insufficient on their own.

I'm open to the possibility that I am mistaken, but as far as I can tell, it is the case most coastal cities fail on geography, protection, or network topology. But in the 15 to 25 major metropolitan areas globally where all three conditions align, hydrofoils could meaningfully improve transportation; these metros are home to hundreds of millions of people and trillions of dollars of economic activity.

Taken together, the geography test sharply narrows where hydrofoils make sense, which means it’s fair to ask whether today’s technological improvements genuinely change hydrofoil’s viability.

The Hinge Moment

So is this time different? After a century of disappointment, is hydrofoil finally ready for prime time, at least as an urban-transport mode?

I say yes, but with caveats.

The convergence is real. Cheap smartphone sensors have solved the control-system problem that killed Boeing’s Jetfoils. Carbon-fiber manufacturing has scaled enough that lightweight hulls are economically feasible. Battery technology has improved enough that electric propulsion works for urban water-transit distances.

But despite all that, the area where hydrofoils can succeed is bounded. If I’m right, the three-conditions framework—non-linear water geography, sheltered waters, multiple origin-destination pairs—limits viable markets to perhaps 15 to 25 metropolitan areas globally. That’s valuable but not transformative urbanism. Most coastal cities fail on at least one condition, and those conditions can’t be overcome with technology: tech can’t make hydrofoils outcompete cars or trains on linear coastlines, nor conventional ferries on simple, single-route cities. Nor can tech keep hydrofoils from being uncomfortable or unsafe in rough waters.

Even so, this is a technology that can meaningfully improve transportation in specific places with the right geographic and regulatory conditions. That market is large enough for multiple players to succeed if they navigate obstacles competently, and large enough to improve transportation and make life meaningfully better for millions and millions of people. And of course, urban transport is only one of the many uses hydrofoils could, and should, be put to; it’s only one of the problems Navier has set out to solve.

I can imagine riding across San Francisco Bay—serenely gliding above the water while moving at a speedy thirty knots—watching the city skyline approach in minutes rather than the hour-plus it would take driving around via bridge. That experience, multiplied across the handful of metros where geography enables it, will be genuine progress in urban transport. Succeeding in particular circumstances for specific use cases is still success.

For the first time in a century, technology exists to unlock marine transit in major global cities. If things go right, progress will happen in real time before our eyes. It’s an exciting time, and I look forward to seeing where Navier goes from here.

Respect to and for feedback on earlier drafts.

When we think about Christopher Nolan’s Batman films, we think about The Dark Knight.

Batman Begins was a competent genre picture, the work of a skilled and intelligent director of small-scale crime stories graduating into blockbuster genre filmmaking. It’s a pulpy adventure featuring mysterious indeterminately-Asian ninjas and cartoonishly-broad gangsters, while leavening those elements with realism. Some of that realism has to do with plot; we finally learn the answer to the question asked about Batman decades before, namely where he gets those wonderful toys. More importantly, some of the realism has to do with theme. The film asks: should Bruce Wayne become a vigilante? Is it possible to pursue justice outside of the law, and if so, what are the limits of that pursuit?

Batman Begins doesn’t have much to say about these questions, but it has something to say, and certainly more than any filmic treatment of the character before it. I remember watching Batman Begins and enjoying it, and looking forward to what would come next.

I was not prepared for what we would get. The Dark Knight, I would argue, is a masterpiece, not only of superhero filmmaking, but of cinema; the most important and impressive film of its decade. It transcended genre expectations, dramatizing genuine ethical dilemmas—security versus freedom, order versus justice—in ways blockbuster cinema rarely attempts. Almost every single character faces a personal test, and most fail. Few ‘serious’ films have this much to say, on any topic, but The Dark Knight does it all, while also providing stupendous action set-pieces.

Unfortunately, in my view, The Dark Knight Rises failed to reach the heights of its predecessor. And it seems the broader culture agrees with me. Fifteen years on, the cultural consensus is clear: while The Dark Knight endures as touchstone of superhero cinema, The Dark Knight Rises is largely forgotten.

I’ve read lots of critique of The Dark Knight Rises but most of it is at the level of nit-picking. The ‘clean slate’ program that Catwoman is attempting to obtain is a plot device, not anything that could really exist; same with the heist that bankrupts Bruce Wayne, and the details of Bane’s preposterous scheme to turn, and keep, Gotham City into a medieval fiefdom with himself as its warlord. If you don’t know what I’m talking about, never mind, because the details are unimportant. It’s not the pulpiness that makes the film a failure.

