He says supply chains send materials around the world many times, and vertical integration can cut that.
"So with vertical integration, maybe you can only go around the world once."
Why is Tesla building a chip fab?
Q4955 answers2020-2026
Short answer
He said the Terafab costs tens of billions and real production would come in late 2028, which is very fast for a fab.
All 55 answers, oldest to newest
2020 · 1 answer
2023 · 7 answers
He says Tesla builds lithium and cathode refining only because others are not doing it.
"we're doing it because others aren't doing it and we wish others would do it."
He said i should clarify that Dojo V1 is highly optimized for vast amounts of video training vs general purpose AI. Dojo V2 will address these limitations.
"I should clarify that Dojo V1 is highly optimized for vast amounts of video training vs general purpose AI. Dojo V2 will address these limitations."
He said not exactly a competitor to Nvidia, as we aren’t selling Dojo computers (although it could be sold as service). We just need a crazy amount of training compute to get full self-driving 10X safer than human, but can’t ...
"Not exactly a competitor to Nvidia, as we aren’t selling Dojo computers (although it could be sold as service). We just need a crazy amount of training compute to get full self-driving 10X safer than human, but can’t get enough GPUs, plus prices are extremely high!"
He said dojo is arguably the biggest computer architecture overhaul since Cray. Doesn’t mean it succeeds, but it doesn’t lack for boldness.
"Dojo is arguably the biggest computer architecture overhaul since Cray. Doesn’t mean it succeeds, but it doesn’t lack for boldness."
He said there is a lot of merit to having your own chip design. We optimize its usage down to the microsecond.
"There is a lot of merit to having your own chip design. We optimize its usage down to the microsecond."
Asked Do you know if Tesla is doing the same with Dojo, he said we’re going with both Dojo and Nvidia. I think we will have the most inference silicon, given that every car has an inference computer.
"We’re going with both Dojo and Nvidia. I think we will have the most inference silicon, given that every car has an inference computer."
He said managing that data is extremely difficult. We ignore (don’t store at all) or delete almost all of it. The training videos that matter are rare ones, like weird & busy intersections, so less than 1% of all videos. Tesl...
"Managing that data is extremely difficult. We ignore (don’t store at all) or delete almost all of it. The training videos that matter are rare ones, like weird & busy intersections, so less than 1% of all videos. Tesla is limited by AI training compute. Nvidia + Dojo fix it."
2024 · 11 answers
He said the governor is correct that this is a Dojo Supercomputer, but $500M, while obviously a large sum of money, is only equivalent to a 10k H100 system from Nvidia. Tesla will spend more than that on Nvidia hardware this ...
"The governor is correct that this is a Dojo Supercomputer, but $500M, while obviously a large sum of money, is only equivalent to a 10k H100 system from Nvidia. Tesla will spend more than that on Nvidia hardware this year. The table stakes for being competitive in AI are at least several billion dollars per year at this point."
He says Tesla's AI inference efficiency is far better than any other company's.
"There's no company even close to the efficiency of Tesla."
He said correct. Also, I can’t overstate the difficulty of making 50k H100s train as a coherent system. No company on Earth has been able to achieve this yet. Once we have that system working, we will order more hardware.
"Correct. Also, I can’t overstate the difficulty of making 50k H100s train as a coherent system. No company on Earth has been able to achieve this yet. Once we have that system working, we will order more hardware."
Asked thanks for the color. very interesting. off topic, but how do you see Dojo comparing to..., he said training compute for Tesla is relatively small compared to inference compute, as the latter scales linearly with size of fleet. Perhaps the best way to think about it is in terms of power consumption. When the Tesla f...
"Training compute for Tesla is relatively small compared to inference compute, as the latter scales linearly with size of fleet. Perhaps the best way to think about it is in terms of power consumption. When the Tesla fleet reaches 100M vehicles, peak power consumption of AI hardware in cars will be ~100GW. Training power consumption is probably <5GW. These are very rough guesses. Obviously, 5GW of AI training compute is enormous by current standards, but is only about 5% of total Tesla AI compute. There is a path for Dojo to exceed Nvidia. It is a long shot, as I’ve said before, but success is one of the possible outcomes."
He said of the roughly $10B in AI-related expenditures I said Tesla would make this year, about half is internal, primarily the Tesla-designed AI inference computer and sensors present in all of our cars, plus Dojo. For build...
