Tesla Just Unleashed $10B AI Monster: 100,000 GPUs Storm Giga Texas!

Added: 2026-05-17

[00:00:00]This past May 13th, Musk declared Tesla's field AI untouchable. The same day, a drone flew over Giga Texas. Four massive construction sites, $10 billion running in parallel. At the center is a supercluster aiming toward 100,000 Nvidia GPUs, devouring 500 megawatts of electricity, enough to run an entire small city. the whole of it only to feed one single brain. That brain is teaching Tesla cars, the Optimus robot, and even the cyber cab at the same time. So, what is it? Let's dive right in.

[00:00:49]Cortex 2 is Tesla's second AI training supercluster cited right on the Giga Texas campus next to its elder sibling Cortex 1 which has been running from before. When the two clusters combined reach full capacity, the total number of Nvidia H100 and H200 chips inside will reach 100,000.

[00:01:15]This is one of the largest AI infrastructures on the planet at this point in time. Total investment for this entire ecosystem reaches $10 billion.

[00:01:27]The initial 250 megawatt phase has been running since April. The full 500 megawatt is projected to be reached by the middle of this year. On May 13th, Joe Tegmire's drone flew over the campus and recorded four construction sites operating in parallel to serve this single brain alone.

[00:01:50]But at this point, there is one thing that makes me sit back and think. Cortex 1 already exists, already runs stably.

[00:01:59]So why must Tesla rush to build a further Cortex 2 right now? Cortex 1 when it first launched brought training power 10 times what Tesla had before with roughly 50,000 GPUs.

[00:02:15]Sounds large, but Tesla is not running one AI project. They are running four projects in parallel at the same time.

[00:02:25]One is FSD version 15 with an AI model 10 times larger than the current architecture. Two is the neural network for the Optimus robot preparing to enter mass production.

[00:02:39]Three is the vision model for Cyber Cab, the robo taxi that just left the factory in April. Four is simulation to shorten the field test cycle. The problem lies here. Each of those projects on its own alone already demands a quantity of compute greater than even Cortex 1 can supply. Four projects in parallel means Tesla must double its infrastructure right now, not to upgrade, but only to be allowed to keep playing. Cortex 1 taught Tesla how to build an AI supercluster. And Cortex 2 is the first time Tesla is permitted not to have to choose which project takes priority over which. Everything seems clear at this point. A second cluster is added, more GPUs added, the problem solved. But in reality, this is only the beginning.

[00:03:36]because 100,000 GPUs running at the same time generate a colossal volume of heat to the extent that conventional data centers cannot handle it. So, how does Tesla solve that problem? A single H100 chip when running at full load emits roughly 700 watts of heat multiplied by 100,000 chips. The result is 70 megawatt of heat. Enough to heat up an entire small neighborhood. If using traditional air cooling, the electricity bill for cooling alone would grow so large that no company in the world could sustain it long term. That is the reason most large data centers today must be cited in regions with naturally cool climates like Iceland or the north of the Nordic countries. Tesla chose a different path.

[00:04:31]Cortex 2 uses Super Microsant running directly to each GPU, not through air. The effect delivered an 89% reduction in cooling electricity costs compared to traditional air cooled systems. 89% not 8% not 20%.

[00:04:55]Converted into dollars, this saving is equivalent to tens of millions of dollars per year, enough to fund the entire R&D of a midsized AI startup.

[00:05:06]Tesla is not simply saving on electricity. They turn it into a strategic advantage that the majority of observers do not recognize. There is one more detail from the drone footage of May 13th that I want to mention. On top of the part two cooling tower of Cortex 2, maintenance vehicles were already parked there. Sounds trivial, but it says something important. This site is no longer in the construction phase. It has stepped into the trial operation phase. Cortex 2 is not a plan on paper.

[00:05:42]It is alive. There is one final point about this infrastructure that almost no one has mentioned. The water source used for the cooling system comes from recaptured rainwater stored in underground tanks beneath the Giga Texas campus. Tesla does not take water from the city water supply grid. They create their own separate circulation loop self- sustaining their own system. Tesla is not building a data center. They are building an inverse thermal energy plant. Input is electricity and rainwater. Output is intelligence. There is the cooling system. There are the 100,000 GPUs.

[00:06:27]There are the 500 megawatt of electricity. But all of those things in the end are only hardware. The more important question is what exactly is this massive apparatus training? Cortex 2 does not receive data from one source.