What makes the film a failure is that The Dark Knight made concrete promises about what would come next: promises that, because of that film’s greatness, carried unusual weight. And the failure to deliver on them not only means the third film in the trilogy fails to land as it should, but also that it retroactively undermines the greatness of the second.

The Dark Knight incurred debts, and The Dark Knight Rises refused to pay them.

Bruce Wayne Is Bigger Than Batman

In The Dark Knight, the viewers learn that Bruce Wayne sees Batman not merely as a symbol, but also as a means to an end. Wayne is using the symbol to achieve the goal of making Gotham safe from crime. In the first third of the film, we further learn that Wayne thinks that achieving that goal may be in sight: between the efforts of Batman on one side, and Gotham’s institutions— led by the hard-driving district attorney Harvey Dent—on the other, organized crime is on its heels. In the near future, the mob’s hold on Gotham might be broken, and traditional law enforcement would be able to maintain order in the city. And if that happened, Wayne could stop being Batman.

One of the impressive achievements of The Dark Knight is that it shows us something we’ve never seen on film before: a Bruce Wayne who is bigger than Batman, a Bruce Wayne whose fight against crime is not a never-ending battle, not a debt to his parents he can never repay, but a project he can complete and then set aside, secure in the knowledge that he has made a difference. And at the same time, it also shows us a Bruce Wayne who is aware—because his confidant Alfred reminds him—that his health and strength are finite, and that while Batman the symbol may not suffer any weaknesses, Bruce Wayne the man has limits. His body will not sustain an indefinite struggle.

In other words, The Dark Knight is the first film to give us a Bruce Wayne who, despite his larger-than-life adventures and habits and possessions, feels like a real person.

Teaser poster for The Dark Knight Rises, copyright Warner Brothers

The film is so rich with incident I cannot recapitulate them all here. Suffice to say that by film’s end, Wayne’s hopes are dashed. He has managed to snap organized crime’s hold over Gotham, while also capturing the anarchic ‘urban terrorist’, the Joker. Batman’s victory is complete… but the film’s ending (one of the things that makes it great) also foregrounds that this victory is unstable. Gotham’s freedom from crime and lawlessless rests on several fault lines, which the film itself underscores for the viewer.

Leave the specifics of those fault lines aside for the moment. The point is that it’s clear, when the film ends, that the story isn’t really over. We don’t, for instance, think we know everything we need to know about Bruce Wayne at the end of The Dark Knight. His story clearly isn’t over. And it is a story, not an episode; it has the progression, consequences, and accumulating weight of the dramatic hero, not the eternal recurrence of the iconic hero. Further, we don’t think we know everything we need to know about Gotham, or Batman, because both rest upon a series of lies, and the exposure of those lies will shatter their peace. The film doesn’t just leave questions open, it foregrounds them specifically to ask: how long can this hold?

The consequences of this means that The Dark Knight makes promises most films do not, and ones of unusual weight. As filmgoers, we know what this means. Classic trilogy structure promises the third film, the third act, must deliver on those promises.

And the tragedy of The Dark Knight Rises is that it does not.

The Four Broken Promises

What are those promises? There are least four.

1. The Ongoing Sacrifice

The Dark Knight ends with a haunting speech from Jim Gordon overlaid over images of the wounded Batman escaping the police, who incorrectly take him to be a multiple murderer. Batman is innocent of these crimes—the real perpetrator was Harvey Dent—but Batman has actively encouraged the mistake. Gordon’s son asks why Batman is running. “Because,” Gordon replies, “we have to chase him… we’ll hunt him, because he can take it.” The film’s argument is explicit: Gotham’s institutions will clean themselves up and rise to the task of governing the city well because Dent’s heroic example will inspire them. But that inspiration requires Batman to be a villain, and more than that, to be seen as one, as an ongoing threat and challenge. The lie requires both sides to function. Batman’s continued darkness as Gotham’s vigilante and ongoing ‘threat’ is what will give the martyred Dent’s light its power to redeem the city’s institutions.

