Figure signed a $3.5B compute agreement with Nscale for 100,000 GPUs
The deployment is targeted for the second half of 2027. It is a two-year forward bet that humanoid robots will have enough training data to keep that hardware busy.
Figure AI has signed a $3.5 billion agreement with compute provider Nscale targeting 100,000 Nvidia GPUs, with deployment planned for the second half of 2027, according to a statement issued on 3 September.
It is the largest compute commitment made by a robotics company, and it is a forward purchase: hardware that does not yet exist, for training runs not yet designed, on a timescale that ends two years from now.
What the commitment implies
Robotics has not historically been compute-constrained. The binding constraint has been data — real-world interaction is slow to collect, expensive to gather, and does not benefit from the internet-scale corpora that made language models work.
A $3.5 billion compute commitment is therefore a statement that Figure believes the data constraint is lifting. The most likely reason is simulation: if enough training experience can be generated synthetically, then compute becomes the limit again, and buying two years ahead of a shortage is rational.
The surrounding evidence supports the direction. Tripo AI raised about $446 million for generative 3D with simulation demand behind it. Lucida, published on 1 September, rebuilds cluttered indoor scenes from video into editable assets specifically for robot simulators. NavMCP reported 78.3 percent success on a Unitree Go2 by pairing a vision-language model with a navigation foundation model.
The transfer problem nobody has solved
Simulation-to-reality transfer is the oldest unsolved problem in the field. Policies trained in simulation reliably work in simulation. Contact dynamics, deformable materials, friction and the accumulated small differences between a rendered world and a physical one degrade them on real hardware, and the gap has closed slowly rather than suddenly.
If synthetic data transfers well enough at scale, 100,000 accelerators is a sensible purchase. If it does not, it is $3.5 billion of capacity producing policies that work beautifully in a simulator.
Figure has not published transfer results that would let an outsider judge which.
The counterparty risk runs both ways
Nscale is simultaneously raising $3.5 billion in pre-IPO financing, about $2 billion of it from Nvidia, and holds a reported $45 billion contract with Anthropic.
So a provider that has not yet listed, whose capacity is being financed by its own supplier, has sold a two-year forward commitment to a robotics company whose product is not shipping at volume. Each party's obligations depend on the other's forecast being right.
That pattern is now general. Nvidia is investing in the providers, leasing the buildings, supplying the chips and booking the revenue, and Jensen Huang's forecast of roughly 70 percent growth for fiscal 2028 rests partly on commitments from customers Nvidia has capitalised. The Figure-Nscale deal is one more node in that structure, and the first large one where the end demand is robots rather than tokens.
Neither company disclosed payment schedules, minimum commitments or what happens to the agreement if either party's timeline slips.
Runs the newsroom. Rename this profile in the studio to your own byline.
Related
Every weekday, the AI stories that moved money or shipped code.
No cross-posting, unsubscribe anytime. See all newsletters