[00:06:43]It receives four entirely different data streams simultaneously.

[00:06:48]And most importantly of all, all four of those streams pour into the same single foundational neural network. The first stream is visual data from more than 1.3 million Tesla cars running on the road.

[00:07:04]The FSD subscriber number as of the first quarter of this year. Each of those cars is a video source running nearly continuously recording roads, pedestrians, signs, those special situations that no AI laboratory can produce. Tesla does not need to buy data. data arrives on its own every hour. No other AI company on the planet is holding this advantage. The second stream comes from the Optimus robots running internal trial runs inside Tesla factories. A few dozen robots sounds like little, but the type of data they generate is very special. Motion data, hand force data, balance data, data on how robots handle obstacles in real environments. This is the type of data that no one else on the planet is in possession of at equivalent scale.

[00:08:01]Boston Dynamics has robots but not linked to a broader AI network. Figure has robot data but at a far smaller scale. The third stream is simulation data. The stream whose real value few people realize. Cortex 2 itself generates driving situations and robot manipulations in a virtual world, then uses the results to pre-train the neural net. One hour of running simulation on Cortex 2 is equivalent to thousands of hours of field testing in real life. A Tesla car may never encounter a specific situation in its entire life cycle, but Cortex 2 can create that situation millions of times in a single night. And the final stream comes from the cyber cab fleet that just left the factory.

[00:08:54]The production spec cyber cabs have been tested running autonomously on the Giga Texas campus. The feedback loop from Cyber Cab back to Cortex 2 is the shortest feedback loop Tesla has ever had in the entire history of the company. A few hours, not a few weeks.

[00:09:14]At this point, the most shocking point emerges. These four data streams do not stand individually. They are poured into the same single foundational neural network. When FSD learns the way to handle pedestrians, Optimus also learns.

[00:09:32]When Optimus learns to keep its balance on slippery surfaces, the simulation system is also improved. When Cyber Cab encounters an unfamiliar situation on the factory campus, Tesla cars out on the road also benefit. This is not four separate AI projects of Tesla. This is one single brain with four senses.

[00:09:56]That brain is fed every second, every hour. So what will its output change on Tesla products in the next 6 months? The first direct product Tesla users will sense from Cortex 2 is precisely FSD version 15. Musk announced a number that made the entire automotive AI industry have to stop and look carefully. FSDV15 will run on a model with 10 times the parameters compared to the current architecture. 10 times, not double, not triple. In the AI world, a 10 times parameter jump is not called an upgrade.

[00:10:39]That is a completely new generation. To help you visualize, GPT3 to GPT4 was about 5 to 10 times the parameters.

[00:10:50]And that jump alone was enough to change the way the entire world looks at AI.

[00:10:55]Tesla is making a similar jump, but for driving. The problem lies here. A model 10 times larger requires a training infrastructure 10 times larger. Cortex 1 is not enough. This is precisely the entire reason Cortex 2 must exist. Not to show off power, but because FSDV15 cannot be born if lacking it. Musk declared the goal for V15 to operate at a safety level far surpassing humans in unsupervised self-driving situations.

[00:11:32]This is the most ambitious statement Tesla has ever made about FSD throughout the 10 years of developing this technology. And here is the detail that links the entire story back together. On the recent Q1 earnings call, Musk admitted a truth hard to say. Hardware 3 on older Tesla cars does not have enough memory bandwidth to run unsupervised FSD.

[00:12:00]So where is the solution? At the new AI5 chip, precisely the chip Tesla is preparing to produce within the Giga Texas campus will be where V15 actually runs. Do you see the connection yet?

[00:12:16]Cortex 2 trains V15.

[00:12:19]AI5 runs V15 on the car. Field data from the car returns to Cortex 2 to continue training. A self-improving loop with no end point. For the first time in Tesla's history, three puzzle pieces, the training brain, the processing chip, and the AI model are upgraded simultaneously, not out of phase. But all compute model chip in the end must serve one thing. A physical product that humans can touch, work with, live with. What is it? Open AAI has chat GPT. Anthropic has Claude.

[00:13:03]Google has Gemini. All are AI that live on servers, process text and images, answer questions through screens. Cortex 2 was not built to compete with those systems. Tesla does not need one more chatbot. Cortex 2 was built to sustain something very different. A truly physical product capable of walking among humans. That product is Optimus, a humanoid robot running a shared neural network with Tesla cars. And this only now is the pivotal point of difference between Tesla and every other humanoid competitor on the market today. Figure AI develops robots and AI separately from cars because they have no cars.

