Anthropic is now running on SpaceX’s supercomputer. The company that lands rockets is now powering your AI agent.
This is the kind of announcement the AI industry produces in 2026 that would have read as fiction in 2024.
On May 6, Anthropic and SpaceX signed a partnership giving Anthropic full and exclusive access to Colossus 1, the data center in Memphis, Tennessee that xAI built in 2024 to train Grok. The deal gives Anthropic access to over 300 megawatts of compute capacity and more than 220,000 Nvidia GPUs, including H100, H200, and the newest GB200 accelerators. New capacity comes online within the month. The two companies also stated they have “expressed interest in partnering to develop multiple gigawatts of orbital AI compute capacity.”
The same Elon Musk who, in February of this year, wrote on X that Anthropic “hates Western civilization.” The same Musk who has called Anthropic “woke,” “misanthropic,” and “evil” on multiple occasions. The same Musk whose company xAI directly competes with Anthropic on every dimension that matters.
That Musk just leased his second-largest AI data center to the company funding the AI lab competing with his own.
His public explanation, delivered with characteristic terseness: “After that, I was ok leasing Colossus 1 to Anthropic, as SpaceXAI had already moved training to Colossus 2.” Or in other reporting: “no one at Anthropic set off my evil detector.”
Both companies got something they urgently needed from each other. The fact that public posturing had to be set aside to make the deal happen tells you more about the state of AI compute in May 2026 than any individual quarterly earnings number.
What Each Side Got
The deal is structurally simple. Anthropic gains immediate access to one of the world’s largest AI compute clusters at a moment when its other compute commitments do not produce capacity for another six to twelve months. SpaceX, having absorbed xAI in a $1.25 trillion merger earlier this year, gets to monetize a data center that xAI’s own training operation no longer needs.
The numbers matter. Anthropic’s prior compute announcements include:
- A 5 gigawatt agreement with Amazon, with nearly 1 GW of new capacity by the end of 2026
- A 5 gigawatt agreement with Google and Broadcom, with capacity coming online in 2027
- A strategic partnership with Microsoft and Nvidia, including $30 billion of Azure capacity
- A $50 billion investment in American AI infrastructure with Fluidstack
Each of these is enormous on its own. Stacked together they constitute the largest forward compute commitment in the history of any AI company. But every one of those agreements has a delivery timeline that runs into late 2026 or 2027 at the earliest. The capacity is real. It just is not real yet.
That gap is the problem the SpaceX deal solves. Anthropic needs compute now, not in eighteen months. Claude usage has grown faster than Anthropic anticipated, with CEO Dario Amodei publicly stating in early May that the company projected 10x revenue growth and saw 80x instead. Paid subscribers have been hitting rate limits. Claude Code users were experiencing peak-hour reductions that frustrated power users. The company needed a compute bridge between today and when its 5 GW capacity contracts come online. SpaceX had a 220,000-GPU bridge sitting in Tennessee, partially idle.
The first practical consequence was visible within hours of the announcement. Anthropic doubled the five-hour rate limits on Claude Code for Pro, Max, Team, and seat-based Enterprise plans. It removed the peak-hours reduction for Pro and Max users. It “considerably” increased API rate limits for Claude Opus models. The compute is not just promised. It is being put to work this week.
For SpaceX, the deal is the clearest signal yet that the company is becoming a real AI infrastructure provider. xAI is reportedly running flops utilization on Colossus 1 around 11%, far below the roughly 40% achieved by rival AI clusters, according to reporting from The Information. xAI moved its own training to Colossus 2. The capacity at Colossus 1 was sitting underused. Leasing it to Anthropic converts an asset that was destroying enterprise value into one that monetizes against the largest forward compute demand in the AI industry.
It also gives SpaceX a marquee customer reference for its broader AI infrastructure ambitions. Cursor, the AI coding company SpaceX is in talks to acquire for $60 billion, is reportedly already using Colossus capacity. With Anthropic now signed, SpaceX has effectively launched a third-party AI compute business inside the xAI shell. The implications for SpaceX’s IPO narrative are direct. The company is not just building rockets and satellites. It is building one of the largest non-hyperscaler compute providers in the United States, on a strategic timeline that aligns with its IPO targeted for later this summer.
The Subtext Almost No One Is Mentioning
Two layers of context make this deal more striking than the simple fact that two rivals are now doing business together.
