A week after we covered Nscale's Anthropic deal, Nscale quietly promised a large, multi-year pool of compute to Figure AI so it can train and run humanoid robots.
The agreement commits at least $3.5 billion in compute, with an intention to scale beyond $6 billion over time. Initial hardware deliveries are slated to begin in the second half of 2027 at Nscale’s Barstow, Texas data center. The plan centers on NVIDIA's Vera Rubin GPUs, and industry reports put as many as 100,000 GPUs in scope. Nscale will take an equity stake in Figure and become its preferred compute provider.
Why this matters: compute has been the choke point for embodied AI. Booking multi-billion-dollar capacity ahead of time turns compute from a scarce, rented resource into something Figure can plan around, which shortens the cycle between training a model and trying it on a real robot.
How it works, simply: big GPUs do the heavy lifting to train Helix models, simulation software validates robot behaviors in virtual worlds, and the same GPU architecture runs the models inside the robots. NVIDIA calls that loop a "robotics flywheel", train, test, deploy on the same family of hardware.
What changes now is practical and strategic. Figure gains planning certainty and a path to scale, and Nscale positions itself as a go-to provider in a more competitive market. This is not immediate access for everyone: reporters note the deal is a capacity commitment with no GPUs or cash exchanged yet, and deliveries begin in 2027, so it's a forward-looking industrial bet, not an overnight shift.
The open question is whether that reserved fleet will actually turn prototypes into reliable, usable robots outside the lab. The next year of hardware deliveries and early deployments will show whether the "robotics flywheel" really spins into products.