[00:13:52]Boston Dynamics has only robots, no field data from millions of cars running on the road. Unitry produces lowcost robots at large scale, but has no central brain at Cortex scale to teach them. Tesla regards Tesla cars and Optimus not as two different products, but as two physical shells together, sharing one single brain. Every hour Cortex 2 runs training, Optimus becomes a little smarter. Not by means of adding new code, not by means of programming more features, but by means of adjusting weights in the same neural network that taught Tesla cars to break when they see pedestrians. When Cortex 2 reaches the full 500 megawws, that is the first time in the history of technology that a company places a physical intelligence brain at national scale into one single purpose, teaching machines to act in the real world. A bet this large always comes with barriers. What is Cortex 2 facing? The most obvious difficulty comes from precisely the GPU supply.

[00:15:06]itself. Nvidia is currently having to allocate H100 and H200 chips for very many large customers at the same time.

[00:15:16]Microsoft, Meta, Google, Amazon, Oracle.

[00:15:22]Tesla is not the only customer and has no special priority, right? The speed at which Cortex 2 reaches full capacity depends directly on the delivery schedule from Nvidia. If delayed by one quarter, the entire FSDV15 plan slows along with it. This is a risk Tesla cannot fully control. The next problem lies right at Texas. 500 megawatt for Cortex 2 plus the remaining part of Giga Texas is pushing the burden onto the local power grid.

[00:15:58]Texas once underwent widespread power outages during the winter of 2021, an event the entire local industry has not yet forgotten. Tesla is investing in its own backup infrastructure, megapac batteries, and internal solar energy grid. But these hidden costs have not been fully counted in the $10 billion investment figure announced. Harder than even supply. and the power grid is the question of model architecture.

[00:16:30]Large compute does not automatically turn into good AI. Tesla must choose the right neural network architecture to fully exploit 100,000 GPUs.

[00:16:45]A wrong decision about architecture on Cortex 2 could cost months of compute passing without creating value, equivalent to tens of millions of dollars burned into the air. This is a technical risk that only very few engineers in the world have enough expertise to assess correctly. And the final pressure comes from outside.

[00:17:08]The same day Musk declared untouchable, Figure AI introduced Helix 2 with humanoid robots running autonomous 8hour work shifts. That same week, Whimo expanded robo taxi coverage to 1,400 square miles across 11 American cities.

[00:17:29]No one among them stands still. All peer competitors are accelerating at the same time. Cortex 2 is not insurance for Tesla. It is the ticket for Tesla to be allowed to continue playing in a race where every competitor is accelerating.

[00:17:46]Now, back to Musk's original declaration.

[00:17:50]Untouchable cannot be touched. After going through everything just covered, I hope you realize one thing. Musk is not talking about a specific product. He is not talking about a single car, a single chip or an individual robot. He is talking about architecture. Cortex 2 is not a massive data center. It is the architecture that allows a company to train physical intelligence at a scale never before existing in the history of mankind.

[00:18:24]This is not a chip race, not a robot race, not a software race. This is a data architecture race. Whichever side has the most realworld data, the largest training infrastructure, and the shortest feedback loop will be the side that defines the physical AI of this entire decade. And Tesla, right at this moment, is in possession of all three of those advantages at the same time. There are three specific time milestones I hope you remember. By the middle of 2026, Cortex 2 must reach the full 500 megawatt. If achieved, this is the largest compute jump the automotive AI industry has ever had. By the end of 2026, the first training results from Cortex 2 must show clearly on FSDv15.

[00:19:22]This is the real test of the entire strategy. And by 2027, Cortex 2 must support Optimus V4 running commercially for the first time. In the next 12 months, if Cortex 2 truly reaches full capacity and all three milestones above are completed, will the world's definition of an AI company still be the same as it is today? Cortex 2 may be redefining physical AI or it may also be an overly large ambition.

[00:19:58]The next 12 months will give the answer.

[00:20:00]Tech Revolution was born to follow alongside you those technology stories like this one. Whether you are an AI engineer or have never heard the word GPU before, I believe everyone deserves to understand where the world is going. If anywhere I have not explained clearly enough, leave a comment so we can together continue exploring. Like if the video brings value, share with people who need to know and subscribe to walk with us. Thank you for staying with me until here.