Elon Musk Is Building AI’s Biggest Weapon

Added: 2026-05-19

[00:00:00]I want to share with you today a deep dive on Elon Musk's real strategy, which is to build more compute than anyone else on Earth. Then I think the plan is he'll either rent it out or he'll use it for himself, Tesla, or SpaceX. So, you might have already heard last week Anthropic just agreed to rent every GPU at Colossus One. That's 300 MW, 220,000 chips. Google, apparently, is now reportedly in discussions to reserve Starship launches for their own space data centers. So, now we have the two biggest AI labs not named to open AI.

[00:00:28]They're lining up to pay Elon for compute. And here's what I want you to sit with. I don't think it's hyperbole anymore to say this is the most important race on the planet right now.

[00:00:38]I think it's more important than robo-taxi, more important than Optimus, maybe even more important than landing humans on Mars.

[00:00:44]Because I think whoever wins compute wins everything else by default. So, today I'm going to give you the most complete status update you'll find anywhere on Tesla and SpaceX compute.

[00:00:54]We're going to talk about the current builds, the future plans, and exactly how they stack up against OpenAI, Meta, Google, and Anthropic. Look, I get the skepticism here. Every 6 months, somebody says Elon just won something.

[00:01:05]He won robo-taxi, Optimus, Starship, and the goalposts keep moving. But I think this one's different, and I'll want to explain to you why. Because compute, raw brute-force compute, is the one thing in AI that you cannot fake. You can't spin a narrative on it. You can't type a model. You can't fake a building full of GPUs that's actually running. Here's what I'm going to walk you through.

[00:01:24]We're going to talk about every supercomputer in the Musk universe right now, Colossus One, Colossus Two, Macro Hard, Macro Harder, Cortex One, and Cortex Two. What they are, how big they are, how fast they got built. Then, we're going to share with you every supercomputer the rest of the Mag 7 is building. Stargate, Prometheus, Hyperion, Google's TPU stack. What's actually operational versus what's still a press release. Then, I'm going to explain something called the bitter lesson. Because once you understand this bitter lesson, you understand why this is the only race that matters. and why the company with the most compute eventually wins everything. Let me start with the one that started it all.

[00:02:02]Remember this? Colossus 1 sits in an old Electrolux factory in Memphis, Tennessee. XAI bought the shell because building from scratch was too slow.

[00:02:10]Here's the timeline that nobody else in the industry has matched, not even now.

[00:02:14]19 days from concept to construction start. 122 days to bring 100,000 H100 GPUs online.

[00:02:22]Then another 92 days to double it to 200,000. For context, a data center this size at Google, at Meta, at Microsoft usually takes two to three years. XAI did it in four months. And the way they did it tells you everything about how they operate. They couldn't get enough power from the grid. So, what did they do? They rolled in mobile gas turbines.

[00:02:40]They couldn't get enough water, so they tapped a wastewater treatment plant next door.

[00:02:45]Tesla Megapacks for battery backup, SpaceX engineering for the supply chain, the whole thing duct-taped together by people who actually ship. So, today Colossus 1 has roughly 220,000 GPUs.

[00:02:56]It's a mix of H100s, H200s, and the newer GB200 Blackwels. This is the one that Anthropic just rented in full. And here's the part most people miss when they ask how Elon got power so fast.

[00:03:07]There's a company called Solaris Energy Infrastructure. They own a fleet of about 600 MW of gas turbines. Roughly 400 of those are already serving XAI.

[00:03:17]And the remaining capacity sits inside a joint venture. So, Solaris owns just over half, XAI owns just under half.

[00:03:24]By the second quarter of next year, Solaris is targeting over 1.1 GW of operating turbines just for XAI. That's how you bypass a slow utility. You don't wait for the grid, you bring your own grid. Tesla Megapacks for battery backup, Solaris turbines for primary power.

[00:03:41]This is the vertical integration nobody else has. Now, here's where it gets serious. Colossus 2 is a different kind of beast. That's 2 GW of power capacity.