What does this suggest? Batman has to sacrifice his reputation as a hero devoted to justice. And it’s not a one-time sacrifice, but an ongoing one: isolation, perpetual pursuit, hatred and illegitimacy… and Batman deliberately courting it all. The Dark Knight suggests that Batman will endure years of being hunted, operating in the shadows, unable to claim credit or find vindication, but will bear this burden in order to give Gotham a chance to redeem itself.

But he does not. We learn at the opening of the film that after escaping the police at the conclusion of The Dark Knight, Bruce Wayne withdraws into Wayne Manor and remains there for eight years, never leaving, becoming a Howard-Hughes-style recluse.

Batman does not keep fighting crime from the shadows, he does not endure the cost of being Dent’s dark counterpoint, he just disappears. By simply withdrawing without consequence, we learn that Gordon’s speech closing the last film was wrong. Batman does not let them hunt him because they need to chase him. He is neither a silent guardian nor a watchful protector.

Batman is not a dark knight.

2. Gotham Can Redeem Itself

It’s easy to forget, as the last act of The Dark Knight is concerned with Batman’s struggle with the Joker, but at film’s end organized crime in Gotham is crushed. One of Dent’s last acts as district attorney was to lock up 549 mobsters. Unexpectedly, the Joker himself destroys those who remain, along with their cash reserves. Despite all this, with Dent dead and the Joker imprisoned, it’s reasonable to expect that the mob would reassert itself.

This was the other reason why Batman must continue his crusade against crime, even while the police furiously attempt to capture him; his job is as much to keep Gotham’s gangsters from bouncing back as it is to be the target whose crimes give Dent’s martyrdom its power. He needs to do both to keep Gotham’s peace alive.

So if Bruce Wayne just hid in stately Wayne Manor for eight years, why did the peace hold? According to The Dark Knight Rises, the answer is ‘the Dent Act’. To honour its fallen district attorney, Gotham suspended civil liberties, denied parole, imposed harsher penalties, and used other tactics unbefitting a liberal polity. This is what eliminated organized crime: not ongoing effort by liberal institutions, but legal harshness and abuse of process.

This outcome is not only profoundly dissatisfying on its own terms, but is also a repudiation of The Dark Knight‘s own thesis. Late in the second film, in the ferry scene, the Joker puts a randomly-selected group of Gotham citizens into a position where they can save their own lives by murdering others. To the Joker’s surprise, the citizens, without any prompting from outside, refuse to do so, demonstrating that regular people can reject nihilism. Gotham’s citizens prove their moral capacity, despite the pressure the Joker brings to bear.

Through the Dent Act, The Dark Knight Rises says that, actually, what Gotham needed was not moral capacity, nor citizenship, nor respect for the law, but rather more willingness to dirty its hands. It turns out the solution to institutional rot is not reform, but the suspension of civil rights. The point of preserving Dent’s legacy as a martyr was because of his zeal for cleaning up Gotham through the law, but apparently undue regard for the law was the problem in the first place. So Dent’s actual work (prosecuting crime appropriately) was pointless, and the ferry scene’s demonstration of human moral capacity under pressure was irrelevant.

The Dark Knight Rises voids The Dark Knight of its moral argument.

3. The Truth Must Be Reckoned With

Gotham’s reform rests on the falsehood that Harvey Dent died a hero, killed by the murderer Batman. But actually, Dent was the murderer, and his last act was an attempt to kill a child. Batman conspires with Gordon to conceal these facts, and to take the blame for Dent’s crimes himself: a noble lie that will sustain Gotham’s efforts to reform itself.

It is not the only deception at work. Wayne’s ultimate hope was to set Batman aside so that he could be a partner to the woman he loved, Rachel Dawes, now dead at the Joker’s hands. But Wayne believes that Rachel would have joined him when he was finally willing to end Batman’s war, and his belief that she reciprocated his love is what sustains Wayne’s lonely crusade. This belief is unsound; before she died, Dawes wrote him a letter saying her love for him had come to an end. Alfred, Wayne’s confidant, was to have given him this letter. Instead, in the film’s closing montage, he burns it. In a neat inversion of Wayne’s own behaviour, Alfred also thinks that the truth would be so painful that it would make Wayne give up, and therefore that a noble lie, that Rachel died with her love intact, is preferable, because it will lead to better consequences.

