On May 6, 2026, on stage at Anthropic's first developer conference in San Francisco, Dario Amodei said the quiet part out loud. The company had planned for 10-fold annualised growth in Q1; it got 80-fold. "Just crazy," he told the room, "too hard to handle." The compute story he announced two minutes later is what 80x looks like when you can't fix it with planning.

Anthropic signed an agreement to rent every watt of compute at SpaceX's Colossus 1 data center in Memphis, Tennessee. "More than 300 megawatts of new capacity (over 220,000 NVIDIA GPUs) within the month," in the company's own words. The mix runs across H100, H200, and Blackwell GB200 accelerators, per The Decoder's reporting. Financial terms were not disclosed.

Translated: Anthropic is renting the cluster Elon Musk's AI lab built six months ago, because Musk's AI lab — now SpaceXAI, post the $1.25 trillion February merger that absorbed xAI into SpaceX — has already moved into Colossus 2 next door, the gigawatt successor that came online in January and scaled to 1.5 GW in April. Colossus 1 is yesterday's frontier cluster. Anthropic is renting yesterday's frontier cluster.

That is not an insult. 220,000 H100/H200/GB200 GPUs is a serious supercomputer in any year, and Memphis is one of the most contested bits of physical infrastructure in the AI build-out — the data center whose load drew a Shelby County energy debate when Colossus 2 came online. Anthropic is now the entity buying the power for the older sibling.

Why Anthropic needs it

Pull on Amodei's "80x" line and the deal's logic falls out. Anthropic's public framing for the past year has been that the bottleneck on revenue is compute, not demand. Q1 made that real. If the Q4 2025 plan was to grow tenfold and the actual delivered run-rate was eight times that, the gap is not closed by ordering more Trainium chips that ship in 2027. The gap is closed by finding hot capacity in May.

Hot capacity in May is exactly what Colossus 1 was. SpaceXAI already paid to build the building, plug in the substations, and rack the GPUs. The marginal cost of selling that block to Anthropic is somewhere between zero and the operating bill, against a brand-new revenue line for SpaceX in the quarter before its June IPO roadshow. That is an efficient transaction for two companies that both want to walk into their public listings with a "demand exceeds supply" story to tell.

How each side framed it

Anthropic's announcement was operational: Claude Code's five-hour rate limits doubled across Pro, Max, Team, and Enterprise plans, peak-hour throttling on Pro and Max gone, Claude Opus API rate limits raised "considerably." That is the customer-visible payoff and the only number the trade press picked up cleanly.

Musk's framing, posted to X and quoted across Al Jazeera, the Cryptopolitan recap, and CNBC's writeup, was less operational. "Everyone I met was highly competent and cared a great deal about doing the right thing," he wrote. "No one set off my evil detector. So long as they engage in critical self-examination, Claude will probably be good."

That is one of the more entertaining sentences a vendor has ever put out about a B2B compute lease. It is also load-bearing. The same post — and Yahoo Finance's wire confirms the wording — included a clause that lets Musk reclaim Colossus 1 if Anthropic "engages in actions that harm humanity." The kill switch is in the contract, written in the language of an AI-safety worry.

A reader can pick the reading. The charitable one: a serious compute provider acknowledges out loud that the workload is novel and reserves an option a normal cloud lease would not. The cynical one: the largest single GPU lease of the year just acquired a unilateral termination clause whose trigger is whatever Musk decides counts as harm to humanity, in the same week he is in active litigation with OpenAI over what he believes Sam Altman did to a charity. Both readings can be right at once.

The orbital pitch

Buried in the same Anthropic post is a sentence that costs many gigawatts: "As part of this agreement, we have also expressed interest in partnering with SpaceX to develop multiple gigawatts of orbital AI compute capacity."

There is no orbital data center. There is no contract for an orbital data center. There is a press-release-friendly hand-wave at the idea, paired with the company that has the only capable launch fleet on Earth and a chairman who has been pitching space-based compute as the only sustainable answer to AI's energy budget. The line tells you nothing about timelines and everything about the optics each side wanted on day one: Anthropic gets to be the Earth-bound AI lab thinking about orbital infrastructure; SpaceX gets a frontier-AI customer signing onto a launch-economy thesis right before its IPO. It is a sentence that does not commit either side to anything and lands as a headline anyway.

What this stacks against

This is now the second time in two weeks Anthropic has bought capacity at a scale that makes the previous "biggest deal ever" line look small. In April it closed Round Three with Amazon at up to 5 GW, with nearly 1 GW available by year-end. The published portfolio also includes a 5 GW Google/Broadcom agreement starting 2027, a $30 billion Microsoft/NVIDIA Azure capacity tranche, and a $50 billion Fluidstack American-AI-infrastructure commitment. The SpaceX deal is the only one of those that delivers in May 2026. Everything else delivers later.

That is the reason this matters. Anthropic's 2027–2030 compute graph is full. Its 2026 graph had a gap, and the gap was the thing customers were noticing — peak-hour throttling on Claude Code, capacity errors on Opus, the friction the 80x growth number explains. The Colossus 1 lease is the patch over the next two quarters. After that, Trainium racks land at Project Rainier and the in-house compute story is the one Anthropic wants to tell. Until then, Anthropic is buying time on an Nvidia cluster that Musk built for a different lab.

What to watch

1. Whether the Colossus 1 capacity actually shows up in product within thirty days. Anthropic's announcement says "within the month." That would be exceptional. The honest measurement is whether Claude Code's peak-hour throttling stays off for paying users in late May and early June. If the rate limits regress on a Tuesday, the deal slipped.
2. The ethics-clause precedent. A "we can take the GPUs back if your AI harms humanity" clause has not, to our knowledge, appeared in any prior public hyperscaler lease. If the next Anthropic disclosure or a Q2 earnings call surfaces equivalent language in another compute contract — Amazon, Google, Microsoft — the bar has moved on what counts as standard AI-vendor due diligence. If it stays unique to SpaceX, it is a Musk artefact, not a market norm.
3. The IPO double-bill. Both companies are listing this year. SpaceX's roadshow targets June at a $1.75–$2 trillion valuation; Anthropic's window is the second half, with secondary trades printing in the trillion-dollar range on Forge. Each side's S-1 prospectus will reference the other company's contract by name. This is the first lease that both parties will need to disclose to the SEC.

The frame the wires reached for this week was "SpaceX strikes data center deal with Anthropic." That undersells it. The story is that one of the two AI labs still in the conversation rented yesterday's frontier supercomputer from the other one's owner, with a contractual ethics clause, an orbital sales pitch, and an evil-detector quote attached, four weeks before both parties try to take their books to Wall Street. Compute is now the product roadmap, the politics, and the bond pitch — sometimes in the same sentence.