[00:03:50]555,000 Nvidia GPUs purchased. $18 billion worth of in hardware, and the target is 1 million GPUs at the single site. But the number that actually matters is one I want you to remember.

[00:04:02]It's coherence. Colossus 2 is the world's first 1 gigawatt coherent AI training cluster. Okay, so what does coherent mean? Here's the simple version. So when you train a frontier model, you want every single GPU talking to every other GPU constantly with almost zero latency because the model is one giant brain split across all those chips.

[00:04:22]If one corner of the cluster is slow, the whole training run slows down. So coherent means they're all stitched together into one machine.

[00:04:30]Non-coherent means you've got smaller clusters spread across multiple buildings or cities and you train in chunks and stitch the results together.

[00:04:37]Non-coherent is way less efficient. You lose a huge amount of training quality.

[00:04:42]Most of the industry, including some of the big open AI sites, is non-coherent at scale.

[00:04:48]xAI partnered with Cisco and Nvidia Spectrum X to solve the networking problem.

[00:04:53]And they're the only ones who've actually done it at a gigawatt. This is a thing competitors are still drafting plans for. xAI just turned it on.

[00:05:01]Now this is where it gets fun. Last fall, Elon announced something called macro hard.

[00:05:06]He literally tweeted it. It's a tongue-in-cheek name, but the project is very real.

[00:05:11]So the idea is this. Microsoft is a software company. They don't make hardware. So in principle, you should be able to simulate Microsoft, a fully AI run software company. Macro hard pairs Grok as the high-level brain, the navigator, then a Tesla-built agent watches the screen, moves the mouse, types on the keyboard just like a human worker would. It runs on Tesla's own AI 4 chip plus xAI's Nvidia hardware. The project's had some bumps. Engineers have rotated and left. A data annotation effort got paused because researchers found flaws in the model. That's normal.

[00:05:45]That's what shipping at the edge looks like. But, the strategic point is huge.

[00:05:49]If MacroHard works, you don't just have an AI company, you have an AI company that replaces every other software company. That's a multi-trillion-dollar wedge. Okay, so here's where it gets confusing, and I want to clear it up.

[00:06:01]MacroHard is the project, the AI software company. MacroHarder, with two Rs, is a building.

[00:06:09]Specifically, it's XAI's third data center in the Memphis region. And it's not actually in Memphis, it's in Southaven, Mississippi, just across the state line.

[00:06:17]The price tag is wild, over $20 billion, the largest private sector investment in Mississippi's history by a mile. This is the third building in Colossus 2 complex. And it's what takes XAI from one supercomputer site to a 2-gigawatt regional cluster.

[00:06:32]Operations targeted to begin around February.

[00:06:36]So, when you put the picture together, Colossus 1 in Memphis, about 300 megawatts. Colossus 2, main building, ramping toward a gigawatt.

[00:06:44]MacroHard the building, around 350 megawatts of cooling. And now, MacroHarder in Southaven, that pushing the total toward 2 gigawatts.

[00:06:53]Three buildings, two states, but one coherent training cluster. That's enough power to light up about 1 and 1/2 million homes.

[00:07:01]Now, the locals are absolutely pushing back, by the way. The NAACP and environmental groups have raised concerns about air pollution and water use.

[00:07:10]That's a real risk for the timeline.

[00:07:13]Permitting can slow things down. But, the capital's committed, the buildings are being built. This is absolutely happening.

[00:07:19]Now, let's flip over to Tesla.

[00:07:22]Most people forget that Tesla has its own supercomputers sitting right inside Giga Texas.

[00:07:28]Cortex 1 came online in Q4 of 2024. It started with 50,000 H100 GPUs. Now, it's scaling toward 100,000 H100s and H200s, about 130 megawatts of power and cooling, growing toward 500 megawatts. Cortex 1 is what trains Tesla's FSD models. Every single mile video from every Tesla on the road feeds it in here.

