On September 7, 2026, The Information reported that Anthropic has signed $517 billion in compute contracts over the past eleven months. In December 2025, the same company told investors it expected to spend around $180 billion renting servers through 2029. The number did not triple because someone typed too many zeroes. It tripled because the definition of "enough" moved.

Data Center Dynamics and Seoul Economic Daily both corroborated the figure the day it broke, with Amazon and Google together supplying roughly eleven gigawatts of the 14.8 total, and Microsoft, SpaceX, AMD, Lambda, Nscale, Fluidstack and Volta filling in the rest. Anthropic entered 2026 with somewhere between one and two gigawatts of live capacity. It leaves the third quarter having pre-committed to more than ten times that.

A number that doesn't fit the model

Anthropic's annualized revenue, per Bloomberg's July figure, is about $65 billion. Its expected IPO — Seoul Economic Daily reports late October or early November — will price against a fundraise that has to make eight years of the current run-rate look like a down payment on already-signed leases. That is not a company raising money to grow into a plan. That is a company signing the plan first and running the fundraise to catch up.

The individual deals, stacked chronologically:

• April 2026: expanded Amazon partnership, an additional 5 GW; Google and Broadcom multi-gigawatt agreement, Krishna Rao (CFO) calls it "a continuation of our disciplined approach to scaling infrastructure."
• May 2026: SpaceX/xAI-adjacent agreement at roughly $1.25 billion per month; total commitment now put at up to $45 billion.
• July 2026: $5 billion with AMD for compute.
• August 2026: $10 billion with Volta over six years in Norway; $45 billion with Nscale over six years, Vera Rubin chips in West Virginia coming online late 2027.
• September 1, 2026: $35 billion with Lambda.
• Plus: about $50 billion with Fluidstack for proprietary data centres in Texas and New York, direct chip purchases from Broadcom, Google and AMD, and roughly $30 billion of Microsoft Azure capacity (1 GW).

Add it up and you land in the same neighbourhood as $517 billion — with enough left over for the "possibility that the company could have signed other deals without making them public" that Data Center Dynamics flags in its write-up. Multiple contracts extend past 2030.

The word disciplined is doing a lot of work in Krishna Rao's April sentence.

The Dario paradox

Early in 2026, Dario Amodei told an interviewer that Anthropic's competitors "don't really understand the risks they're taking." He said it about capital pace, not model behaviour. The frame was that other frontier labs were making commitments they could not underwrite.

Nine months later, Anthropic has locked in more compute than it has raised money for. The company's CEO has spent the same period publishing safety essays, appearing before Congress, and putting the "existential risk" register on the table as a policy tool. Both things are true at once, and neither cancels the other: you can believe that scaling responsibly is the safer path and still be the entity that scaled hardest, if your read is that whoever gets to the frontier first sets the rules.

But the tension is worth naming. The public position — that everyone else is moving recklessly — is now paired with a private ledger that dwarfs everyone else's disclosed commitments except OpenAI's target of up to $750 billion for 30 GW by 2030. Sam Altman, for his part, described the neo-cloud sector this summer as "unsustainable silliness." Two of the loudest voices about compute-market froth are simultaneously the two largest buyers.

The IPO squeeze

Neil Campling of Mirabaud Securities put it plainly to 24/7 Wall Street after the Lambda deal on September 1: "Anthropic needs so much compute power. So they're looking to get it from any way that they can." And on the funding side: "This is another example of how we're likely to need that IPO to raise money."

The IPO becomes not a milestone but a mechanism. If Anthropic prices in late October or early November against a revenue run-rate of $65 billion, it needs the market to underwrite a decade of leases already on the books at a multiple that treats those leases as competitive moat rather than fixed drag. Every quarterly earnings call for the rest of the decade will be graded against the pace at which those 14.8 gigawatts convert into billed inference. Miss the ramp and the same leases become the story.

Jensen Huang, whose customers are on both sides of these deals, described the shape of the flywheel to shareholders this quarter: "we get paid twice, once on the hardware sale and again through the share of rental revenue." Nvidia's data-centre revenue for the quarter came in at $89 billion, up more than a hundred percent year-on-year. The neo-cloud tenants signing decade-long leases and the chip vendor selling into their fit-out are, in a real sense, the same P&L viewed from opposite ends.

The compute-buildout beat, one bullet point

For readers following [[AI compute buildout - Loop beat]]: the change here is not another gigawatt-scale announcement — those had become weekly. The change is that when you total the announcements plus the leaked contracts, Anthropic alone is now the single largest compute counterparty in the market, ahead of Meta's 2030 target and second only to OpenAI's aspirational 30 GW. And Anthropic is doing it from a base of one to two gigawatts as recently as this January.

There is no equivalent moment in enterprise IT history. AWS spent its first decade building the compute the industry then rented; Anthropic is renting the compute the industry has not yet built, on the bet that both the demand curve and the fabrication supply chain will meet the schedule. Fluidstack's Texas and New York proprietary sites, the Nscale West Virginia build, and the Amazon and Google expansions all carry 2027-2030 online dates. The company is not compute-constrained in the second half of 2026. It is compute-constrained in the second half of the decade, in advance.

What to watch

1. The S-1 disclosure of lease obligations. Anthropic's IPO filing will have to enumerate these commitments as operating-lease and take-or-pay liabilities. That table is the first place the $517 billion becomes forensically legible rather than journalistically reported.
2. The gap between announced and unannounced. DCD flagged that private contracts likely exist. The delta between the sum of announced deals and the reported $517 billion is the size of the shadow book.
3. Whether OpenAI counters with a matching signed number. Its $750 billion / 30 GW figure remains a target, not a contract stack. If it converts targets to signed leases in the next 60 days, the framing shifts from "Anthropic is out over its skis" to "both labs are, together, out over the industry's fabrication capacity."
4. The Amodei rhetoric. If the public tone shifts — from "rivals don't understand the risk" to something quieter — that is the tell that even the buyer thinks the buyer has moved too fast.

The last twelve months of AI capex have been a competition to sign the biggest contract before the S-1 has to explain it. Anthropic just took the lead. The IPO will settle whether that was the right lead to take.