Anthropic Leases All of SpaceX's Colossus 1 to Ease Claude Capacity Crunch

Photo by Mehan Talukder on Unsplash

Anthropic has taken over the entire compute capacity of Elon Musk's Colossus 1 data center, an arrangement that would have been almost unthinkable three months ago when Musk was publicly accusing the company of hating Western civilization. The Anthropic SpaceX Colossus compute deal, announced May 6 at Anthropic's developer conference in San Francisco, gives Claude access to a Memphis facility packed with NVIDIA hardware that until recently was training Grok. It is the strangest pairing in the AI infrastructure race so far, and it is already changing what paying Claude users get.

What the deal actually does

Anthropic confirmed in its announcement that the agreement provides access to more than 300 megawatts of new capacity, equivalent to over 220,000 NVIDIA GPUs, coming online within the month. The hardware mix matters here. According to SpaceXAI's partnership page, Colossus 1 runs a combination of H100, H200, and next-generation GB200 accelerators, and the company describes it as one of the world's largest and fastest-deployed AI supercomputers.

The site itself sits in Memphis, Tennessee. CNBC reported that Colossus 1 was powered partly by natural gas-burning turbines, which drew protests from local residents over air quality long before Anthropic's logo got attached to the facility. Anthropic now inherits that political baggage along with the GPUs. For a company whose public messaging leans heavily on responsible AI development, the Memphis turbines are a complication that will not stay quiet.

The handover was possible because xAI no longer needs the building. Data Center Dynamics reported that xAI had already shifted training workloads to Colossus 2 before leasing the original site to Anthropic, with Musk confirming the move on X. The Colossus 2 site was purchased in March 2025 and came online in January 2026. Data Center Dynamics also noted that in April 2026, xAI had previously announced Colossus capacity would be used by Cursor, suggesting the building's path to becoming a multi-tenant facility had been in motion for weeks before the Anthropic deal closed.

Why Claude users were hitting walls

The practical reason for the deal is that Claude has been buckling under its own popularity. TechStartups reported that Anthropic experienced 80x growth in Q1 2026, causing reliability issues for Claude Pro and Max subscribers throughout the quarter. That is not a sustainable curve to ride on borrowed AWS Trainium and Google TPU capacity alone, particularly when the heaviest users are running long-horizon agentic sessions that pin GPUs for hours at a time.

Anthropic's head of product Ami Vora used the May 6 event to spell out what the new capacity buys customers. Axios reported that peak-hours usage limit reductions on Claude Code were removed for Pro and Max users effective the same day. The change goes deeper for developers running heavy agentic workloads.

CoinDesk reported that Claude Code's five-hour rate limits doubled for Pro, Max, Team, and Enterprise plans starting May 6, with Claude Opus API rate limits raised significantly as part of the same announcement. For anyone who has watched a Claude Code session die mid-refactor at the four-hour mark, that is the headline number. The combination of doubled session limits and lifted peak-hours throttling is the most concrete user-facing change Anthropic has shipped in months, and it is directly traceable to the Colossus capacity coming online.

How Colossus fits into Anthropic's compute stack

This is not a one-off. Anthropic has been signing capacity deals at a pace that suggests its planners believe demand keeps compounding through 2027. The company's existing arrangements include a 5 GW agreement with Amazon, with nearly 1 GW expected online by the end of 2026, plus a 5 GW agreement with Google and Broadcom set to come online in 2027. There is also a Microsoft-NVIDIA strategic partnership covering $30 billion of Azure capacity, and a $50 billion U.S. AI infrastructure investment with Fluidstack.

Against that backdrop, 300 megawatts is small. What it is not is slow. Colossus 1 comes online in weeks rather than the multi-year build cycles attached to the Amazon and Google deals, which is why Anthropic was willing to take a short-term lease from a landlord whose CEO was publicly hostile a quarter earlier. The Amazon and Google commitments will eventually dwarf Colossus, but neither will help a Claude Max subscriber whose session is timing out tonight.

The hardware diversity is the quieter part of the story. Anthropic trains and runs Claude on AWS Trainium, Google TPUs, and NVIDIA GPUs, and Colossus 1 plugs cleanly into the GPU side of that stack without forcing the model team to retarget anything. That portability has become a strategic asset on its own. Labs that committed early to a single silicon vendor are now negotiating from a weaker position when capacity gets tight.

