Hi Andrew, I am impressed enough with some of your writing to reply pointing out several important factual errors that I believe are leading your analysis astray. First relying on poor-quality sourcing gave you completely false information about what Dojo is/was, its relationship to AI training, and Tesla’s in-house AI efforts.
Dojo is a Tesla designed computer chip specifically for AI model training. It is a competitor to Nvidia AI/GPU chip. Other companies doing this are: Meta, Google, Amazon, Microsoft, and OpenAI. (1) All of these companies design the chips and use foundry companies like TSMC or Samsung to build the actual chip hardware.
Tesla’s on-car computers use its in-house FSD Computer. These are AI inference chips. Originally Tesla used a Nvidia chip in their “Hardware 2” when Autopilot 2 first debuted. In 2016 they started work on “Hardware 3” or “FSD Computer 1” and began installing them in all their cars in 2019.(2) You can think of this as Apple switching from Dragon and Intel chips for iPhones and Macs to their “A” series of in-house chip designs. AI3 was replaced with “Hardware4” with “AI4” chips in 2021.(3) Both generations fabbed by Samsung. Next year Tesla is switching to “AI5” built by TSMC. Finally Tesla just inked a #13 billion + contract with Samsung to build their “AI6” chips in Samsung’s new Texas Fab (at 2 nm process) by 2027.(4)
The news a few weeks ago was that Tesla is cancelling the development effort for DOJO version 2 and 3 and using the AI6 chips for their next, next supercomputers in 2027. Remember both are chip designs not computers.
Tesla’s latest supercomputers are Colossus 1 (operating since Dec. 2024) and Colossus 2 (under construction). These are build with Nvidia AI/GPU training chips and Colossus 1 is what is currently used to train new versions of FSD. (5)
So your paragraph:
“The edge for Tesla’s robotaxis was supposed to be its superior driving-automation software, but Tesla has just shut down its Dojo supercomputer project. Musk has positioned Dojo as "the cornerstone of Tesla's AI ambitions" since 2019, claiming it would process "truly vast amounts of video data" essential for achieving full self-driving capability. Its abandonment signals Tesla's retreat from in-house AI development in favour of external partners, which is hardly the mark of a company about to revolutionize automated driving.”
Is wrong. Colossus is not shutting down. FSD AI is not being outsourced. Tesla was iterating on two chips families the AI series and Dojo and they decided to ax Dojo and focus
By the way, look at the list of who does in-house AI chips designs. See any Auto OEMs there? Only the biggest tech/AI firms. Maybe GM and Ford are nothing like Tesla in important ways that you have not researched.
Maybe Elon Musk, who co-founded Open AI and then founded X.ai and took it from zero to the highest scoring AI chatbot in last than three years, might know something more about AI than Mary Barra. Things to think about.
Brian: I appreciate the thoughtful pushback. Some notes:
1) I'm not sure what you're getting at with your Dojo claims as the text cuts off, but I am sure we can agree that Tesla has disbanded the Dojo team, its head is leaving, and that in the past Musk claimed Dojo was one of Tesla's competitive advantages
2) Colossus isn't Tesla, but xAI, built to train Grok. Do you have any source that shows Tesla is using Colossus for FSD? I'm not aware of any
3) I agree that Tesla continues to design new chips while fabbing them elsewhere; Nvidia, and apparently Samsung too
I'll update the Dojo paragraph to add some nuance, but I'm afraid my priors aren't shifting: killing Dojo and relying on outsourced chips isn't consistent with a claim that Tesla has a distinct training moat here... and a distinct training moat is what the firm needs to sustain its valuation premium.
Thanks for catching my mistake, Robert. I switched between devices and linked to the wrong article. My point by point responses (take two after first version accidently deleted):
1) I'm not sure what you're getting at with your Dojo claims as the text cuts off, but I am sure we can agree that Tesla has disbanded the Dojo team, its head is leaving, and that in the past Musk claimed Dojo was one of Tesla's competitive advantages
Dojo is a chip design not a supercomputer. Unlike Waymo of GM (or Ford, Mercedes etc. failed attempts) Tesla is designing custom AI chips for both their automobiles and their AI training. What happened recently is that they decided to give up on Dojo 2 and 3 and instead use their AI6 inference chip in future AI training supercomputers as well as in their cars. Tesla signed a $13 billion+ contract with Samsung for the manufacture of the AI6 chips, include deep Tesla involvement in the manufacturing. No other Autonomous vehicle effort is investing so much.
