Figure sizes its humanoid training run at up to 1,389 Vera Rubin racks

Trace what the agreement actually wires together. Helix trains in Nscale's cloud, gets validated in NVIDIA Isaac Sim, then runs on NVIDIA hardware inside the robot ^[5]. Figure says its binding constraint is the data and compute needed to train Helix, the vision-language-action system that drives its machines ^[4].

Then the rack math. Vera Rubin bundles Rubin GPUs with Vera CPUs, networking, storage processing and rack-scale systems ^[6], and NVIDIA puts 72 Rubin GPUs and 36 Vera CPUs in an NVL72 rack ^[7]. Divide the ceiling by the rack and you get about 1,389 racks ^[1]. Treat that as an upper bound on one robotics company's rack-scale demand and nothing more, because neither release says how many GPUs land in the initial phase or whether every system uses that configuration ^[8].

The per-GPU numbers are easy to compute and easy to misread. An initial $3.5 billion across 100,000 GPUs is $35,000 each ^[2]; the intended expansion past $6 billion is $60,000 each ^[3]. For either figure to describe a price, the money would have to cover the whole ceiling, and the covered share is not stated ^[13]. What the commitment does anchor is scale against Figure's own budget: roughly 3.5 times the more than $1 billion it said in August it would spend on data and compute over the following twelve months ^[10][7].

The data side is measurable in a way the compute side is not. Index has paid contributors $15 million against 16 million uploaded videos ^[9], about 94 cents a video ^[6]. At 35 minutes of uploaded video per second, ingest runs near 2,100 times real time ^[5]. Ninety-four cents a clip is cheap training data, assuming the clips show what the label says they show.

Power is where the public record thins out. Nscale said on August 31 it had closed approximately $3 billion in delayed draw term loan commitments for campuses in Ward County, Texas and Madison, North Carolina, allocating up to $1.85 billion to Ward County ^[14], roughly 62 percent of the total ^[8]. An Ionic Digital supplemental prospectus puts Nscale's Ward County lease at 234 megawatts with another 89 megawatts contractually obligated if that capacity becomes available ^[15], so up to 323 megawatts ^[4]. Nscale's financing release does not identify that campus as Figure's deployment site ^[16].

Which leaves the schedule carrying the whole argument. Runtimewire reads the GPU count as a ceiling and the installation as at least several quarters away ^[19]. For the rack arithmetic above to transfer into anyone else's capacity planning, the initial money would need to buy a stated fraction of the 100,000, the systems would need to ship as NVL72 racks, and the site would need interconnect for them. None of the three is on the record yet, which is the honest reason to hold the parties to the delivery date rather than the headline count.