Anthropic Signs Deal for All of xAI Colossus One Data Center Capacity in Unlikely Rivals Partnership

--- headline: "Anthropic Signs Deal for All of xAI Colossus One Data Center Capacity in Unlikely Rivals Partnership" slug: anthropic-xai-colossus-compute-deal category: business story_number: "03" date: 2026-05-10 ---

In one of the most unexpected partnerships in recent AI history, Anthropic has struck a deal with SpaceX subsidiary xAI to lease the entire compute capacity of the Colossus 1 data center in Memphis, Tennessee -- gaining access to more than 220,000 NVIDIA GPUs and over 300 megawatts of power that could reshape the competitive landscape of frontier AI development.

The agreement, announced May 6 at Anthropic's Code w/ Claude event, hands the safety-focused AI lab a massive injection of compute at a moment when GPU capacity remains one of the most constrained resources in the industry. For xAI, the deal transforms an underutilized asset into a revenue engine, though it raises pointed questions about the company's ambitions as an AI model developer just as parent company SpaceX barrels toward a public listing.

The Numbers Behind the Deal

Colossus 1, which xAI built at breakneck speed in late 2024 and early 2025, houses a formidable cluster of NVIDIA H100, H200, and next-generation GB200 accelerator chips spread across its Memphis campus. The facility's 300-plus megawatts of capacity make it one of the largest single-site AI compute installations in the world.

While neither side has disclosed the financial terms, analyst estimates vary widely. New Street Research, cited by Fortune, pegs the arrangement at $3 billion to $4 billion in annual revenue for xAI. CryptoBriefing projects a higher range of $5 billion to $6 billion per year. Either figure would represent a significant new revenue stream for SpaceX ahead of its anticipated IPO.

Elon Musk, who previously enjoyed calling Anthropic "Misanthropic" on social media, struck a notably different tone this time. "I 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 was impressed," Musk posted on X. He added that he was comfortable leasing Colossus 1 because "SpaceXAI had already moved training to Colossus 2."

Musk followed up with a pointed caveat: "We reserve the right to reclaim the compute if their AI engages in actions that harm humanity." The criteria for what constitutes such harm, presumably, would be decided by Musk himself -- a detail that tech commentator Simon Willison flagged as "a new form of supply chain risk for Anthropic."

Why Anthropic Needs the Capacity

Anthropic has been visibly compute-constrained for months. The company's Claude models have surged in popularity among enterprise customers and developers, but usage limits have frustrated power users. Alongside the xAI deal, Anthropic announced higher usage limits for Claude subscribers -- a move that would be impossible without a substantial bump in available infrastructure.

The deal gives Anthropic access to the new capacity within weeks rather than the months or years required to build a data center from scratch. In the current AI arms race, where training runs for frontier models can consume tens of thousands of GPUs for months at a stretch, speed of access matters as much as raw scale.

The xAI Dilemma

For xAI, the partnership is both a lifeline and a concession. TechCrunch's Equity podcast hosts did not mince words about the implications. "This seems like a major heat check before the IPO," said reporter Sean O'Kane, noting that while becoming a neocloud -- a company that buys GPUs and rents them out rather than using them for its own model training -- might be "a more believable business in the near term," it is less likely to excite long-term investors.

Kirsten Korosec, also on the podcast, observed that when companies have a choice between renting out data center capacity and training their own frontier models, they overwhelmingly choose the latter. The fact that xAI is leasing its entire Colossus 1 facility "suggests that maybe xAI isn't doing so much on the AI model training" front.

The timing adds another layer. Just days before the Anthropic announcement, xAI sent deprecation notices for Grok 4.1 Fast and several other models, giving developers less than two weeks before shutdown. Internally, reports have surfaced that xAI employees were using rival models rather than Grok for their own work. All of xAI's co-founders except Musk have departed the company, and Musk has signaled plans to dissolve xAI as a separate entity and fold it into SpaceX under the new moniker "SpaceXAI."

Environmental Baggage

The deal also carries environmental risk for Anthropic's reputation. Colossus 1 has been dogged by controversy since its construction. The facility's gas turbines initially operated without Clean Air Act permits or pollution control devices, classified under a "temporary" exemption. Credible reports have linked the site to increased hospital admissions in the Memphis area related to poor air quality, and xAI faces an active environmental lawsuit over the facility.

Data center analyst Andy Masley, who has pushed back on exaggerated environmental claims about AI infrastructure elsewhere, was blunt about Colossus 1 specifically: "I would simply not run my computing out of this specific data center."

For Anthropic, a company that has built its brand around safety and responsible development, the optics of powering Claude with a facility under environmental litigation present a communications challenge that no amount of compute can solve.

What Comes Next

The partnership signals a broader shift in the AI industry. As the gap widens between companies that can build and fill massive GPU clusters and those that need compute but lack infrastructure, deals like this one will become more common. SpaceX's looming IPO will test whether investors see neocloud revenue as a strength or an admission that xAI's model development has stalled.

For Anthropic, the calculus is simpler: compute now, reputation management later. In a race where the next training run could determine market leadership, 220,000 GPUs are hard to turn down -- no matter whose name is on the building.