Nscale bought into the robot maker that just committed $3.5bn to its compute

Somewhere in Nscale's data room there is a line that reads Figure, $3.5bn. Anyone pricing it needs two numbers the announcement does not carry: the length of the term, and the size of the cheque Nscale wrote back the other way.

Start with the disclosed figures. Spread $3.5bn across up to 100,000 Vera Rubin GPUs ^[1][2] and you get $35,000 per GPU for the whole multi-year term, or $60,000 at the $6bn ambition ^[19]. Without a term length, neither number tells you what an hour of Barstow capacity costs. Against Nscale's stated backlog of around $51bn ^[17], the initial commitment is about 6.9 percent, rising to roughly 11.8 percent if the larger version signs ^[20]. The rest of that backlog may be plain arm's-length business; this portion is not, because the buyer is now part-owned by the seller ^[3].

The direction of the money is what makes it hard to read. Nscale sends Figure an undisclosed sum and takes shares, and Figure commits at least $3.5bn back over a period nobody has specified ^[1][3]. Bloomberg's Lynn Doan described this as the playbook Nvidia has run for two years, and noted that some on Wall Street treat circular arrangements as a way of making demand look larger than it is ^[9]. Nvidia's own version of it is a $99bn equity portfolio built largely on positions in the labs and clouds that buy its chips ^[10]. Here Nvidia has money in Nscale and money in Figure ^[11], which puts it on both sides of a contract for its own hardware.

The pitch is easy to separate from the work. The pitch is Jensen Huang's robotics flywheel: Figure's models train on Vera Rubin through Nscale's cloud, get validated in Nvidia's Isaac Sim, then run on Nvidia GPUs inside the robots ^[4]. The work is duller and more real. Brett Adcock has said Figure's constraint is data and compute rather than money ^[13], and the money part checks out, given a round above $1bn at a $39bn valuation in September 2025 ^[15]. The data part is already running: Figure's Index programme ingests thirty minutes of human training video every second from paid creators using a phone app ^[14], which is 1,800 times real time, or about 43,200 hours of footage a day ^[21]. Figure says the sustained runs needed to turn that footage into a robot that can work a full shift were beyond the scale it could reach ^[18]. That is a genuine constraint, and Barstow capacity is a genuine answer to it, though the first systems are not due until the second half of 2027 ^[6].

One line in the release has no number and no date attached: the two companies say they will explore scaling Nscale's supply chain with humanoids ^[12]. Read plainly, that is Figure robots working shifts in the buildings that train Figure robots. Barstow is a 234MW site Nscale leases from Bitcoin miner Ionic Digital, where it is also deploying infrastructure for Microsoft ^[8], so the staffing question in that building is real even if the humanoid answer stays an ambition.

The forcing function, for anyone handed a vendor backlog before a listing: sort each large contract on two axes. Does the seller hold equity in the buyer, and does delivery start before the seller's next capital event? The Figure line answers yes to the first and no to the second ^[3][6][5]. Contracts in that box cannot be netted out by anyone outside the deal, which is why the undisclosed investment matters more to a stranger than the $3.5bn does. Until Nscale puts a figure on its stake, the accurate way to read that backlog line is as a commitment from a company Nscale part-owns, priced in a currency only the two signatories can see.