Elena VossEight.
That is the multiple by which Anthropic missed its own 2026 growth forecast. Dario Amodei told the New York Times the company had planned to grow 10x this year. Q1 came in at an 80x annualized rate. The forecast itself was already an aggressive number — Anthropic raised its 2026 revenue projection 20% in January, to $18 billion. Q1 said the raised forecast was still 8x too low.
A forecast miss of 8x is not a procurement problem. It is a structural one. Anthropic responded the way a company responds when it learns its own model of itself is wrong: by buying compute from every supplier that has it, including the one it was founded to escape.
The Vendor List
In the six months before May 6, Anthropic committed to four large compute deals and disclosed the fifth on the day. Each was disclosed separately. The combined list reads differently than any one of them.
A year ago, Anthropic made AWS its primary training and inference partner as part of a $4 billion Amazon investment. The Information has since reported Anthropic expects to pay Amazon, Google, and Microsoft a combined $80B+ to run its models. By April this year it had added Google and Broadcom for gigawatts of next-gen TPU. By May 6 it had added Musk.
The Three-Month Window
The procurement velocity is the second number in the story. The Information's $200B Google figure was disclosed on May 6. The SpaceX agreement was disclosed the same day. The 100%-of-Colossus-1 detail, including the 300+ MW available "within the month," came hours later.
Three months separate two facts. In February, Musk called Anthropic "evil." On May 6, SpaceX agreed to supply Anthropic's compute. The interval is shorter than most enterprise procurement cycles. It is shorter than the lead time on a single Nvidia rack.
The simplest explanation for the speed is that both sides had something the other needed urgently. xAI's Colossus 1, built in Memphis and expanded to a million GPUs, was sized for Grok. Reporting on the day of the deal noted that Grok never grew to utilize the capacity. The fastest-deployed AI supercomputer in the world had a demand problem. Anthropic, growing 80x, had the inverse one. They found each other.
The Founding Premise, Inverted
Anthropic's founding story, as documented by the New York Times in 2023 and corroborated in subsequent reporting, is in part a story about distance from Musk. The Amodei siblings and the early Anthropic team left OpenAI in 2021 over safety differences with the leadership shaped by — among others — the man who had been OpenAI's largest early funder and was now suing it. Anthropic's brand is "the safety lab." Its commercial moat, by its own framing, is being the lab that gets it right.
This is the company that just agreed to take 100% of Elon Musk's data center. The same week Mira Murati testified under oath that Sam Altman lied to her about safety standards on an OpenAI model, in a trial Musk himself was funding. The trial that re-litigates whether Musk was right to leave OpenAI ran on the same news cycle as the deal that made him Anthropic's compute landlord.
The forecast missed in one direction. The rivalry collapsed in the other.
The Counter
One reading of all this is that the compute deficit is positioning. AI labs have every incentive to frame demand as overflow — it justifies the capex, signals scarcity to enterprise buyers, and supports the valuation. Amodei describing 80x growth as "too hard to handle" is the kind of complaint a CFO would script. Saying you can absorb 300 MW within the month, on capacity a competitor failed to fill, is a flex.
The flex reading has real weight. But two facts cut against it. First, the financial commitment is not a posture — $200B to Google, $30B to Microsoft, and a "100% of Colossus 1" lease are line items on actual contracts, not investor-day talking points. Second, the supplier list now includes the one supplier Anthropic had structural reasons to avoid. Companies do not lease compute from people they were founded to escape unless the alternative — slowing growth, missing the demand window, ceding the enterprise share — is worse.
Compared to What
The base rate for compute deficits in this cycle is high. OpenAI has been on its own procurement march — Oracle, Nvidia, AMD, Broadcom, Microsoft, more recently a joint MRC protocol with five of them. Google is reporting $400B+ of remaining performance obligations. Microsoft has been quietly delaying some of its own data center expansion. Demand is exceeding the buildout in some places and lapping it in others.
What is unusual about Anthropic's May 6 disclosure is not that the deficit exists. It is the magnitude of the forecast miss that drove it. 10x was the plan. 80x was the realization. The 8x ratio between the two is the structural fact. Everything else — the Google commitment, the SpaceX deal, the suspended hostility with Musk, the headlines about "evil" turning into headlines about partnership — follows from a company learning, in a single quarter, that its own model of itself was off by an order of magnitude.
A growth-rate miss of that size doesn't get fixed by hiring a procurement team. It gets fixed by buying compute from whoever has it, including the supplier you were founded to avoid. The number to remember is eight. Forecast misses of that magnitude rearrange more than infrastructure budgets. They rearrange founding premises.