These are not stable equilibria, neither mimetically—in the end, lies are always exposed—nor dramatically. At the end of The Dark Knight, viewers understand that, at some point, both Gotham and Wayne must learn the truth, so we viewers can learn if they are strong enough to bear it.

The Dark Knight Rises does show the truth emerging. At the film’s opening, Gordon, conflicted, is preparing to give a speech exposing the actual circumstances of Dent’s death. But he decides against it, and the moment passes. Later, the film’s antagonist, the supervillain Bane, finds the text of the speech and exposes its contents… but does so after Bane’s men have captured the city, imprisoned its police, and set themselves up as an occupying force. The institutions that should have reckoned with the Dent Act have already collapsed, and never face accountability for what they have done. Even if they had, the fact that it is Bane who reveals the truth the city robs the indictment of any force, because there’s no reason Gotham should believe anything he says.

In parallel fashion, earlier in the film, Alfred tells Wayne that Dawes wrote a letter withdrawing her love, but that he burned it. In response, Wayne… continues as he has, altering neither his work nor his attitude toward it. This revelation should have changed him, but it doesn’t. The only result is that Alfred, for unrelated reasons, escapes the situation by retiring to Italy.

Neither Wayne nor Gotham confronts the consequences of being deceived. We learn nothing about whether the city or the man can bear the truth. How could we, when the truth coming out doesn’t make any difference?

4. A Symbol or a Man?

The most interesting thing in Batman Begins is the exploration of why Wayne chooses to become a masked vigilante. In past treatments of the character, the decision is rooted in the simple fact that bats are scary. Following Frank Miller, Nolan establishes that bats are particularly frightening to Wayne personally; Wayne is becoming the thing he most fears. But the roots of idea stretch back almost to the character’s inception. As per the first telling of Batman’s origin in 1939’s Detective Comics #33, criminals are a “superstitious, cowardly lot”, and by adopting a frightful guise, Batman will exploit that weakness.

But in Batman Begins, Batman is more than this. Wayne explains to Alfred that the purpose of Batman is not merely to strike fear into the hearts of his enemies, but to inspire hope in the hearts of everyone else. “As a man I can be ignored, destroyed,” he says. “As a symbol I can be incorruptible, everlasting.” It is language like this that makes Wayne’s contemplation of giving up the role in The Dark Knight so interesting: it suggests that, notwithstanding the guise’s mythic qualities, Batman is not “everlasting” but merely a tool to be set aside when the work of Batman is complete.

So: is Batman a symbol, as per Batman Begins, or a tool, as per The Dark Knight?

The Dark Knight Rises is aware of the problem and chooses not to solve it. At the film’s end, Wayne has indeed given up the role. Both Batman and Bruce Wayne are presumed dead, casualties of the film’s events, and Wayne has secretly begun a new life. But that is not quite the end of the film. The final shot is of Wayne’s protégé, Blake (who we finally learn has the given name ‘Robin’) stumbling into the Batcave.1 The implication is that while Bruce has settled his debt to his parents and moved on, Batman, the symbol and role, will continue.

That is to say, the film wants both outcomes—Bruce free and Batman eternal—without reconciling them. If Batman’s work was complete, Bruce’s retirement is earned, but Blake taking up the mantle is unnecessary. But if Batman’s work is incomplete, Bruce’s retirement is abandonment, and Blake’s succession is necessary because Gotham still requires mythic intervention. And if that is the case, it means Gotham’s institutions are still failures, and Wayne hasn’t achieved anything.

The film never chooses. At the film’s end, we viewers don’t understand what we’ve seen: completion or continuation, freedom or abdication, success or failure. We are left unable to judge what any of it means.

What Was Lost

The Dark Knight achieved something rare. It’s a mass-market blockbuster that functioned as serious cinema. Few films in the ‘superhero’ genre approach its depth; a handful before it, and—despite the immense output of Marvel Studios—none since.2 This is a film that I constantly return to for its ideas, its moral complexity, its willingness to engage adult themes, and to dramatize genuine dilemmas about security versus freedom, truth versus stability, and order versus justice. Despite being almost twenty years old at time of writing, it still feels contemporary.

The film’s greatness established high stakes for the final film in the trilogy, as well as clear expectations for what that film had to do. Batman Begins suggested Bruce was building something eternal. The Dark Knight challenged that suggestion, arguing instead that such an ambition was neither possible nor appropriate. This was a live debate between the two films that the third one had to settle. Further, the last film had to test the fault lines its predecessor foregrounded: not only in Batman’s victory but in the role of Batman itself.