Further Tesla has been clear that Dojo was a hedge and a maybe for over a year now (1) What does “But I would — think of Dojo as … a long shot. Um, it’s a long shot worth taking, because the payoff is potentially very high, but it’s not something that is a high probability — it’s not like a sure thing, at all. It’s a higher-risk, high-payoff program. Um, Dojo is working, and it is doing training jobs. So, and we are scaling it up. And we have plans for Dojo 1.5, Dojo 2, Dojo 3, and what not. So, you know, I think it’s got potential. [Elon Musk sighs at this point.] But, we can’t emphasize enough: high risk, high payoff.” Sound like to you?
2) Colossus isn't Tesla, but xAI, built to train Grok. Do you have any source that shows Tesla is using Colossus for FSD? I'm not aware of any
Good catch I used a link for Colossus when I should of used this article (2) covering Tesla’s Cortex. That article (and many others show that Cortex 1 is currently training FSD and a bigger Cortex 2 is under construction. These are training the FSD neutral networks and no change to those plans.
3) I agree that Tesla continues to design new chips while fabbing them elsewhere; Nvidia, and apparently Samsung too
No this is a factual wrong statement. Most world leading chip companies are fabless: Nvidia, Apple, Qualcomm, AMD. They design the chips then take the designs to foundry companies such as TSMC to build the actual chips. Some companies such as Samsung and Intel are both chip designers and foundries.
No other autonomous vehicle companies are custom designers. GM tried, but their Origin vehicle was cancelled and Cruise failed. Having custom chips with millions of copies, heading into the third version next year, with a multi-billion dollar contract for the fourth version, is a large advantage. The Austin Robotaxis are estimated at a third to a fifth the price of Waymo's robotaxis. Tesla has nearly a thousand times the scaling capacity over Waymo too (2000/year vs. 3 million/year).
“I'll update the Dojo paragraph to add some nuance, but I'm afraid my priors aren't shifting: killing Dojo and relying on outsourced chips isn't consistent with a claim that Tesla has a distinct training moat here... and a distinct training moat is what the firm needs to sustain its valuation premium.”
This paragraph has multiple factual and logical errors.
1. Outsourcing chip manufacture is standard. Are Apple’s or Nvidia’s custom silicon “not real” because they use TSMC fabs?
2. All leading AI developers rely mostly on Nvidia chips for their training supercomputers. That is industry standard.
3. The data moat has everything to do with the millions of Teslas on the road with cameras and AI computers in them, and nothing to do with Dojo. The data is collected and uploaded by the cars and the connectivity that lets Tesla ping them for usual driving situations. Also, the ability to run new versions in “shadow mode” while the current version of driver is driving. Dojo was just a way to build a faster supercomputer. Tesla is solving that without Dojo see page 8 of their latest earnings deck. (3)
I understand this Substack started with your take on why Tesla cannot succeed. Maybe you think the Waymo way is the only way, and Lidar and essential, and everything flows from there. I suggest that if you dig deeper into what is happening, there is more to Tesla than what your writing reflects.
"Multiple factual and logical errors". Can you point me to these, so I can fix them?
It is indeed the case that outsourcing production of chips is standard, and that very few designers are also producers. In this regard, Tesla's decision not to produce its own chips wouldn't be blameworthy... except that the CEO spent years insisting that making its own chips WAS a competitive advantage. It's not the decision, it's the backtracking. Was Musk wrong then, or is he wrong now?
Sure, you write "...killing Dojo and relying on outsourced chips isn't consistent with a claim that Tesla has a distinct training moat here... and a distinct training moat is what the firm needs to sustain its valuation premium.”