The first is the ongoing litigation between Anthropic and the Trump administration. In March, the Pentagon designated Anthropic a “supply chain risk” and effectively blacklisted the company from working with the U.S. military. Anthropic responded by suing the administration in San Francisco and Washington. That litigation is ongoing. Meanwhile, the Defense Department has been actively integrating xAI’s Grok model into its workflows. Musk has been a vocal supporter of the Trump administration. And Anthropic just signed a deal with the Musk-controlled company whose AI model the Pentagon is currently choosing over Anthropic’s. The political dynamics there are unusual. Both sides apparently set them aside.
The second layer is the local controversy at the Memphis site itself. Colossus 1 sits in Boxtown, a historically Black neighborhood. xAI installed dozens of natural gas turbines to power the facility, claiming no federal permit was required because the equipment was temporary. Civil rights groups, including the NAACP, have been protesting the air quality impacts since 2024. Hours after the Anthropic announcement, the NAACP’s Director of Environmental and Climate Justice issued a statement criticizing the deal: “Any company that disregards the obvious environmental and health concerns of Black communities to supposedly power a future that will help us all is sending a clear message about who it intends to serve.” Anthropic’s public response includes a commitment to “cover any consumer electricity price increases caused by our data centers in the US.” How the company handles the existing turbine controversy in Memphis specifically is not yet detailed.
Both subtexts suggest that the urgency to close the deal was high enough to override considerations that, in any normal industry, would have killed it.
What This Tells You About AI Compute Right Now
Strip out the personalities, the politics, and the local controversy. Three structural reads remain.
The first is that the largest AI lab in the world by ARR (Anthropic now exceeds OpenAI in annualized revenue) is so compute-constrained that it has signed deals with every conceivable partner. Amazon, Google, Microsoft, Nvidia, Fluidstack, and now SpaceX. The fact that the SpaceX deal happened despite the political and personal tensions between the two companies tells you that traditional considerations are being subordinated to the compute supply problem. In 2026 AI, compute scarcity is the constraint that overrides every other strategic preference, including who you have to do business with to solve it.
The second is that SpaceX is positioning itself as an AI infrastructure provider with a credible offering that hyperscalers cannot easily replicate. Memphis was built fast (xAI launched Colossus in record time and has built capacity faster than any single hyperscaler in the same period). The 220,000-GPU cluster is one of the largest single-site GPU concentrations in the world. SpaceX’s broader vertical integration (Starlink, Starship, Terafab, and now data centers at scale) gives it a unique stack. The SpaceX IPO narrative just got a major data point. The company is not just selling a future. It is selling current capacity to the customer that is currently winning the enterprise AI race.
The third is that the orbital compute interest, while still in the “expressed interest” stage, is no longer a curiosity. SpaceX has been signaling for months that 80% of its planned chip output from the Terafab project would go to satellites in low Earth orbit. The company has filed with the FCC for a license to launch up to one million data center satellites. Anthropic’s “interest” in partnering on multiple gigawatts of orbital compute capacity is the first time a frontier AI lab has publicly committed to evaluating space-based AI infrastructure as a real procurement option. If even one of those gigawatts gets built and operational in the late 2020s, it changes the structure of where AI is computed for the next several decades.
The companies that benefit from the structural pattern this deal exposes are not necessarily the ones in the headlines. The frontier AI labs benefit because compute is unlocking. The hyperscalers benefit because they are the only entities with the scale to deliver on the multi-year capacity commitments. The chip suppliers (Nvidia, Broadcom, Marvell, Astera Labs, the connectivity ecosystem) benefit because every gigawatt of new capacity multiplies their addressable market. The launch and space companies benefit because the orbital compute story moves from speculation toward procurement. SpaceX is the company that captures multiple of those buckets in a single integrated stack.
What is genuinely new here is the demonstration that, when compute supply is the binding constraint of an industry, even the deepest public rivalries can be subordinated to it. The Anthropic-SpaceX deal is not friendship. It is not strategy. It is the clearest signal yet that the AI industry has hit the moment where there is not enough compute to go around, and the companies that need it most will sign with whoever has it, regardless of what they have said about each other in public.
That dynamic does not last forever. It lasts until the supply curve catches up with the demand curve, which on current build timelines is sometime in 2028 or beyond. Until then, the deals that get done are the ones that solve the compute problem. The Anthropic-SpaceX agreement is the most visible example of the playbook, and probably not the last one investors will see this year.
Read the full article here