THURSDAY, MAY 21, 2026 · BRISBANESUBSCRIBE →

THE AI POST

INTELLIGENCE. CURATED.

Rows of server racks in a large data center with blue LED lighting
May 19, 2026

Anthropic Is Running Claude on Elon Musk's Supercomputer. Yes, That Elon Musk.

Three months after Musk called Anthropic a threat to Western civilization, SpaceX is handing over its entire Colossus 1 data center to power Claude. 220,000 GPUs. 300 megawatts. The irony writes itself.

In February, Elon Musk declared on X that "Anthropic hates Western Civilization." He was sharing a false claim from a Trump administration official about the company's AI safety practices. Three months later, SpaceX is handing Anthropic the keys to its entire Colossus 1 data center in Memphis, Tennessee.

The deal, announced May 6 at Anthropic's Code with Claude developer conference in San Francisco, gives Anthropic access to more than 220,000 NVIDIA GPUs and 300 megawatts of compute capacity. The hardware includes dense deployments of H100, H200, and next-generation GB200 accelerators. In practical terms: this is one of the largest single compute deals in AI history.

Anthropic CEO Dario Amodei took the stage to frame this as a direct response to exploding demand. Claude Code users on Pro and Max plans immediately got doubled five-hour rate limits. Peak-hour throttling was removed entirely. API limits for Opus were raised. For the developers who have been screaming on Hacker News and Reddit for months about hitting walls mid-session, this was the answer.

The timing matters. This newspaper reported last week that Uber burned its entire 2026 AI budget in four months, largely on Claude Code. That is not an isolated story. Enterprise adoption of agentic coding tools has outpaced every infrastructure projection Anthropic had. Multi-agent workflows, which burn through compute at rates that make single-chat sessions look like pocket change, are now the default pattern for serious engineering organizations.

The SpaceX deal sits alongside a staggering run of compute agreements. Last month Anthropic signed deals with Google (up to $40 billion in investment), Amazon ($5 billion, earmarked for Amazon's custom chips), Microsoft, and NVIDIA. After subtly criticizing Sam Altman last year for hoarding compute contracts, Amodei has now out-Altman'd Altman. The man who warned about the AI arms race is in the middle of one.

And then there is the Musk angle, which is impossible to ignore. In the same week that a jury unanimously rejected Musk's lawsuit against OpenAI in under two hours, his own company became the compute backbone for OpenAI's biggest competitor. Musk's public explanation for the pivot was characteristically breezy: "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. No one set off my evil detector."

The business logic is straightforward. SpaceX built Colossus 1 for xAI's Grok models, but the facility has excess capacity. Anthropic needs more GPUs than it can build data centers for. SpaceX gets revenue from idle hardware. Anthropic gets immediate compute without the 18-month construction timeline. Both companies get what they want.

There is also a more speculative dimension. Both companies "expressed interest" in building orbital data centers together, with Anthropic noting that "compute required to train and operate the next generation of these systems is outpacing what terrestrial power, land, and cooling can deliver on the timelines that matter." Ars Technica's reporting notes that orbital data center economics remain unproven, but the fact that two companies with this level of resources are publicly exploring the idea suggests the compute ceiling is closer than most people realize.

The broader pattern is clear: AI demand is outrunning supply so fast that companies that publicly despise each other are making deals anyway. Ideology follows infrastructure. Musk's companies have now provided compute to the AI lab he tried to destroy in court, funded the AI company he called a threat to civilization, and lost every legal challenge he filed against the industry he helped create.

For Claude Code users, none of this corporate drama matters much. What matters is that the five-hour rate limit just doubled and peak throttling is gone. For everyone else watching the AI compute wars, this deal is a signal: the scramble for GPU capacity has gotten so desperate that the industry's most vocal enemies are now business partners. That should tell you something about where this is heading.