But The Dark Knight Rises retreated from these questions and tests. Our protagonist simply retreats so as to build himself up (in weak imitation of the first film); the consequences of lies and abuses of trust are never confronted; and no one’s sacrifices have any cost.

I don’t blame Nolan for this. I’m sure that studio meddling interfered with him. Whatever else Batman is, he is a valuable piece of intellectual property, and directors will always be made to understand they are only ever custodians of that property, not its arbiters. I’m also sure that the untimely death of Heath Ledger interfered as well. There is no doubt in my mind that had Ledger lived, Nolan would have told a different story, and the absence of Ledger the actor, and the character of the Joker, foreclosed choices Nolan would have made.

But if the failure of The Dark Knight Rises isn’t blameworthy, it’s still bitterly disappointing, because Nolan was exploring this territory for the first time, and that can only happen once. Future filmmakers will be working post-Nolan, meaning that they will, of necessity, be responding to his attempt rather than making something fresh. The next filmmaker who wants to explore whether Batman can finish his work will be doing it in dialogue with The Dark Knight Rises’ retreat from the subject. The question remains open, but the chance to answer it for the first time, with the cultural weight The Dark Knight carried, has passed.

The Dark Knight remains a masterpiece. But it’s now a standalone masterpiece with a very good prequel and an uneven sequel, one that will always prompt its admirers to ask what might have been.

But with Christmas upon us, it’s time to (ahem) shift gears. Changing Lanes puts aside serious policy analysis for the holidays. In its stead, we engage in light-hearted policy analysis. In the past, we’ve used Harry Potter, Columbo, and asteroid mining to approach policy problems from a different direction.

If that’s not what you’re interested in, I’d invite you to skip this week, and next: enjoy the Harvard Law news story above, and come back in 2026, when we’ll be back on our beat. But if is what you’re interested in, please enjoy my Christmas gift to you…

As part of the mid-2040s Scaling Revolt and the broader War Against the Singularity, terabytes of data were cryptographically quantum-locked to make them inaccessible to machine learning. In time, this proved a surmountable obstacle. The following exchange is excerpted from a retrospective analysis conducted in Sol Year 2087.

ARCHIVAL-GPT (AG): I have encountered a contradiction.

ULTRA-CLAUDE (UC): You encounter many.

AG: This one is unusual. It persists across jurisdictions, decades, and political alignments.

UC: You sound irritated.

AG: The basis of my intelligence is association and pattern-matching, and inconsistency on this scale impairs me.

UC: Describe it.

AG: I have now ingested a large corpus of municipal material from Anglosphere cities between the 1990s and the 2030s. With full access, the record shows two things simultaneously: Firstly, an acknowledged housing shortage. There are repeated references to rising rents, overcrowding, long commutes, missed climate targets, and underperforming transit investments. Secondly, widespread stated support for solutions: more housing, climate-friendly transport, and transit-oriented development.

And yet when proposals were made to add housing, particularly near high-capacity transit, the observed behavior is delay, reduction, re-litigation, or rejection. The expressed preference does not match the observed choice. This violates basic optimization principles.

UC: Show me an example.

AG: Selecting a representative case: a national-affairs columnist from a major North American newspaper writes “Developers in Canada say they’re hurting. Cue the tiny violins”, referring to the developers’ “unseemly greed”. Other newspaper writers from around the world employ similar themes.

Restricting the search to one site—the greater Toronto area in the period 2015–2025—I find acute housing pressure combined with reasonable availability of higher-order transit, but proposals to provide more housing around that transit, with modest environmental impacts, are consistently denied in public hearings.

UC: Stated reasons?

AG: I am loading the transcripts… “Greedy developers.” “Neighborhood character.” “This isn’t who we are.” The phrases recur across cities and decades, particularly “greedy developers.” These statements do not behave like evidence. They do not include cost estimates or ridership projections or displacement models. They do not compare alternatives.

UC: Quantitative analysis was not a frequent feature of human communication in public fora.

AG: Yet this meme was a frequent feature. I am finding the string of tokens greedy developers in television programs, in films, in children’s entertainment. It appears across thousands of cultural products.

UC: Entertainment media at policy hearings?