[00:07:51]This is where version 14 came from. This is where the unsupervised version is being trained right now. And then there's Cortex 2 being built on the north side of Giga Texas right now. 250 megawatts coming online next month. 500 megawatts by mid-year.

[00:08:05]And another roughly 100,000 GPUs. Here's the thing that connects all of this. So, Cortex 1 trains FSD.

[00:08:11]Cortex 2 trains FSD and will be the brains for Optimus. And Optimus is the bigger workload.

[00:08:18]A human or robot has way more degrees of freedom than a car. Way more video to learn from. Way more edge cases. So, when you add it up, you got Colossus 1 plus Colossus 2 plus you got Cortex 1 plus Cortex 2.

[00:08:31]You're looking at well over a million GPUs across the Musk Empire by the end of next year.

[00:08:36]Concentrated in two locations. All stitched together with coherent networking.

[00:08:41]Nobody else has done anything even close to this.

[00:08:44]Okay, so let's do the comparison because the headline numbers are one thing.

[00:08:48]Coherent, working, training.

[00:08:51]Today, those numbers are another.

[00:08:53]Let me start with OpenAI's target, the big one. The flagship site is Abilene, Texas. The plan is 1.2 gigawatts and 450,000 Blackwell GPUs built with Oracle, built by Crusoe.

[00:09:05]But, here's where it gets interesting.

[00:09:07]As of right now only about half of the eight buildings are operational.

[00:09:11]And the planned 600 megawatt expansion just got scrapped in March.

[00:09:15]Financing dispute plus winter weather messing with the liquid cooling build out or something like that. So, OpenAI is going big on paper, 10 gigawatts by 2029, but on the ground reality is they're still ramping the first gigawatt and pieces are slipping. Meanwhile, xAI already has a coherent cluster running and is renting it to Anthropic. Now, let's talk about Meta. This one's actually serious. Meta's building two superclusters with names. Prometheus is 1 GW in New Albany, Ohio coming online this year. Hyperion is the monster up to 5 GW in Louisiana. Meta literally ordered 10 gas-fired power plants for Hyperion.

[00:09:55]But Hyperion won't be fully built out until 2030.

[00:09:58]So, Meta has the biggest blueprint of anyone, but blueprints don't train models, operational GW do. And in 2026, xAI has more operational coherent compute than Meta does. Now, Google, they have TPUs, their own custom silicon, TPUv7.

[00:10:15]Very efficient, very powerful. Google's also spending 175 to 185 billion dollars in CapEx this year. But the irony is Google's reportedly in discussions to have SpaceX help them launch to space.

[00:10:29]Or maybe it'll rent compute from xAI too. Let that sink in for a second.

[00:10:33]Google with own TPUs renting launch and GPUs from Elon. Amazon. Amazon has Trainium chips. Anthropic has been using them. And as of last week, Anthropic is also using Colossus 1 because Anthropic ran out. So, when Elon was asked why he agreed to sell compute to Anthropic, he said, "No one set off my evil detector."

[00:10:52]That's it. That's the standard now. The same guy who called them anti-Western 3 months ago, compute makes strange bedfellows. And Anthropic didn't just want GPUs, they asked Elon about renting orbital compute, data centers in space, which we'll come back to. Okay, now we get to the part that ties this all together. This is the most important idea in AI, and almost nobody in the mainstream press talks about it. So, in 2019, there's a researcher named Richard Sutton, and he wrote a short essay called The Bitter Lesson. His argument in one sentence: For 70 years, every time someone tried to be clever in AI by hand-coding rules, by injecting human knowledge, by being elegant, the simple brute-force approach that just used more compute eventually beat them every time. Chess, Go, speech recognition, image recognition, language models, every single time. The bitter part is this.

[00:11:44]Researchers spend years building elegant systems, and then somebody comes along with 10 times more compute and a dumb algorithm and wins.

[00:11:53]It's bitter because the lesson is anti-human in a way. Your cleverness doesn't matter. The compute does. And here's the implication for Tesla, for SpaceX, for everything Elon's doing.

