Figure commits $3.5bn to compute, roughly twice everything it has raised

Figure and the cloud provider Nscale announced on September 3, 2026 a strategic partnership to deploy up to 100,000 GPUs on NVIDIA’s Vera Rubin platform for training and inference of Figure’s humanoid robot AI. Figure is committing $3.5 billion initially, with stated plans to scale beyond $6 billion. Nscale becomes Figure’s preferred compute provider and takes a shareholding in Figure. The companies name Barstow, Texas as the site, with initial deployment in the second half of 2027.

The number to hold onto is not 100,000. It is $3.5 billion against roughly $1.9 billion — Figure’s disclosed lifetime funding across seed, Series A, Series B and a Series C of over $1 billion at a $39 billion post-money valuation announced in September 2025. A humanoid developer has just committed close to twice its entire raise to a single input, before that input exists.

What this repriced

Physical AI raised $47.4 billion across 521 deals in the first half of 2026, on Crunchbase data published August 18 — and, as we set out in our read of that half, four rounds took more than half of it. Against that backdrop a single humanoid company’s compute line at $3.5 billion sits inside the top handful of physical-AI transactions of the year, and it is not a fundraise. It is a cost.

That changes the arithmetic other humanoid programmes are judged on. Brett Adcock’s stated constraint is data and compute, not mechanics; Figure’s Helix models get better with more of both. If that framing is right, the entry price for a frontier humanoid programme is now a multi-billion-dollar training commitment on top of hardware, supply chain and field operations. Companies that raised a few hundred million to build a robot were competing on a different budget line than they thought. It also hands NVIDIA a second win in the same week: the silicon is its own, and the buyer is a portfolio company, NVIDIA having participated in Figure’s Series C.

What revenue supports it

Almost none, yet. Figure’s disclosed commercial footprint is pilot-stage: Figure 03 units began arriving at BMW’s Plant Spartanburg in late June, with no fleet size published by either party. Sacra puts Figure’s model at a robot-as-a-service subscription around $1,000 per robot per month. At that rate, $3.5 billion is not a number any plausible near-term fleet retires. As we argued in what physical AI actually earns today, industrial humanoid revenue is being underwritten on 2028 expectations, not current invoices.

The defence of the deal is that this is exactly how frontier-model economics have worked everywhere else: capacity is contracted years ahead because it is scarce, and the cost is carried by equity rather than by revenue. Nscale taking stock in Figure fits that pattern — part of the commitment is being paid in ownership, not cash, and the announced total is a ceiling on intent rather than a signed cheque cleared today. Neither company has published a payment schedule, so the cash timing is unknown. What that defence does not settle is duration: model developers contracting capacity ahead of demand have revenue curves; a humanoid fleet at pilot scale does not yet have one to extrapolate.

What to watch

Three things. Whether Figure’s next raise is sized to cover this commitment, which would tell you how much of the $3.5 billion is genuinely contracted. Whether other humanoid developers announce comparable compute deals within two quarters, which would confirm the reprice rather than mark Figure as an outlier. And whether anything at Barstow is energised on the H2 2027 schedule, because a compute commitment that slips is a commitment that quietly shrinks.