"Of the roughly $10B in AI-related expenditures I said Tesla would make this year, about half is internal, primarily the Tesla-designed AI inference computer and sensors present in all of our cars, plus Dojo. For building the AI training superclusters, NVidia hardware is about 2/3 of the cost. My current best guess for Nvidia purchases by Tesla are $3B to $4B this year."
He said tesla had no place to send the Nvidia chips to turn them on, so they would have just sat in a warehouse. The south extension of Giga Texas is almost complete. This will house 50k H100s for FSD training.
"Tesla had no place to send the Nvidia chips to turn them on, so they would have just sat in a warehouse. The south extension of Giga Texas is almost complete. This will house 50k H100s for FSD training."
He said the existing factory space was already allocated to vehicle, battery and cell production. And this is not a matter of tucking a few computers into a corner. You need a kW of power and cooling for each GPU, which would...
"The existing factory space was already allocated to vehicle, battery and cell production. And this is not a matter of tucking a few computers into a corner. You need a kW of power and cooling for each GPU, which would mean 12MW of power and 12MW of cooling. The south extension is custom-built for heavy power compute and cooling (the cooling towers are huge). Initial system is 50k H100s, plus 20k Tesla HW4 AI computers and massive video storage. Important to note that FSD training requires massive liquid-cooled Tesla AI compute and video storage, not just Nvidia."
He said then HW5, which has been renamed to AI5, in the second half of next year. The Tesla AI5 computer has ~10X the capability of HW4 computer and Tesla makes the whole software stack.
"Then HW5, which has been renamed to AI5, in the second half of next year. The Tesla AI5 computer has ~10X the capability of HW4 computer and Tesla makes the whole software stack."
He said and Dojo 1 will have roughly 8k H100-equivalent of training online by end of year. Not massive, but not trivial either.
"And Dojo 1 will have roughly 8k H100-equivalent of training online by end of year. Not massive, but not trivial either."
In July 2024 he said Tesla aimed to have its next-generation AI5 chip in production at the end of 2025 and scale it in 2026.
"we're aiming to have that in production at the end of next year and scale production in 26."
He said AI5 is ~18 months away from high volume production
"AI5 is ~18 months away from high volume production"
2025 · 22 answers
He said tesla Dojo AI training computer making progress. We start bringing Dojo 2 online later this year. It takes three major iterations for a new technology to be great. Dojo 2 is good, but Dojo 3 will be great.
"Tesla Dojo AI training computer making progress. We start bringing Dojo 2 online later this year. It takes three major iterations for a new technology to be great. Dojo 2 is good, but Dojo 3 will be great."
He said the $16.5B Samsung number is just the bare minimum and actual output is likely several times higher.
"The $16.5B number is just the bare minimum. Actual output is likely to be several times higher."
He said Samsung agreed to let Tesla help maximize manufacturing efficiency at the Texas fab, and he will walk the line personally.
"Samsung agreed to allow Tesla to assist in maximizing manufacturing efficiency. This is a critical point, as I will walk the line personally to accelerate the pace of progress. And the fab is conveniently located not far from my house 😃"
He said Samsung’s giant new Texas fab will be dedicated to Tesla’s AI6 chip, Samsung currently makes AI4, and TSMC will make AI5 initially in Taiwan then Arizona.
"Samsung’s giant new Texas fab will be dedicated to making Tesla’s next-generation AI6 chip. The strategic importance of this is hard to overstate. Samsung currently makes AI4. TSMC will make AI5, which just finished design, initially in Taiwan and then Arizona."
He said most training in the future will actually be inference for synthetic data generation.
"Most training in the future will actually be inference for the sole purpose of training (aka synthetic data generation)"
He said the real-world performance gap between Tesla AI4 and AI5 is far more than any chip version he has ever heard of.
"The difference in real-world performance between AI4 and AI5 is far more than any chip version I’ve ever heard of by a lot. It’s real good."
He said in a supercomputer cluster it makes sense to put many AI5/AI6 chips on a board to cut network cabling cost, and one could call that Dojo 3.
"In a supercomputer cluster, it would make sense to put many AI5/AI6 chips on a board, whether for inference or training, simply to reduce network cabling complexity & cost by a few orders of magnitude. One could call that Dojo 3, I suppose."
He said Tesla should not scale two different AI chip designs, and AI5, AI6 and later chips will be excellent for inference and at least pretty good for training.
"It doesn’t make sense for Tesla to divide its resources and scale two quite different AI chip designs. The Tesla AI5, AI6 and subsequent chips will be excellent for inference and at least pretty good for training. All effort is focused on that."
