i
DATAIST
News · 2026-09-05

Figure promises Nscale $3.5B after raising $1.9B in total

@neuronium_ai @neuronium_ai

Figure has committed to buy $3.5 billion of AI compute from Nscale, which in the same deal becomes the humanoid robot maker's preferred compute supplier and takes an equity stake in it on undisclosed terms. The commitment can expand to $6 billion, covers 100,000 Nvidia GPUs at a site in Texas, and starts delivering in the second half of 2027. Figure has raised $1.9 billion in its entire existence and has never disclosed revenue.

Cover: Figure promises Nscale $3.5B after raising $1.9B in total

Figure has committed to buy $3.5 billion of AI compute from Nscale, which in the same deal becomes the humanoid robot maker's preferred compute supplier and takes an equity stake in it on undisclosed terms. The commitment can expand to $6 billion, covers 100,000 Nvidia GPUs at a site in Texas, and starts delivering in the second half of 2027. Figure has raised $1.9 billion in its entire existence and has never disclosed revenue.

The compute is earmarked for Helix, Figure's physical AI model, which runs the Figure 03 humanoid and the coming Figure 04. Brett Adcock, Figure's founder and chief executive, has named data and compute as the company's two binding constraints. For the first of them Figure built Index, an app through which people perform household chores to generate training data: 264,000 downloads, 44,000 weekly active users, 16 million videos from 108 countries, and 35 minutes of uploaded video ingested every second.

The $1.9 billion raised so far breaks down as seed money from Adcock himself, a $70 million Series A, a $675 million Series B, and more than $1 billion raised last September at a $39 billion valuation. The new obligation is close to twice all of it, entered into by a company that has reportedly built just over 1,000 Figure 03 robots. By the account of the analysis this is drawn from, no money has changed hands yet, and none is scheduled to before the first GPUs arrive in the second half of 2027.

The Texas site is not a new campus. It is Cedarvale, in Ward County, owned by Ionic Digital, a former bitcoin mining operator. Nscale leased the entire facility last October on a ten-year triple-net lease worth roughly $2 billion. It was originally lined up for Microsoft, which planned to put about 104,000 Nvidia GB300 GPUs there. Cedarvale's current capacity is estimated at 234–240 MW, spread across four 50 MW buildings first built to house some 66,000 bitcoin mining rigs.

Those numbers do not accommodate what Figure announced. A fleet of 100,000 Vera Rubin GPUs works out to roughly 1,389 NVL72 racks. At 190–230 kW per rack, the compute alone draws 264–320 MW; with cooling and infrastructure losses, about 340–420 MW. That exceeds the whole site's present capacity, before Microsoft's own claim on it is considered. Microsoft holds the right to an additional 700 MW Cedarvale phase that could come online from late 2027, and the full campus could in theory reach 1.2 GW — but that right belongs to Microsoft, not Figure. Adding power is also harder than it was a year ago: Texas paused review of data centre interconnection applications in August, with 474 GW of requests sitting in the ERCOT queue, about 90% of them data centres according to Governor Greg Abbott.

The hardware arithmetic is no kinder. Morgan Stanley values a single Vera Rubin NVL72 rack at roughly $7.8 million, with other estimates reaching $8.8 million. For 1,389 racks that is $10.8–12.2 billion in equipment alone, before land, buildings, electrical plant or Nscale's margin. The reading that fits the money is that $3.5 billion is rent across the life of the contract rather than a purchase — which corresponds to something like 32,000–45,000 GPUs. Even the $6 billion ceiling buys about half the fleet in the headline.

Timing explains more than physics does. Nscale is preparing a US listing, reportedly aiming to raise up to $3 billion at a valuation near $25 billion, though no public filing has been made. Its contracted revenue backlog went from about $51 billion in early August to $103 billion after the Anthropic deal. Its recognised revenue was about $33 million for all of 2025, about $37 million in the first quarter of 2026, and a little over $100 million in the second. A person familiar with the discussions called the $103 billion figure illustrative rather than an official revenue forecast. Figure's $3.5 billion joined that pile days before an expected pre-IPO investor presentation.

The shape of the relationships is by now familiar. Nvidia took part in Nscale's $1.1 billion Series B. Nscale buys Nvidia GPUs, invests in Figure, and becomes Figure's preferred supplier. Figure, which counts Nvidia among its investors, commits to buy Nscale capacity. Figure's obligation becomes Nscale's backlog, the backlog supports the listing and the debt financing, and the proceeds buy more Nvidia GPUs. Jensen Huang calls this a physical AI flywheel. The short seller Jim Chanos has criticised financing structures of this kind; Huang has called the description of such deals as circular financing absurd.

My own read is that this is a roadshow number more than a procurement decision. Every quantity in it — the GPU count, the site, the delivery date — falls apart on contact with the constraint next to it, while the one quantity that holds up is the one printed in the press release and counted in somebody else's backlog. A pre-revenue robotics company and a pre-revenue cloud provider exchanged large figures shortly before one of them goes public. That is not proof of anything improper, but it is the most economical explanation of why the announcement reads the way it does.

What would make it real is robots earning money. To pay $3.5 billion out of operations rather than out of the next round, Figure would need a revenue-generating fleet in the tens of thousands deployed in roughly two years. Its most mature deployment is the BMW plant in Spartanburg, where a Figure 02 logged about 1,250 hours over some 11 months and moved more than 90,000 sheet metal parts. BotQ, Figure's factory, is designed for 12,000 robots a year; Adcock has spoken of 100,000 robots in four years and describes Figure 04 as the version built for manufacturing scale. The gap between one robot's 1,250 hours and a fleet large enough to service this commitment is the entire business.

None of this makes the plan impossible. Adcock has an entrepreneurial record, and in May he closed a separate $700 million round for another AI hardware company at a $6 billion valuation. But three questions put to Figure and Nscale — whether the $3.5 billion agreement is legally binding, how large Nscale's stake in Figure is, and whether 100,000 Figure GPUs can physically sit at Cedarvale alongside Microsoft's 104,000 — decide whether there is a deal here or an entry in a backlog. Until they are answered, the most concrete thing the commitment has produced is a bigger number on Nscale's IPO slide.