By Marco Hugo Barsotti MHB 
This is a readable-outside-X repost from an original article by Farzad (@farzyness)
Anthropic announced that they’ve agreed to a partnership with SpaceX that will substantially increase their compute capacity, by providing 300 megawatts of compute power (a lot), which will allow the Claude models to operate with vastly higher rate limits. Basically, people will be able to use Claude without running into issues related to available compute.
Anthropic was running into massive issues due to the sudden surge of the popularity of their models, which was driving a lot of customers and businesses completely insane. This is because the company massively underestimated the amount of compute they would need to power their AI models. Lucky for them, SpaceX, which now is combined with xAI, had a massive surplus of compute capacity.
Elon posted that he had spent a lot of time last week with senior members of the Anthropic team to understand what they do to ensure Claude is good for humanity, and he was impressed.
Thus, the partnership was born.
What this means in practice is that Anthropic gets all of the compute capacity at Colossus 1, the data center in Memphis, Tennessee that was built to train Grok. That is over 300 megawatts of additional capacity and more than two hundred and twenty thousands NVIDIA GPUs, mostly H100, H200, and GB200 chips, and It should come online within the month. Claude Pro, Claude Max, Claude Code, Team, Enterprise, all of them, the rate limits effectively will double starting today. Peak hour throttling on Pro and Max accounts for Claude are removed, and API limits on Opus models are raised.
Now to understand why this deal is bigger than a simple rate limit increase, you need to understand what an AI Hyperscaler actually is.
An AI Hyperscaler is just the modern word for the small handful of companies that own enough compute, enough chips, power, and physical infrastructure to actually train and run frontier AI models at scale. Microsoft Azure plus OpenAI is one. Google Cloud with Gemini is another. Amazon Web Services with Anthropic as a tenant is another. CoreWeave, Oracle, Meta are all on there own internal stack.
So for most of the modern AI era, going back to GPT 3 in 2020, the assumption was that Big Tech, meaning Microsoft, Google, Amazon, and Meta, owns the power, the data centers, and the AI labs are tenants on top of the structure.
The model is the product. The compute is the landlord. That was the universe up until basically right now.
So to understand why this matters, we need to walk through 3 different puzzle pieces
Puzzle piece number one.
Two months ago, Elon Musk was calling Anthropic literally misanthropic and evil. He was mocking Dario Amodei’s writing – the CEO of Anthropic – about AI consciousness. There was no daylight whatsoever between Elon and Dario.
Now what’s interesting is that they both left OpenAI. Musk in 2018, Amodei in 2021, both citing concerns related to Sam Altman’s character. And both went and built competitors to Sam Altman’s company.
But Elon and Dario themselves, publicly, were enemies. But now, at the very least, they are partners. Perhaps a little bit of keep your enemies close kind of thing. We’ll never truly know, probably.
Puzzle piece number two.
On February 2nd, 2026, SpaceX absorbed xAI in an all stock transaction valuing the combined entity at roughly $1.25 trillion – which is now much closer to $2 trillion based on today’s estimated valuation. Reuters called it the largest corporate merger ever recorded.
xAI was Elon’s AI lab, the one that builds Grok. And Colossus, the giant Memphis data center complex with about 555,000 GPUs across two buildings, became part of SpaceX in that deal. SpaceX is no longer just a rocket company but also an AI lab. And it owns the compute that AI lab runs on.
Puzzle piece number three.
The Pentagon earlier this year reached AI use agreements with SpaceX, OpenAI, Google, Microsoft, NVIDIA, AWS, Oracle, and a smaller player called Reflection. Anthropic was explicitly cut out. Why? Because of that whole debacle between Anthropic and the Department of War that went down a month or so ago. Remember that disaster? This is important. I’ll explain why later.
Now, while all of that was happening in the background, Anthropic had a public capacity problem. Throughout the first half of 2026, Claude Code users, which is the developer product that competes with Cursor and Codex and the like – were hitting their rate limits in hours instead of days. Articles started appearing all making the same point. Anthropic is demand strangled. They have the best product on certain dimensions, but they cannot get enough megawatts plugged in fast enough, which means not enough compute.
Which means customers are pissed.
Anthropic, with the best model on Earth in many people’s view, is having to throttle their own customers.
Their valuation is somewhere around $1.2 trillion as of this recording – which is absolutely ridiculous. Their revenue is climbing violently. Roughly $9 billion in annual run rate at the end of 2025, then about $19 billion by March, and $30 billion by the first week of April. Honestly they’re probably closer to $40 billion by now.
Anthropic’s Annual Revenue Run rate ramp from $9 to $30 billion in roughly one hundred days, is the fastest revenue ramp of any company in history.
So that’s the puzzle.Anthropic, demand strangled, frozen out of the Pentagon, with the steepest revenue curve software has ever seen. SpaceX, fresh off absorbing xAI, sitting on a Memphis data center the size of a small city, with a model called Grok that is, frankly, not commercially competitive with the other guys. And in the middle, the personal Elon and Dario beef that looks like an episode out of a telenovela.