He said he shut down Dojo because all paths converged to AI6 and Dojo 2 was an evolutionary dead end, while Dojo 3 arguably lives on as many AI6 SoCs on a single board.
"Once it became clear that all paths converged to AI6, I had to shut down Dojo and make some tough personnel choices, as Dojo 2 was now an evolutionary dead end. Dojo 3 arguably lives on in the form of a large number of AI6 SoCs on a single board."
Asked if AI5 is the best chip for cars, he said AI5 will probably be the best inference chip of any kind for models below about 250B parameters, lowest cost silicon and best performance per watt, and AI6 will take that further.
"Could be wrong, but I think AI5 will probably be the best inference chip of any kind for models below ~250B params. By far lowest cost silicon and best performance/Watt. AI6 will take that much further."
He said Tesla's AI5 design review went well, AI6 has a shot at being the best AI chip by far, and they switched from two chip architectures to one.
"Just had a great design review today with the Tesla AI5 chip design team! This is going to be an epic chip. And AI6 to follow has a shot at being the best by AI chip by far. Switching from doing 2 chip architectures to 1 means all our silicon talent is focused on making 1 incredible chip. No-brainer in retrospect. Please join the Tesla silicon team if you want to work on chips that save lives. Milliseconds matter."
He said you can fit more total RAM on a board with ordinary memory than HBM and it is super cheap, so using HBM is not the slam dunk many people think.
"You can fit more total RAM on the board if you use “normal” memory than high-bandwidth memory and it is super cheap. Maybe high-bandwidth memory is still the right choice, but using HBM isn’t the slam dunk many people think it is."
He said the Samsung fab in Texas will make AI6 not AI5, HBM is not yet decided, and as the ratio of total to active parameters rises, conventional RAM may win on cost and compute per watt.
"The Samsung fab in Texas will make AI6, not AI5, and the decision regarding using HBM has not been made yet. As the ratio of total parameters to active (frequently used) parameters increases, the cost per unit of useful compute and compute per Watt may favor relatively conventional RAM."
He said AI5 design review went well, AI6 and AI7 will follow in fast succession, and AI8 will be out of this world.
"Just finished a long AI5 design review with the Tesla California and Texas chip engineers. It’s going to be great. And AI6 and AI7 will follow in fast succession. AI8 will be out of this world."
He said Tesla AI5 chips will be made at TSMC and Samsung, with samples or a small number of units in 2026 and high volume only in 2027; AI6 aims for roughly 2X and volume mid-2028; AI7 will need different fabs.
"Slightly different versions of the Tesla AI5 chip will be made at TSMC and Samsung simply because they translate designs to physical form differently, but the goal is that our AI software works identically. We will have samples and maybe a small number of units in 2026, but high volume production is only possible in 2027. AI6 will use the same fabs, but achieve roughly 2X performance. Aiming for a fast follow to AI5, so hopefully mid 2028 for volume production of AI6. AI7 will need different fabs, as it is more adventurous."
He said there is a clear path to doubling AI6 performance on all metrics within 10 to 12 months of AI5 shipping.
"And there is a clear path to a doubling of performance on all metrics for AI6 within 10 to 12 months of AI5 shipping"
He said even the suppliers' best-case chip output would not be enough, so Tesla may have to build a giant 'Terafab'.
"even when we extrapolate the best case scenario for chip production from our suppliers, uh it's still not enough"
He said AI5 will not be available in sufficient volume to switch Tesla production lines until mid-2027, because they need several hundred thousand completed AI5 boards line-side, and they have started work on AI6.
"Yup. Just wrapped up the AI5 Saturday chip design review a few hours ago. We’re starting to do some work on AI6 too. @MichaelDell and I happened to be meeting just beforehand, so I invited him to join. Hopefully, he found it interesting. Btw, AI5 will not be available in sufficient volume to switch over Tesla production lines until mid 2027, as we need several hundred thousand completed AI5 boards line side."
He said he is deeply involved in Tesla chip design, meeting the team every Tuesday and Saturday, and Saturday meetings will not be needed in a few months when AI5 is taped out.
"FWIW, I’m deeply involved in the chip design and meet with the engineering team every Tuesday and Saturday. The Saturday meetings are short-term and will no longer be needed in a few months when AI5 is taped out."
He said Tesla has designed and deployed several million AI chips, is close to taping out AI5 and starting AI6, aims for a new chip to volume every 12 months, and expects to build more AI chips than everyone else combined, saving lives via FSD and medical care via Optimus.