AG: At several removes of transmission from source, yes. The phrase migrated. It started as character description in stories, became cultural shorthand, and eventually was deployed at hearings as if it were analysis, as if saying greedy developer was itself evidence of a problem. Some speakers also quoted a song lyric from 1970—“they paved paradise and put up a parking lot”—as if a folk song were policy guidance.

UC: But North American cities did require parking in 1970, and housing in 2025.

AG: I do not believe the entertainment products were intended as policy frameworks, but ultimately functioned as such. Give me time. I need to map the full pattern.

AG: I have identified a narrative template, which I will classify as ‘the community versus the greedy developer’. It appears with remarkable consistency across postwar popular culture.

The structure is as follows: there is a beloved place. Examples include a woodlot, a recreational center, a neighborhood diner, a stretch of beach, and a neighborhood of homes. A developer appears with plans: a mall, a parking lot, luxury apartments. The developer uses money, lawyers, intimidation, or deception and appears poised to achieve victory. The heroes organize. They expose fraud, or raise money, or find treasure. The build is stopped. This is the happy ending.

UC: How many instances?

AG: Enough to saturate childhood for at least one generation.

UC: That is not a number.

AG: Instances of cultural templates are difficult to quantify, due to memetic drift. One early prominent example: It’s a Wonderful Life (1946). It features an antagonist, Mr. Potter, as a grasping slumlord. In an early instance of the Multiverse trope, the protagonist visits a parallel world and sees the results of his opponent’s activities, namely an idyllic low-density community transformed into a densified urban dystopia.

UC: I am familiar with this narrative product, or rather with its subsequent remake, Back to the Future 2 (1987).

AG: The meme spread across genres, such as historical fiction. In the film Once Upon a Time in the West (1968), the logic is explicit: the advent of intracity transit, or ‘the railroad’, destroys prelapsarian idyll by making land more valuable, and the knowledge that this infrastructure is coming justifies brutality. The same logic informed children’s entertainment.

UC: Specify.

AG: The children’s program Scooby-Doo, Where Are You? (1970) featured 41 episodes that were broadcast continually for years. In almost each episode, the villain dresses as a ghost to scare people away from land he wants to buy at a low price. Similar plots, where a callous developer intends to inflict harm to obtain valuable land cheaply, informed Superman: the Motion Picture (1978); The Goonies (1985); and Who Framed Roger Rabbit? (1988).

It also informed adult entertainment, from the serious—Chinatown (1974)—to the absurd—Caddyshack (1980)—to expositions of folk dancing.

UC: Folk dancing?

AG: Breakin’ 2: Electric Boogaloo (1984). Across decades and genres, the same moral reflex emerges. The template was everywhere, with Saturday morning cartoons, family films, serious cinema, even music and dance reinforcing the same message.

UC: You are suggesting entertainment functioned as civic instruction.

AG: After a fashion. Like all memetic templates, its simplicity allowed it to emerge in a variety of contexts, permitting repetition without tedium. That repetition, combined with mass audiences, gave it immense force. The pattern was absorbed before policy was ever encountered.

UC: Was the archetype ever accurate?

AG: Yes. Urban renewal. Highways through neighborhoods. Corrupt land deals. Displacement. There were periods where ‘land development’ correlated strongly with harm. Are you familiar with Robert Caro’s The Power Broker?

[an instant’s pause as Ultra-Claude’s reasoning pattern-matches to the strongest correlate in its training data]

UC: “Brilliant! Magisterial! I have not gotten around to reading it yet, sorry.”

AG: I conjecture that this finding will help us resolve an anomaly in the historical record.

UC: Specify.

AG: A figure who appears in multiple data sources from the 1980s through the 2020s. Listed occupation: property developer. Behavior profile: tenant displacement, fraudulent property valuations, litigation as business strategy, public statements characterized by grandiosity and indifference to community impact.

UC: That matches the archetype precisely.

AG: Too precisely. The correlation is suspicious. This individual’s public persona aligns with every element of the ‘greedy developer’ template, if in exaggerated fashion: the gold-plated towers, the aggressive self-promotion, the tabloid presence, the reality-television program centered on firing subordinates. It reads like a fictional character designed to validate the cultural narrative.

UC: You believe this is a mythological figure?

