Anthropic just leased the full compute capacity of Colossus 1, the Memphis supercomputer Elon Musk built last year to compete with it. More than 300 megawatts. More than 220,000 Nvidia GPUs across H100, H200, and GB200 generations. All of it coming online for the maker of Claude within a month, per the joint announcement on May 6, 2026.
Three months ago, Musk was on X calling Anthropic “misanthropic” and accusing the company of hating Western civilization. This week, after a few sit-downs with CEO Dario Amodei’s team, he downgraded the rhetoric to “no one set off my evil detector.” For that vibe shift, SpaceX gets between $3 billion and $4 billion in annual revenue from the deal, per estimates from New Street Research analyst Antoine Chkaiban, with more than $2.5 billion of that running through to cash profit. Principle held up for as long as the price was wrong.
Why Anthropic Wrote the Check
The blunt answer is that demand has outrun Anthropic’s compute supply. Amodei told a developer conference Wednesday the company saw eighty times year-over-year revenue and usage growth in Q1 2026, against a plan that assumed ten times. When you mis-forecast demand by an order of magnitude, you do not have months to sit through a procurement cycle. You take the warm GPUs that already exist, in a building that is already powered up, and you sign the lease.
Colossus 1 is exactly that asset. Built last year to train xAI’s Grok models, the Memphis facility is already energized, already racked, and already pulling more than 300 MW from the grid. Anthropic gets the keys inside a month. Building greenfield capacity at this scale in 2026 is a multi-year project, gated by utility power approvals, transformer lead times, land entitlements, and a Nvidia allocation queue that nobody is jumping. Renting from a competitor’s already-built supercomputer is the only door open at the velocity Anthropic needs.
The signal users will feel is already showing. The company doubled Claude Code’s five-hour rate limits across all paid tiers, removed the peak-hours throttle for Pro and Max subscribers, and raised the API rate limits on Claude Opus models considerably. Read together, those are the specific symptoms of a company that has been rationing supply and just unlocked the next shipment.
Why Musk Took the Money
The other side of the trade is a roadshow that begins the week of June 8, per CoinDesk’s reporting on the IPO calendar. SpaceX filed confidentially with the SEC on April 1 targeting a $1.75 trillion to $2 trillion valuation. That is a number you defend with diversified revenue, not with one product line, and definitely not with a money-losing AI lab that was struggling to keep its own data center utilized.
That money-losing AI lab is now the same Colossus 1 footprint it was three months ago, except that on paper it is a fully-subscribed infrastructure asset generating real annualized revenue from the most commercially serious enterprise AI workload in the market. Independent analyst commentary at Next Big Future has already pointed out the obvious read: at this run-rate, the former xAI compute business inside SpaceX breaks even on the back of one Anthropic contract. That is a useful narrative to carry into a $2 trillion roadshow.
What makes the trade easier still is that xAI no longer exists as a separate company. Musk folded it into SpaceX in February, and the combined entity was valued at $1.25 trillion in that transaction. The same compute that was once “subsidized R&D for Grok” is now “infrastructure-as-a-service for Anthropic.” The language change is doing real work when the investment bankers price the offering.
The Real Story Is the Megawatt
Step back from the personalities and the deal exposes the actual bottleneck in AI right now. It is not models. It is not researchers. It is the megawatt. Anthropic, by every public metric the most commercially serious frontier lab outside of OpenAI, was capacity-constrained badly enough to lease a building from a competitor whose CEO had spent the previous quarter publicly insulting it. Companies do not write multi-billion-dollar checks to people they hate unless the alternative is worse.
The same build-out story is showing up across the stack. Datadog crossed a billion dollars of quarterly revenue this week on the back of AI-workload observability spend, per BTN’s reporting on the Q1 print. Nvidia continues to clear allocations on its hyperscaler queue every quarter. The dividing line is being drawn between companies that own or can rent the megawatts and the rest of the market, which is paying full price to use them.
Anthropic’s parallel interest in working with SpaceX on multiple gigawatts of orbital compute capacity, called out in Al Jazeera’s coverage of the deal, is the same thesis projected forward. If the compute roadmap requires gigawatt-scale facilities and the terrestrial grid cannot deliver them in time, the next negotiation runs with whoever can put racks somewhere the grid does not need to be involved.
What’s Left Open
Two questions hang over the deal that the press release does not address. The first is federal. The Pentagon declared Anthropic a supply chain risk in March, and last week’s federal AI test bed agreements went to Microsoft, Google, and SpaceX-AI without Anthropic in the room. The company now buys the bulk of its compute from one of the three vendors that did make the cut. Investors writing into Anthropic’s reported $900 billion round will want to understand how that triangle is supposed to resolve.
The second is the safety brand. Anthropic has built its public posture on the argument that a more cautious frontier lab can still beat the industry’s most aggressive operator on the commercial scoreboard. That argument got harder to tell this week. When the same building is producing your inference and your most strident critic’s rhetorical victory tour, “principled alternative” needs an updated footnote.
The market does not appear to mind. Anthropic gets the GPUs. SpaceX gets the revenue. Musk gets a story to carry into the IPO. Whatever any of them said about each other in February will be edited out of the prospectus.