Nscale has agreed to sell Figure at least $3.5 billion of computing power — and it has bought a stake in Figure at the same time. [1] Figure announced the strategic partnership on September 3: up to 100,000 Nvidia Vera Rubin GPUs, deployment targeted from the second half of 2027 in Barstow, Texas, an initial $3.5 billion compute commitment with stated intent to scale past $6 billion. [3] The Next Web's Sunday reading names what that structure is: the playbook Nvidia has run for two years, one layer down — the vendor finances the customer, then owns a piece of him. [1] Nvidia, for the record, is an investor in both companies. [1]

Start with the physics, because the numbers need it. Training a robot to fold laundry, stack groceries, or walk a factory floor for eight hours takes a strange amount of computing power — more, in some cases, than training a large language model. [2] That is the demand story behind the deal: Figure, chasing a $39 billion valuation in the humanoid race, signed a multi-year agreement for at least $3.5 billion of AI cloud capacity with room to grow past $6 billion. [2] Compute is the robot's calorie count. Somebody has to sell the calories.

Now the structure, because the structure is the story. Nscale is the seller of the calories and a new shareholder in the eater. [1] The stake's size and price are undisclosed. [1] The $6 billion is intent, not commitment. [2][3] Barstow is a 2027 target, not a poured foundation. [3] Every hard number in the announcement describes the future; every present-tense fact describes a financing relationship. Until Texas has racks, the GPUs are paper and the equity is a bet.

The bet's shape is familiar. Nvidia spent two years pairing chip supply with checks into its own customers — lessor, backer, and hardware supplier in a single relationship — and the market is still arguing over where enablement ends and circularity begins. [1] Nscale runs that play one layer down: infrastructure vendor plus shareholder, with Nvidia itself sitting on both sides of the table as an investor in each company. [1] When the seller's upside depends on the buyer's valuation holding, the press release's capacity figures and the cap table are the same story told twice.

None of this makes the deal fake. A multi-year compute agreement at this scale, if delivered, is real industrial capacity for a real robotics program, and Vera Rubin silicon allocated to Barstow would be a genuine deployment. [2][3] But delivered is doing all the work in that sentence. Announcements of capacity arrive in September; GPUs arrive, if they arrive, in the second half of 2027. [3] Financing announcement and delivered capacity have been separate stages in every AI infrastructure story this paper has tracked, and this one gives no reason to merge them early.

The two chief executives posed with the hardware — Payne and Adcock flanking a Figure 03 robot in Nscale livery — which is what announcements look like when both sides need the photograph more than the electrons. [1] The London infrastructure firm needs the demand signal; the humanoid company needs the capacity story for the $39 billion conversation. [1][2] Everybody in the picture needs everybody else in the picture. That is not fraud. It is the geometry of vendor finance: the photograph is the product until the data center is.

Watch three things. The stake terms, if they ever surface — price and percentage turn a vibe into a valuation opinion. The Barstow build — permits, power, construction photos, the boring receipts. And Figure's next raise — whether $39 billion survives contact with a market that has now seen the vendor holding equity. [2] Until then, file this under vendor finance with excellent robots: $3.5 billion of compute sold, one customer part-owned, and the GPU industry's favorite playbook running one layer down. running one layer down. The alternative reading — that this is simply how frontier hardware gets financed now, with risk shared across the stack — deserves its sentence too: nobody builds a hundred thousand GPUs of capacity on purchase orders alone. [3] Shared risk or circular finance, the test is identical and eighteen months out. Barstow will grade them both.

-- THEO KAPLAN, San Francisco