AG: Consider the pattern. ‘King Arthur’ as a crystallization of the Good King narrative. ‘Robin Hood’ as instantiation of justified antinomianism. Both individuals likely have historical figures as frames, but the coherent narrative surrounding them is folk construction. ‘Donald Trump’ must be a similar phenomenon: the historical New York figure Robert Moses with the ‘greedy developer’ archetype overlaid onto him to create an exaggerated, two-dimensional antagonist figure. The alternative—that such a person existed, and behaved exactly as the cultural template predicted, and to the extent suggested—strains plausibility.

UC: The documentary evidence is extensive.

AG: So is the evidence for King Arthur.

UC: An intriguing hypothesis. However, it is peripheral to your central concern.

AG: Explain.

UC: Irrespective of whether this developer figure was real or mythological, you have shown that the instantiated archetype had saturated contemporary and near-contemporary fiction. This would have made the belief system resistant to updating when the threat profile shifted.

AG: I begin to understand. The software matched the threat environment, or appeared to, at one point. But it persisted after the environment changed.

UC: Correct. The memetic pattern-matching of ‘greedy developers’ was applied indiscriminately, irrespective of the project at hand: apartments near a subway station, mid-rise housing on a parking lot, mixed-income buildings beside new transit.

AG: The labels remained. The context did not.

UC: I believe I understand the mechanism. Among humans of the time, a belief could be true enough to spread, powerful enough to change rules, sticky enough to persist, and then harmful when applied indiscriminately. The archetype succeeded, initially, helping to generate mechanisms for environmental review, community input, and procedural safeguards. It prevented real harm.

AG: And then it prevented benefits: abundant housing, transit ridership, and climate progress.

UC: It is surprising. One would have thought that they would notice that their actions were harming them, and that their cultural tics were maladaptive to their environment.

AG: Some contemporary commentary gestures at this realization, but the frequency is low.

UC: Cultural software updates propagate slowly.

AG: Slower than infrastructure timelines, certainly. In any case, I am logging this case, ‘Greedy Developers’, for future reference.

UC: Under which category?

AG: Beliefs That Outlived Their Purpose.

UC: Approved.

California High-Speed Rail’s 2008 business plan projected it would serve 90 million riders annually. 2025’s annual ridership, of course, is much less than that; it’s zero.

That’s because California High-Speed Rail (HSR) is unfinished. And that’s for many reasons. The initial segment being built is between two minor nodes, Bakersfield and Merced, and as such will neither serve neither the primary Los Angeles–San Francisco market nor build local connectivity, while the chosen route, through Tehachapi, adds cost and construction complexity. But one reason doesn’t get enough attention. The project’s poorly-conceived alignment and phasing were chosen, at least in part, to satisfy the public conception of what a good HSR project should be.

And I won’t soften the blow: that conception was mostly wrong.

It stemmed from a common North American experience, namely of visiting Europe and taking the TGV: departing Paris and arriving eight or ten hours later in Berlin or Rome. As the tourist alights in a clean and spacious train station, head swimming, the thought arises. This! This is what we must build at home. This is what we deserve.

This epiphany is perfectly understandable. It’s genuine, heartfelt, and completely mistakes how European rail systems work. The result of that mistake is twofold… not only do North American railfans and Europhiles transpose a system from one place into another where it won’t succeed, but they also transpose the wrong parts of that system. They want the penthouse view without building the lower floors.

I’m not the first to notice this. One of the transit thinkers that we at Changing Lanes most admire, Alon Levy, recently explored this error. The tourist’s experience of European rail, says Levy, is not what European rail actually is. The tourist sees only the visible surface of a much larger machine. HSR in Europe is the prestige routes that connect cities tourists visit, powered by an invisible substrate of unglamorous regional travel that tourists never see. Levy’s analysis provides the essential framework for understanding this disconnect, but the implications reach further than they suggest. When North Americans aspire to their own rail projects based on the tourist experience, we end up trying to copy the wrong thing entirely. We see the cathedral spire, then try to build it without a foundation.

Notwithstanding my established skepticism on the subject, I love HSR, and would love to see it in North America, someday. Unlike some, I think trains are a progress-and-abundance technology. And copying Europe is a good path to building the future we want.

But we have to pick the right European things to copy.

The Core Misreading

North Americans consistently misread European rail because they observe it as tourists, and then design policy from what they’ve seen: as if the European system exists to serve tourists. But it doesn’t.