The Musk reversal

The political subtext is hard to ignore. CNBC noted that Musk had previously called Anthropic "misanthropic" and accused the company of hating Western civilization in February 2026. One month later, in March 2026, the Pentagon declared Anthropic a supply chain risk and blacklisted it from military contracts. Whether those two events were related has not been publicly established, but the timing was uncomfortable enough that some industry observers connected them.

Then the IPO calendar arrived. Axios described SpaceX's expected IPO this fall as slated to be the largest in corporate history. Renting Colossus 1 to Anthropic at scale is a clean way to demonstrate that SpaceX's compute infrastructure has external demand from buyers who are not named xAI. Investment bankers preparing the prospectus have a tangible commercial relationship to point at, and Anthropic gets the GPUs it needs.

The timing on the SpaceX side is no coincidence. CNBC also flagged that xAI announced May 6 it will be dissolved as a separate company, with the combined entity now called SpaceXAI. The Colossus 1 lease is effectively the first major commercial transaction the new entity is announcing, and it lands during the same week as the rebrand. According to The Information cited by TechStartups, xAI's Grok model was achieving roughly 11 percent flops utilization versus around 40 percent at rivals, which goes some way toward explaining why those GPUs were available to lease at all. Empty silicon is a balance-sheet problem, regardless of who owns it.

CNBC also reported that Anthropic is in talks to raise funds at a $900 billion valuation. Compute commitments at the 5 GW scale need that kind of equity story behind them, and the Colossus lease is, among other things, evidence to investors that Anthropic can move fast when reliability is on the line. A lab that cannot keep its paying users online during a growth spike does not get valued at $900 billion, no matter how good the model benchmarks look.

The orbital wrinkle nobody is talking about

Buried inside SpaceXAI's announcement is a sentence that reframes the entire deal. SpaceXAI noted that Anthropic expressed interest in partnering with SpaceX to develop multiple gigawatts of orbital AI compute capacity. That is not a throwaway line. It hooks Anthropic into Musk's broader thesis that terrestrial AI training will run out of power and water before it runs out of demand, and that the next frontier is space-based compute powered by near-continuous solar exposure.

For a company that has built its brand on AI safety, signing on to space-based compute with the world's most controversial CEO is a position Anthropic would not have taken lightly. The constraint driving it is power. Terrestrial AI training is hitting interconnection queues, water permits, and grid limits faster than any model team can absorb. Whether the orbital bet pays off is a separate question, but Anthropic is now publicly inside the tent and will be expected to either follow through or quietly walk back the language in a future filing.

What this signals about Anthropic's 2026

The Colossus deal arrived in the same week as another sizable announcement. SiliconRepublic reported that Anthropic partnered with Blackstone, Hellman & Friedman, and Goldman Sachs to form a new AI services company, with future capacity expansion targeting international infrastructure for data residency compliance in financial services, healthcare, and government sectors.

That is the enterprise playbook in plain view. Regulated buyers care about latency, residency, and uptime. They do not care which billionaire's data center is processing their payroll inferences, as long as the contract clears legal review and the SLAs hold. By stacking a fast-turnaround GPU lease on top of long-term hyperscale commitments and a new financial-services joint venture, Anthropic is building a compute story that lets enterprise sales teams promise both capacity today and residency tomorrow.

The Pentagon situation remains the awkward outlier. Anthropic is now buying compute from a Musk-controlled entity that holds extensive defense contracts, while simultaneously sitting on a March 2026 supply-chain blacklist from the same Department of Defense. Whether that creates pressure to revisit the blacklist, or simply hardens it, is one of the more interesting open questions the deal creates.

What to watch next

Three things are worth tracking over the next 60 days. First, whether Claude Pro and Max users actually see the reliability improvements promised, or whether 300 megawatts gets absorbed inside a week of Q2 demand. Anthropic's 80x Q1 growth figure suggests the latter is a real risk. Second, whether the SpaceX IPO prospectus names Anthropic as a customer, since that disclosure would lock in the commercial relationship beyond a one-quarter handshake. Third, whether the Pentagon revisits its March blacklist now that Anthropic is buying compute from a Musk-controlled entity that holds extensive defense contracts.

The deeper question is whether this deal is a template or an exception. If Anthropic can lease 300 megawatts in a month from a competitor's spare data center, every frontier lab now has a new option for absorbing demand spikes without waiting on multi-year hyperscale builds. The Cursor deal that Data Center Dynamics flagged from April suggests the model is already replicating. The economics of the Colossus 1 turbines and the air-quality fights in Memphis suggest that option comes with costs the press releases will not mention, and that local communities sitting next to spare GPU capacity may end up bearing most of them.