You are stating that relying to Nvidia chips means a company does not have an AI training data advantage. What chips a company uses to train their AI has zero relationship to the quality and quantity of training data. There is zero relationship between the two, yet you are claiming a causal relationship. That is the heart of the logical error.
I already pointed out the factual error that Cortex is the supercomputer, processing "truly vast amounts of video data," not Dojo (a chip design). Also that Dojo 1 was built by TSMC, so "intend to proceed entirely with external silicon (Nvidia and Samsung)" is completely understanding the industry you are talking about. That is like saying Apple is abandoning their Ax series of custom chips when they switched between Samsung and TSMC, because you didn't know they were already using a foundry company to manufacture the chips.
Brian, it was Musk, not I, who insisted that Tesla's reliance on internal chips and Dojo were comparative advantages that would help Tesla to win. My point was, and remains, that Musk made these commitments and then reneged on them, and external observers should make predictions about Tesla's future success accordingly.
starting at 58:30 where he calls it "a high risk, high payoff program" do you have video or a transcript of him saying Tesla's entire AI strategy depends on Dojo?
>and external observers should make predictions about Tesla's future success accordingly.
You completely changed your point now. In the article it was that without Dojo Tesla lost any AI advantage.
Now that I pointed out the factual errors in that claim, you are saying instead that Musk changed his custom chip strategy after six years. Instead of one internal chip design he may use another internal chip design to build a training computer cluster. Somehow changing strategy after six years of immense change means he is an unreliable executive?
Frankly, what you or I think "a good executive should do" is irrelevant. What matters is how effective an executive is over time.
Tesla robotaxis are operating in Austin right now with safely passengers riding in them. Whether that means Tesla is where Waymo was in Chandler in 2017, or SF in 2022 we don't know. The performance of Teslas in Austin and other cities over the next year will tell us.
While the financial fundamentals don't make sense, which has been the case to different degrees for a long time, there are two things you didn't mention that I think are important:
1) Elon's personality attracts top technical talent, and while his reputation elsewhere has diminished, I don't think that has impacts top talent still wanting to work for him. Moths to a flame.
2) The performance, technology, and user experience of Tesla's cars is top notch, including the Supercharging network. BYD and others may come close, or even match/exceed it at some point, but this still a major factor. And while FSD has some issues, having tested it every day for almost a year, I can say it's very good. The edge cases are being addressed quite rapidly, so we'll see if that continues.
I'm not saying the confluence of factors you've mentioned won't necessarily create a doom loop, but I believe these are non-trivial advantages in Tesla's favour.
1) Like so much else, this is true until it isn't; I think lots of people want to work at SpaceX, but the churn at Tesla is non-trivial
2) I drive a Tesla myself, and I use its Supercharger network, and take advantage of FSD from time to time, so I will freely grant the quality of the driving experience. Against this, I haven't driven a BYD because they aren't available in Canada, yet, but I'm not aware of anyone claiming that BYD's performance, technology, nor user experience are worse than Tesla's
But let's be clear: the argument before us is not 'Tesla has no advantages'. It is that 'Tesla is trading far, far above its fundamentals and has done so for particular reasons that don't pertain any longer'. Tesla can still make fine vehicles, as you say, and I think that would be consistent with trading at twice Toyota's value. I remind you it is trading at twenty-five times Toyota's value. My argument is, that will not persist.
Hi Andrew, I am impressed enough with some of your writing to reply pointing out several important factual errors that I believe are leading your analysis astray. First relying on poor-quality sourcing gave you completely false information about what Dojo is/was, its relationship to AI training, and Tesla’s in-house AI efforts.
Dojo is a Tesla designed computer chip specifically for AI model training. It is a competitor to Nvidia AI/GPU chip. Other companies doing this are: Meta, Google, Amazon, Microsoft, and OpenAI. (1) All of these companies design the chips and use foundry companies like TSMC or Samsung to build the actual chip hardware.