"Most people don’t know that Tesla has had an advanced AI chip and board engineering team for many years. That team has already designed and deployed several million AI chips in our cars and data centers. These chips are what enable Tesla to be the leader in real-world AI. The current version in cars is AI4, we are close to taping out AI5 and are starting work on AI6. Our goal is to bring a new AI chip design to volume production every 12 months. We expect to build chips at higher volumes ultimately than all other AI chips combined. Read that sentence again, as I’m not kidding. These chips will profoundly change the world in positive ways, saving millions of lives due to safer driving and providing advanced medical care to all people via Optimus. Send an email with three bullet points describing evidence of your exceptional ability to [email protected]. We are particularly interested in applying cutting edge AI to chip design. Thanks, Elon"
He said chip costs per trillion operations will drop several orders of magnitude.
"Chip costs per trillion operations will drop several orders of magnitude"
He said AI5/AI6 engineering is his biggest time allocation at Tesla, and that AI5 will be good and AI6 will be great.
"AI5/AI6 engineering is my biggest time allocation at Tesla. AI5 will be good, AI6 will be great."
2026 · 14 answers
Asked whether Tesla should acquire Intel fab assets, he said Terafab is needed.
"Terafab is needed"
He said solving AI5 was existential to Tesla; it is roughly Hopper-class as a single SoC and Blackwell as dual, cheaper and lower power; with AI5 looking good they have headroom for Dojo3.
"Solving AI5 was existential to Tesla, which is why I had to focus both the teams on that chip and I’ve personally spent every Saturday for several months working on it. This will be a very capable chip. Roughly Hopper class as single SoC and Blackwell as dual, but it costs peanuts and uses much less power. Now that AI5 is looking good, we have some headroom to work on Dojo3 too."
He recruited Koreans to join Tesla on chip design, fabrication or AI software.
"If you’re in Korea and want to work on chip design, fabrication or AI software, join Tesla!"
He said the goal requires SpaceX, xAI and Tesla working together on the Terafab project.
"really requires a combination of efforts of SpaceX, XAI and Tesla working together"
He said there is no other way to reach extreme scale: either they build Terafab or stay stuck at the industry's ~20% annual chip/memory output growth.
"In any event, there is no other way to reach extreme scale, so either we make Terafab or we will be stuck at the ~20% chip/memory output growth per year of the current industry."
He said the Terafab is not a negotiating tactic. Tesla simply sees no other path to enough AI chips at high production volumes.
"It's just literally we don't see a path to having enough, sufficient quantity of AI chips down the road as we scale production to high levels."
Getting to a terawatt needs a gigantic chip factory. The terafab is expected around 100 million square feet, ten times Tesla's Texas gigafactory. He says it is fundamentally about scale: even without new breakthroughs you could, with difficulty, scale existing chipmaking to a terawatt of chip output per year.
"the terafab is going to be around a hundred million square feet"
He said Terafab is necessary or Tesla will not have enough AI chips to scale Optimus, and they have ordered equipment for an Austin development fab with lithography, mask, logic, memory, packaging and test under one roof.
"we will be constrained in our ability to scale Optimus production because we simply won't have enough AI chips"
He guessed Terafab AI compute output would be about 25% for Tesla Optimus and 75% for AI spacecraft.
"My very approximate guess is that Terafab AI compute output would be ~25% for Tesla Optimus and ~75% for AI spacecraft"
He worries Taiwan chips may stop arriving for geopolitical reasons, and existing fabs are already at max capacity as AI, cars, and humanoid robots scale. Tesla and SpaceX are building an R&D fab at Giga Texas and hope to make something useful, at least the 'crawl' step, by the end of 2027.
"it's either build terrafab or or fail to scale"
He does not yet know how to make chips but expects to figure it out, starting with standard tools then customizing them; the goal is a fast on-site loop of about 10 chip designs a year instead of two years per design.
"I have no idea how to make chips, but we'll figure it out"
Correcting a report that TSMC might own and run Terafab, he said Tesla/SpaceX will build and run the fab; TSMC might sublease part of it and nothing more.
"No, we will build and run the fab. Let there be ZERO doubt about that. Maybe TSMC subleases part of the Terafab if they want, but nothing more than that."
He recruited for Terafab and said the value of its output will rank in the top 20 of country GDPs.
"Join the Terafab project. The value of the output of this massive structure will rank in the top 20 of country GDPs!"
He said the Terafab costs tens of billions and real production would come in late 2028, which is very fast for a fab.
"real production would be late 28, which is lightning fast by fab standards"