Now here are the real implications.
Layer one.
The bottleneck on Anthropic’s growth is, at least temporarily, broken. Anthropic gets more capacity. SpaceX gets a massive recurring revenue stream from a paying frontier model tenant. Two former enemies decide their shared problem, OpenAI, is bigger than their personal problem with each other. And Anthropic now has the runway to release Mythos, their latest model, which apparently is so good that it’ll destroy the world.
We all know that’s hyperbolic BS, but the benchmarks don’t lie. This model appears to be incredibly strong, and now Anthropic will be able to serve it to customers to further their lead.
Layer two.
The thing most people are not seeing yet is that SpaceX is no longer a rocket company in any meaningful sense of the word. SpaceX has very quietly assembled, layer by layer, the only fully vertically integrated AI infrastructure stack on planet Earth.
Starting with launches.
SpaceX owns Falcon 9, which is dropping launch costs by literal orders of magnitude. But more importantly, they own Starship, which when it actually starts flying at cadence is targeting between $10 and $100 per kilogram to orbit. From the current Falcon 9 prices that is somewhere between a thirty times and a three hundred times reduction. From where the industry was ten years ago it is over a thousand times reduction. In time, Starship will be sending data centers to space to power these AI systems to harness the infite, free energy from the sun to power super intelligent AI.
Then you’ve got Chips.
SpaceX along with Tesla is building Terafab, a 2 nanometer chip fab in Austin, Texas. Originally announced as a $25 billion dollar project, now on pace to be closer to $200 billion.
The Kobeissi letter, three hours before the Anthropic announcement today, reported that SpaceX is now actually proposing $55 billion of capital expenditure for phase one alone, which could rise to $119 billion across all phases, in addition to Intel joining the project in April as a manufacturing partner. The roadmap is 100,000 wafer starts per month ramping to 1,000,000 per month, which would be roughly 70% of TSMC’s entire global output.
Now look at Power & Compute.
They own Colossus 1 with 220,000 GPUs. They own Colossus 2, a gigawatt scale facility with another 335,000 plus GPUs. They have a third site coming online in Southaven, Mississippi, with a roadmap to a million GPUs by late 2026. About $18 billion invested in the Memphis complex alone.
Or what about the Network.
Starlink, with over 10 thousands satellites in orbit, the largest commercial satellite constellation in human history, and ramping to a million in the coming years.
Or what about the Models.
Not only does SpaceX own xAI and Grok. As of today, they are renting frontier model compute to Anthropic. So Anthropic and Grok are co-tenants in the same data center. And one floor up, on the developer surface, SpaceX has a $60 billion option to buy Cursor by the end of 2026, with a $10 billion breakup fee if they walk. Cursor is the leading integrated development environment built natively on Claude, by the way.
Stack those layers up: Launches. Chips. Power. Satellites. Models. A Developer surface.
The Musk industrial stack, with SpaceX at the center, has six. The rest of the field maxes out at four, if they are lucky.
Layer three.
The winners and losers from this deal. Start with Anthropic.
They were demand strangled. Now they are not, at least less so. Claude Code users get more capacity instantly. Their $30 billion ARR economics improve because compute cost per token drops as utilization rises across more capacity. Their next model, called Mythos, which they previewed in April to limited partners with a one million token context window and which is reportedly the most capable model anyone has ever benchmarked, becomes deployable to a much larger user base.
Anthropic wins, but they pay a real cost, likely billions of dollars per year to xAI to rent their compute. They are now structurally tied to a contractor that is helping the Pentagon build military AI, while they themselves are blacklisted from that work. Kind of ironic.
Now with SpaceX. They win in the most obvious way.
They just publicly proved their thesis as an AI infrastructure company. They have a paying frontier model tenant. That is the same business model Microsoft has with OpenAI, which is what justifies a $1 trillion plus valuation on Microsoft Azure alone.
You have to keep in mind SpaceX is targeting an IPO sometime mid 2026. Bank of America is projecting $1.5 to $2 trillion or more on the public listing. This deal makes that math much easier to defend. You’ve got recurring AI compute revenue, on top of launch revenue, on top of Starlink, on top of the Cursor option, on top of Terafab equity, on top of a potential merger with Tesla. Absolutely unbelievable position.
Now the loser, to some extent. xAI and Grok.
xAI built Colossus 1 to train Grok. SpaceX bought xAI. Today, SpaceX rented every megawatt of Colossus 1 to Anthropic. Grok’s training pipeline gets reshuffled to Colossus 2 and to the Mississippi site, but xAI is not crippled. They still have Colossus 2, they have Mississippi with a roadmap to one million GPUs, they still have the Cursor option.
So I am not going to say Grok is dead. I will say this through. Essentially every original xAI cofounder besides Elon has left in 2026, basically the entire founding team. The implicit message of today’s deal is loud and clear. SpaceX values Anthropic dollars more than Grok’s position. Grok’s internal status looks fragile. Whether they can climb back to frontier model status is a real open question, and no one should ever count Elon out of anything. But as of today, the rented compute is worth far more than the tokens Grok generates.