Tesla’s on-car computers use its in-house FSD Computer. These are AI inference chips. Originally Tesla used a Nvidia chip in their “Hardware 2” when Autopilot 2 first debuted. In 2016 they started work on “Hardware 3” or “FSD Computer 1” and began installing them in all their cars in 2019.(2) You can think of this as Apple switching from Dragon and Intel chips for iPhones and Macs to their “A” series of in-house chip designs. AI3 was replaced with “Hardware4” with “AI4” chips in 2021.(3) Both generations fabbed by Samsung. Next year Tesla is switching to “AI5” built by TSMC. Finally Tesla just inked a #13 billion + contract with Samsung to build their “AI6” chips in Samsung’s new Texas Fab (at 2 nm process) by 2027.(4)
The news a few weeks ago was that Tesla is cancelling the development effort for DOJO version 2 and 3 and using the AI6 chips for their next, next supercomputers in 2027. Remember both are chip designs not computers.
Tesla’s latest supercomputers are Colossus 1 (operating since Dec. 2024) and Colossus 2 (under construction). These are build with Nvidia AI/GPU training chips and Colossus 1 is what is currently used to train new versions of FSD. (5)
So your paragraph:
“The edge for Tesla’s robotaxis was supposed to be its superior driving-automation software, but Tesla has just shut down its Dojo supercomputer project. Musk has positioned Dojo as "the cornerstone of Tesla's AI ambitions" since 2019, claiming it would process "truly vast amounts of video data" essential for achieving full self-driving capability. Its abandonment signals Tesla's retreat from in-house AI development in favour of external partners, which is hardly the mark of a company about to revolutionize automated driving.”
Is wrong. Colossus is not shutting down. FSD AI is not being outsourced. Tesla was iterating on two chips families the AI series and Dojo and they decided to ax Dojo and focus
By the way, look at the list of who does in-house AI chips designs. See any Auto OEMs there? Only the biggest tech/AI firms. Maybe GM and Ford are nothing like Tesla in important ways that you have not researched.
Maybe Elon Musk, who co-founded Open AI and then founded X.ai and took it from zero to the highest scoring AI chatbot in last than three years, might know something more about AI than Mary Barra. Things to think about.
(1) https://medium.com/%40frulouis/6-tech-giants-dominating-the-2025-semiconductor-ai-chip-race-b9b3dac7e498
(2) https://en.wikichip.org/wiki/tesla_%28car_company%29/fsd_chip
(3) https://www.autopilotreview.com/tesla-hardware-4-rolling-out-to-new-vehicles/
(4) https://techcrunch.com/2025/07/28/tesla-signs-16-5b-deal-with-samsung-to-make-ai-chips/
(5) https://www.teslarati.com/xai-tesla-megapack-colossus-2/
Brian: I appreciate the thoughtful pushback. Some notes:
1) I'm not sure what you're getting at with your Dojo claims as the text cuts off, but I am sure we can agree that Tesla has disbanded the Dojo team, its head is leaving, and that in the past Musk claimed Dojo was one of Tesla's competitive advantages
2) Colossus isn't Tesla, but xAI, built to train Grok. Do you have any source that shows Tesla is using Colossus for FSD? I'm not aware of any
3) I agree that Tesla continues to design new chips while fabbing them elsewhere; Nvidia, and apparently Samsung too
I'll update the Dojo paragraph to add some nuance, but I'm afraid my priors aren't shifting: killing Dojo and relying on outsourced chips isn't consistent with a claim that Tesla has a distinct training moat here... and a distinct training moat is what the firm needs to sustain its valuation premium.
Thanks for writing!
Thanks for catching my mistake, Robert. I switched between devices and linked to the wrong article. My point by point responses (take two after first version accidently deleted):
1) I'm not sure what you're getting at with your Dojo claims as the text cuts off, but I am sure we can agree that Tesla has disbanded the Dojo team, its head is leaving, and that in the past Musk claimed Dojo was one of Tesla's competitive advantages
Dojo is a chip design not a supercomputer. Unlike Waymo of GM (or Ford, Mercedes etc. failed attempts) Tesla is designing custom AI chips for both their automobiles and their AI training. What happened recently is that they decided to give up on Dojo 2 and 3 and instead use their AI6 inference chip in future AI training supercomputers as well as in their cars. Tesla signed a $13 billion+ contract with Samsung for the manufacture of the AI6 chips, include deep Tesla involvement in the manufacturing. No other Autonomous vehicle effort is investing so much.