[00:12:04]If the bitter lesson is right, and 70 years of evidence says it is, then the company with the most compute wins. Not the company with the smartest researchers. Not the company with the best data set.

[00:12:14]The company with the most GPUs running coherent training. That's why every move Elon's made for the last 3 years makes sense. The factory in Memphis, Tesla Megapacks for power, SpaceX engineering for speed, the merger of xAI into SpaceX, the Tesla $2 billion investment in xAI. None of it is being clever. It's about scale. Now, here's where most people stop thinking, and I want you to keep going. Because if you think this is just about who has a better chatbot, you're missing it. The same compute that trains Grok trains FSD. The same compute that trains FSD trains Optimus. The same compute that trains Optimus can run a fully autonomous AI software company called MacroHard. And the same compute that does MacroHard, they can run AI scientists, AI doctors, AI engineers, AI lawyers, maybe even make new physics, new metals.

[00:13:02]This is the part Elon keeps trying to tell people, and it doesn't land because it sounds science fictional.

[00:13:08]But think about it like this. What is a business at its core?

[00:13:12]A business is a bunch of humans making decisions and producing outputs.

[00:13:16]If you can run human-level intelligence on a chip at scale, you can run businesses on chips.

[00:13:22]You can develop new physics on chips.

[00:13:24]You can run drug discovery on chips and cure diseases that have killed people for 10,000 years.

[00:13:30]The demand for intelligence is effectively infinite. There's no point at which a civilization says, "We have enough intelligence, thanks." You always want more.

[00:13:39]And whoever can sell intelligence by the hour, at scale, at low cost, is selling the most valuable commodity that has ever existed.

[00:13:48]This is what Anthropic figured out. This is why Google's in talks.

[00:13:52]Compute is oil. Compute is the new electricity.

[00:13:55]And Elon just built the biggest refinery. Last piece, and this one's wild, and it's not hypothetical anymore.

[00:14:02]In January, SpaceX filed with the FCC.

[00:14:04]The filing was for a million satellite orbital AI data center constellation. A million satellites in orbit as one supercomputer.

[00:14:12]That's not a tweet. That's a slide. That is a regulatory filing with the United States government.

[00:14:18]Then in May, when Anthropic signed the Colossus 1 deal, they didn't just rent GPUs on the ground. They expressed interest in working with SpaceX on multiple gigawatts of compute in space.

[00:14:28]Why orbital data centers? The power is effectively unlimited because solar runs 24/7 up there.

[00:14:35]The cooling is basically free because space is 270° below zero.

[00:14:39]There's no zoning board. There's no NAACP lawsuit. No water rights fights.

[00:14:46]SpaceX can launch what they ever they want. 200 Starships a year by 2028 on current targets. Each one carrying hundreds of tons of to orbit.

[00:14:55]You start stacking GPU clusters in space and a scaling curve looks nothing like what anyone else can do.

[00:15:01]Microsoft can't do this. They don't have a rocket. Google can't do this. They don't have a rocket. Meta, they don't have a rocket. Only one company on Earth has the launch capacity to put compute in orbit at scale.

[00:15:13]That company is now the parent company of xAI. This is what people don't price in. Earth compute has limits. Power grids, water, land, permits. Space compute doesn't. The first company to crack orbital data centers gets an effectively unlimited scaling curve.

[00:15:29]SpaceX has already filed for that spectrum.

[00:15:32]Okay, so let me give the bear case a fair shot because there are real risks here. The first one is power. Colossus 2 needs gigawatts. That's small city scale. Satellite imagery suggests the cooling capacity is currently behind schedule. Some analysts think Colossus 2 won't hit a full gigawatt until later this year. That's real.

[00:15:50]Power and cooling are the bottleneck, not chips. The second one is macro heart. Projects had leadership turnover.

[00:15:56]Engineers have moved teams. A big annotation effort got paused. The thesis is sound, but execution is hard. The third one is competitive response.