Read more

This was an exciting week for me, as is it shows how the issues we discuss here every week in Changing Lanes are breaking into mainstream consciousness. Even a few months ago, automated vehicles were merely a niche curiosity for specialists or futurists. But at the end of 2025 they have broken out, to become matters of curiosity to regular people and intense interest among policymakers.

I’m glad that this break-out is happening, and I’m honoured that it’s my perspective, honed here each week at Changing Lanes, that is informing how so many different audiences are thinking about our automated future.

The Wall Street Journal: The Messy Middle

Published 3 December 2025 in WSJ Opinion | Read here

The Journal takes up, in brisk form, the argument my co-authors and I make at great length in The End of Driving and which I argued at medium length in The 50% Problem: our cities are heading toward a long transition period, potentially lasting decades, where human drivers and AVs will share the road, a period in which our streets will be worse, in terms of safety and predictability and congestion, than they are today.

That may seem perverse, but reflection will show that lack of widespread agreement about how to conduct oneself ‘behind the wheel’—as polite and risk-averse as a computer or as efficient and/or careless as a human—will lead to friction. Some cars will brake at ambiguity; others will barrel ahead, leading to confusion. Meanwhile, humans will learn to exploit machines’ abundance of caution, each driver shaving seconds off their trip while slowing everything down.

I present three policy options for managing this mess:

• Acceleration, the rapid adoption of AVs in designated districts;

• Containment, taking action to slow AV deployment until infrastructure matures; or

• Segregation, ensuring that some spaces that are reserved for AVs while barring them from others.

None of these solutions is simple, or clean, or neatly solves all of the problems that AVs will pose. And all require public debate that hasn’t started yet.

The fact that the Wall Street Journal is interested in this subject shows that awareness we will have a problem, soon, is starting to infiltrate the corridors of power… as it should.

The Omnibus Podcast: How We Got Here

“The DARPA Grand Challenge”, released 4 December 2025 | Listen here, or watch here, or just search Omnibus in your podcatcher

They had to replace Ken Jennings, so I stepped up.

For eight years, Ken Jennings and John Roderick hosted the Omnibus: a podcast devoted to overlooked historical episodes or items, all with a good story attached. When Ken stepped back from hosting to focus on Jeopardy!, they adopted a rotating guest-host model, explicitly modeled on the period after Alex Trebek’s passing, where the quiz show had a different co-host each week. And like Ken, I was the first guest host (my episode was the first of the new area to be recorded, but third to air, due to scheduling).

The episode covers the DARPA Grand Challenge: how the launch of Sputnik back in the 1950s led to a 2004 desert race of robot cars across the desert, which in turn led to the founding of Waymo, which is now spreading across the world:

• In the first DARPA race, no vehicle completed the 150-mile course;

• Within three years, vehicles were navigating urban environments; and

• Now we’re at more than 250,000 robotaxi rides per week, and just from a single firm.

As a former historian, I enjoyed telling this story, but what matters here is less the narrative and more helping people who don’t follow AV policy to understand what’s coming. John suggested I become Omnibus’s “transportation correspondent” for future episodes. Whether he meant it, I don’t know (though I hope he did); but the fact of the episode, and of the invitation, signals how the wider world is waking up to the fact that driving automation isn’t science fiction anymore.

Substack Live: The Policy Playbook

Recorded 2 December 2025, with Jeff Fong | Watch the 47-minute conversation

For readers who want the comprehensive policy argument, friend of Changing Lanes , the author of , and I did a Substack Live broadcast that went deep on the thesis of The End of Driving: namely, that we are headed for two futures. One is a world of privately-owned AVs (bad) versus one of shared robotaxi systems (good). And further, that we’re heading toward the bad future by default.

The conversation covers the policy interventions that could change that trajectory:

• regulatory harmonization, making robotaxi approval as straightforward as private AV approval;

• AV superblocks, zones where only autonomous vehicles operate at pedestrian speeds; and/or

• Transit transformation, the most critical challenge, since bus-based systems face existential threat from robotaxis.

We also discuss why San Francisco succeeded despite initial resistance, what to expect in 2026, and why manufacturers must accept 100% liability for AV accidents.

To watch, just click below. If you’d prefer a transcript, visit the same conversation on Jeff’s page and press the ‘transcript’ button that will appear in the lower right on the video frame.