Further Tesla has been clear that Dojo was a hedge and a maybe for over a year now (1) What does “But I would — think of Dojo as … a long shot. Um, it’s a long shot worth taking, because the payoff is potentially very high, but it’s not something that is a high probability — it’s not like a sure thing, at all. It’s a higher-risk, high-payoff program. Um, Dojo is working, and it is doing training jobs. So, and we are scaling it up. And we have plans for Dojo 1.5, Dojo 2, Dojo 3, and what not. So, you know, I think it’s got potential. [Elon Musk sighs at this point.] But, we can’t emphasize enough: high risk, high payoff.” Sound like to you?
2) Colossus isn't Tesla, but xAI, built to train Grok. Do you have any source that shows Tesla is using Colossus for FSD? I'm not aware of any
Good catch I used a link for Colossus when I should of used this article (2) covering Tesla’s Cortex. That article (and many others show that Cortex 1 is currently training FSD and a bigger Cortex 2 is under construction. These are training the FSD neutral networks and no change to those plans.
3) I agree that Tesla continues to design new chips while fabbing them elsewhere; Nvidia, and apparently Samsung too
No this is a factual wrong statement. Most world leading chip companies are fabless: Nvidia, Apple, Qualcomm, AMD. They design the chips then take the designs to foundry companies such as TSMC to build the actual chips. Some companies such as Samsung and Intel are both chip designers and foundries.
No other autonomous vehicle companies are custom designers. GM tried, but their Origin vehicle was cancelled and Cruise failed. Having custom chips with millions of copies, heading into the third version next year, with a multi-billion dollar contract for the fourth version, is a large advantage. The Austin Robotaxis are estimated at a third to a fifth the price of Waymo's robotaxis. Tesla has nearly a thousand times the scaling capacity over Waymo too (2000/year vs. 3 million/year).
“I'll update the Dojo paragraph to add some nuance, but I'm afraid my priors aren't shifting: killing Dojo and relying on outsourced chips isn't consistent with a claim that Tesla has a distinct training moat here... and a distinct training moat is what the firm needs to sustain its valuation premium.”
This paragraph has multiple factual and logical errors.
1. Outsourcing chip manufacture is standard. Are Apple’s or Nvidia’s custom silicon “not real” because they use TSMC fabs?
2. All leading AI developers rely mostly on Nvidia chips for their training supercomputers. That is industry standard.
3. The data moat has everything to do with the millions of Teslas on the road with cameras and AI computers in them, and nothing to do with Dojo. The data is collected and uploaded by the cars and the connectivity that lets Tesla ping them for usual driving situations. Also, the ability to run new versions in “shadow mode” while the current version of driver is driving. Dojo was just a way to build a faster supercomputer. Tesla is solving that without Dojo see page 8 of their latest earnings deck. (3)
I understand this Substack started with your take on why Tesla cannot succeed. Maybe you think the Waymo way is the only way, and Lidar and essential, and everything flows from there. I suggest that if you dig deeper into what is happening, there is more to Tesla than what your writing reflects.
(1) https://cleantechnica.com/2024/02/08/what-is-going-on-with-teslas-dojo/
(2) https://interestingengineering.com/innovation/tesla-supercomputer-cortex-2-texas
(3) https://www.tesla.com/sites/default/files/downloads/TSLA-Q2-2025-Update.pdf
"Multiple factual and logical errors". Can you point me to these, so I can fix them?
It is indeed the case that outsourcing production of chips is standard, and that very few designers are also producers. In this regard, Tesla's decision not to produce its own chips wouldn't be blameworthy... except that the CEO spent years insisting that making its own chips WAS a competitive advantage. It's not the decision, it's the backtracking. Was Musk wrong then, or is he wrong now?