[00:16:05]Google CPOs are very good, really good.

[00:16:08]OpenAI plus Microsoft have more total CapEx committed than xAI does. If Starlink hits its build targets, they could match or exceed Colossus on raw GPU count. The fourth one, and this is the honest one, Elon's timeline slip, always, on everything.

[00:16:22]So, when he says 1 million GPUs by the end of this year, I'd add at least 9 months at the minimum. But here's the reframe. Even with slippage, even with cooling delays, xAI has the only operational gigawatt scale coherent cluster on Earth right now. Anthropic is paying for it. Google's in talks for it.

[00:16:38]That is not a hypothetical. That is happening this month.

[00:16:41]The lead doesn't have to be permanent for the thesis to work. It just has to last long enough to train two or three generations of frontier models. And then the network effects and data flywheels take over. So, if you own Tesla stock, here's how I think about this. The market still prices Tesla as a car company that maybe building a robot.

[00:16:58]Some analysts are starting to add a robotaxi premium, fine. But almost nobody is pricing the compute story.

[00:17:04]Tesla's Cortex clusters are training FSD, that's understood. What's not understood is that Tesla owns roughly 9% of xAI now, and xAI just merged into SpaceX, which Tesla is also a major customer of. So, you've got two companies, Tesla, SpaceX, and xAI sharing engineers, sharing compute, sharing supply chains with one guy at the top who points them all in the same direction. If you think AGI is real, and the bitter lesson says it is, then the compute leader captures most of the value.

[00:17:31]And Tesla shareholders own a piece of that compute leader, either directly through xAI ownership or indirectly through Cortex training the FSD and Optimus models to monetize the compute.

[00:17:40]Three scenarios I think through.

[00:17:42]One, let's say that AGI doesn't happen for 10 plus years. Tesla trades on cars and FSD, stock goes up but slowly. Two, AGI happens by 2030. xAI is a top three lab, Tesla's ownership stake plus Optimus deployment make Tesla the most valuable company on Earth.

[00:17:58]Three, the compute lead compounds. Space data centers come online. xAI becomes the default intelligence provider for the planet. Tesla rides that wave plus owns the physical AI layer with Optimus.

[00:18:09]Market is pricing scenario one right now.

[00:18:12]Here's what I'd actually do with this information.

[00:18:14]The first thing is track the GPU count.

[00:18:16]Colossus 2 is supposed to hit 1 GW by mid-year. Satellite imagery and Epoch AI publish updates. Watch that number.

[00:18:24]The second one is watch the customer list. Anthropic was first, Google's reportedly in talks right now. If a third major AI lab signs up, that's a signal that xAI compute is becoming the default rental market.

[00:18:36]Then there's Cortex 2. 250 MW in April, 500 by mid-year.

[00:18:41]Tesla doesn't usually announce these milestones cleanly.

[00:18:44]Watch for satellite imagery, watch for utility filings in Travis County, Texas.

[00:18:49]Fourth, the macro hard product. If a public beta drops this year, pay attention. Even a small launch matters because it proves the concept. And the last one is the space angle. If SpaceX announces an orbital data center test mission, this changes everything. That's the black swan that no other company can match. So, I've been covering Tesla since 2012. I watch a lot of cycles, a A of milestones, a lot of hype.

[00:19:12]What's happening right now in compute is different. This is not a product launch, not a press conference, it's not infrastructure.

[00:19:19]It's built running already generating revenue.

[00:19:23]For most of AI history, scale has been the answer and whoever has the most scale wins.

[00:19:29]Elon has the most scale. He's pulling away.

[00:19:31]Anthropic is rented from him, Google's in talks and the next data centers are going to orbit. If you understand the bitter lesson and you understand convergence, you understand why Tesla, SpaceX and XAI being one effective unit is the most important corporate structure on Earth right now.

[00:19:47]I'll keep tracking the milestones. You know the drill. If you got something out of this, please hit subscribe. I'll see you in the next one.