Anthropic Committed About $90 Billion in Four Weeks and Did Not Buy a Single Computer

Between August 4 and September 1, one AI lab committed roughly $90 billion. Not one dollar of it bought hardware.

It bought time on someone else's machines, in someone else's building, drawing someone else's power.

The three deals, in order:

- August 4: about $10 billion over six years to Volta, for Nvidia powered cloud delivered out of Bitdeer's Norway data center. - August 27: $45 billion over six years to lease 460 MW of Nvidia Vera Rubin capacity at Nscale's Monarch campus in West Virginia. - September 1: an estimated $35 billion over six years for 350MW from Lambda, at a new Texas data center leased by Nvidia and developed by Hut 8.

Four weeks. Three counterparties. Same six-year term on all three.

## A lease is not a purchase

This distinction gets flattened in the coverage, and it should not be.

When a lab buys GPUs, it owns a depreciating asset. It can resell it, redeploy it, write it down, or simply stop the next order. The decision is reversible and the exposure is bounded by what it already spent.

A six-year capacity commitment is a different animal. It is a dated, fixed obligation to pay, and it sits ahead of the buyer's own flexibility. Cash goes out whether or not demand shows up. In shape, if not in accounting treatment, it looks less like capex and more like a bond the company issued against its own forecast. We do not have the contract terms for any of these three, so treat that as a description of shape rather than a claim about the balance sheet.

The reason it matters: capex can be cut in a bad quarter. A signed six-year lease cannot.

## The vendor is on every side of the table

Look at the Lambda structure again. Nvidia backs the cloud provider. Nvidia leases the site. Nvidia sells the silicon that goes into it. One name, three positions, one transaction.

That is not an isolated arrangement. In July, Nvidia was reported to be negotiating a $250 billion financing guarantee to back OpenAI's lease and construction debt for a 10 gigawatt campus. Across the deals we track, Nvidia appears in 75 of them, 56 in 2026 alone.

When the supplier finances the buyer's obligation to the supplier, revenue and credit exposure stop being separate lines. Demand looks strong right up until the moment the financing has to be tested.

## The unit that actually matters is megawatts

Here is the number that made me write this. Of the 615 AI deals in our database, exactly 25 state a power capacity figure. Those 25 alone commit 58.2 GW, and 21 of them are compute deals accounting for about 41 GW of that. As of September 2, 2026.

Dollars get quoted because dollars are easy to write down. Power is the thing in short supply, and roughly four percent of announced deals will tell you how much of it they have claimed.

That asymmetry is doing real work. A $35 billion headline with no capacity figure and no term attached cannot be compared to anything, which is convenient for everyone announcing one.

## What to do with this

**Ask for three numbers, not one.** Dollars, megawatts, and years. Any compute announcement missing two of the three is a press release, not a datapoint. The 350MW and the six years tell you more about the Lambda deal than the $35 billion does.

**Track the financier, not just the buyer.** The question worth asking about any large 2026 compute deal is who is standing behind the counterparty, and whether that party also books the revenue. When the answer is the same company twice, the deal is telling you something about the vendor's confidence in organic demand.

**The question to sit with:** if AI revenue growth disappoints in 2028, which of these six-year commitments can actually be cancelled, and what happens to the provider that built a data center on the strength of one signature?