As for the data moat, I wrote more about that here: https://www.changinglanesnewsletter.com/p/contra-pueyo-on-robotaxis
Sure, you write "...killing Dojo and relying on outsourced chips isn't consistent with a claim that Tesla has a distinct training moat here... and a distinct training moat is what the firm needs to sustain its valuation premium.”
You are stating that relying to Nvidia chips means a company does not have an AI training data advantage. What chips a company uses to train their AI has zero relationship to the quality and quantity of training data. There is zero relationship between the two, yet you are claiming a causal relationship. That is the heart of the logical error.
I already pointed out the factual error that Cortex is the supercomputer, processing "truly vast amounts of video data," not Dojo (a chip design). Also that Dojo 1 was built by TSMC, so "intend to proceed entirely with external silicon (Nvidia and Samsung)" is completely understanding the industry you are talking about. That is like saying Apple is abandoning their Ax series of custom chips when they switched between Samsung and TSMC, because you didn't know they were already using a foundry company to manufacture the chips.
Brian, it was Musk, not I, who insisted that Tesla's reliance on internal chips and Dojo were comparative advantages that would help Tesla to win. My point was, and remains, that Musk made these commitments and then reneged on them, and external observers should make predictions about Tesla's future success accordingly.
>Brian, it was Musk, not I, who insisted that Tesla's reliance on internal chips
>and Dojo were comparative advantages that would help Tesla to win.
> My point was, and remains, that Musk made these commitments and then reneged on them,
[Citation to primary source needed.] Here is a link to Musk discussing Dojo in late Jan. 2024:
https://www.youtube.com/live/A_2Z-AEtdAg
starting at 58:30 where he calls it "a high risk, high payoff program" do you have video or a transcript of him saying Tesla's entire AI strategy depends on Dojo?
>and external observers should make predictions about Tesla's future success accordingly.
You completely changed your point now. In the article it was that without Dojo Tesla lost any AI advantage.
Now that I pointed out the factual errors in that claim, you are saying instead that Musk changed his custom chip strategy after six years. Instead of one internal chip design he may use another internal chip design to build a training computer cluster. Somehow changing strategy after six years of immense change means he is an unreliable executive?
Frankly, what you or I think "a good executive should do" is irrelevant. What matters is how effective an executive is over time.
Tesla robotaxis are operating in Austin right now with safely passengers riding in them. Whether that means Tesla is where Waymo was in Chandler in 2017, or SF in 2022 we don't know. The performance of Teslas in Austin and other cities over the next year will tell us.
While the financial fundamentals don't make sense, which has been the case to different degrees for a long time, there are two things you didn't mention that I think are important:
1) Elon's personality attracts top technical talent, and while his reputation elsewhere has diminished, I don't think that has impacts top talent still wanting to work for him. Moths to a flame.
2) The performance, technology, and user experience of Tesla's cars is top notch, including the Supercharging network. BYD and others may come close, or even match/exceed it at some point, but this still a major factor. And while FSD has some issues, having tested it every day for almost a year, I can say it's very good. The edge cases are being addressed quite rapidly, so we'll see if that continues.
I'm not saying the confluence of factors you've mentioned won't necessarily create a doom loop, but I believe these are non-trivial advantages in Tesla's favour.
Daniel:
1) Like so much else, this is true until it isn't; I think lots of people want to work at SpaceX, but the churn at Tesla is non-trivial
2) I drive a Tesla myself, and I use its Supercharger network, and take advantage of FSD from time to time, so I will freely grant the quality of the driving experience. Against this, I haven't driven a BYD because they aren't available in Canada, yet, but I'm not aware of anyone claiming that BYD's performance, technology, nor user experience are worse than Tesla's
But let's be clear: the argument before us is not 'Tesla has no advantages'. It is that 'Tesla is trading far, far above its fundamentals and has done so for particular reasons that don't pertain any longer'. Tesla can still make fine vehicles, as you say, and I think that would be consistent with trading at twice Toyota's value. I remind you it is trading at twenty-five times Toyota's value. My argument is, that will not persist.
Thanks for writing!