Layer four: The entire system.
Watch this.
AI needs energy. Energy needs gigawatts. Gigawatts need land, permits, and time, twenty four to thirty six months minimum to build a serious data center – unless you’re Elon which you can then throw one up in like 4 months.
AI also needs chips, and frontier chips need fabs that take three to four years to spin up, minimum, which is massively bottlenecked by ASML, who makes the machines to make the chips.
AI also needs networks to move the data. AI needs developer surfaces to actually deliver itself to humans, and developer surfaces need an installed base. Then you’ve got the Models, which everyone obsesses about because models are the visible part, but are now actually the easiest layer to commoditize. Especially with Open source releases from China, from Mistral, from Meta itself, which are compressing the gap between best in class and adequate down from years to months.
The moat used to be the model. Now it is everything below the model.
Whoever owns the launches, the chips, the fabs, the gigawatts, the satellites, the developer surface, and the model is positioned to capture margin from everybody upstream of them, including frontier labs who used to be peers.
The labs that do not own those layers, OpenAI for instance, become structurally dependent on whoever does. Microsoft owns OpenAI’s compute. Google partly owns Anthropic’s compute. Amazon partly owns Anthropic’s compute. And now, as of today, SpaceX owns part of Anthropic’s compute. SpaceX owns Grok’s compute. SpaceX has the option to own Cursor. SpaceX is building the chips along with Tesla and Terafab. SpaceX is launching the satellites that increasingly will host the data centers themselves.
Wild stuff.
Now for the last ten years, the assumption was that AI was won by the company with the smartest researchers and the best benchmarks.
That story is over.
Today, AI is won by the company that owns the most layers of the physical stack.
Compute is the new oil.
Megawatts are the new pipeline. Chips are the new refinery. Launches are the new shipping fleet. Developer surfaces are the new gas station. Models are the new gasoline, which is now commoditizing, so the gasoline is actually the cheap part. So that means whoever owns the rest of the chain, wins – and wins big.
At this point, you might be thinking the obvious read is, ah, Dario and Elon buried the hatchet, they are ganging up on Sam Altman, and OpenAI is in trouble, especially with the current lawsuit between Elon and Sam.
You’ve got two former OpenAI colleagues team up against the third. And there is some truth to it. OpenAI is in real trouble in some ways. Their CFO capped Microsoft’s revenue share recently. They pulled back from the Stargate Norway data center, and Microsoft picked it up directly via a company called Nscale, contracting thirty thousand Rubin GPUs themselves.
OpenAI has committed $250 billion to Microsoft Azure over six years, but Microsoft’s 2026 capex of $190 billion and a doubling of AI infrastructure in two years tells you Microsoft has the upper hand in that relationship.
So yes, OpenAI is squeezed.
But here’s the reality – two months ago, the entire reason SpaceX absorbed xAI for about $250 billion in stock was because they wanted to be in the AI race. They wanted Grok to compete head on with Claude and ChatGPT and Gemini.
As of today, that whole calculus has changed. Grok is now the relatively underutilized internal product. SpaceX’s actual go forward business model in AI is not winning the model race.
It is being the landlord for everyone else’s models.
They do not need to beat Claude. They just need Claude, and whoever is next, to live on their compute, on their chips, on their power, in their satellites, especially now that they are massively ramping up their Starship capacity and Terafab capacity. The amount of chips and rockets SpaceX is going to build is going to be next level absurd.
This is exactly what Tesla did to legacy automakers by the way, just applied to AI.
Think about it. Tesla vertically integrated the supply chain. They built their own battery cells, their own chip designs for FSD, their own software stack, their own factories. By the time GM and Ford and Volkswagen tried to catch up, the entire industrial logic underneath had moved.
Legacy auto could not win the EV race because they did not own the layers underneath the car in an autonomous EV world. SpaceX is now doing the exact same play in AI. They are trying to own everything below the model, so that whoever wins at the model layer pays them rent. Forever.
In the AI race, you have way more competent players, so the competition will go on for decades, but the reality is that SpaceX is now the only fully vertically integrated AI Hyperscaler on Earth, and that means SpaceX is best positioned to capture the largest single share of the AI economy through the rest of this decade.
And this is something that affects basically every person.
If you work in software, your tools are about to get a lot more powerful, and the labs that ship those tools are increasingly going to be tenants on top of SpaceX infrastructure.
If you work in any kind of knowledge work, the AI assistants that displace or augment your workflow are increasingly going to be running on SpaceX compute.
If you are an investor in any of the Mag 7 names, you need to update your model. Microsoft, Google, Amazon, NVIDIA, Meta, Apple, Tesla. Incredible names. But those names assumed a four player race for AI infrastructure. There are now five players. And the fifth, SpaceX, is the only one with launches and chip manufacturing under its own roof.
What does the world actually look like on a Tuesday morning in 2030 when one company owns the rockets, owns the chip fab, owns the megawatts, owns the satellites, owns the broadband, owns the developer surface for AI, and rents out the smartest minds on the planet to every other company on Earth?
Completely wild to really think about.