Figure Signs $6B Deal for NVIDIA Rubin GPUs
Figure has secured a $6 billion compute deal with Nscale for 100,000 NVIDIA Vera Rubin GPUs, signaling that humanoid robotics is now bottlenecked by training scale rather than physical hardware.
Figure's agreement with UK-based Nscale starts at $3.5 billion and is projected to scale past $6 billion. Under the contract, Nscale will deploy up to 100,000 of NVIDIA's upcoming Vera Rubin GPUs at a 240-megawatt facility in Barstow, Texas, which is leased from Bitcoin miner Ionic Digital. Deliveries are scheduled to begin in the second half of 2027, with the site eventually scaling to 1.2 gigawatts. As part of the transaction, Nscale is also making a strategic equity investment in Figure, and the companies will explore deploying Figure's humanoids within Nscale's data center supply chain.
The massive compute cluster will train Helix, Figure's vision-language-action model. Figure's CEO Brett Adcock recently highlighted the company's compute-bound status following the launch of Index, a data collection initiative that generates 35 minutes of humanoid training data every second. Training Helix on this continuous stream of teleoperation and real-world video requires immense processing power. The deal mirrors the scaling strategies of frontier large language models, applying the belief that scaling data and compute will unlock advanced physical capabilities.
For AI practitioners, the transition to NVIDIA's Vera Rubin architecture represents a major leap over the Blackwell generation. Rubin pairs the new Vera CPU with Rubin GPUs, offering up to 10 times lower inference costs per token and requiring four times fewer GPUs to train mixture-of-experts models. Crucially for physical AI, where latency dictates real-world reaction times, the platform features HBM4 memory that triples bandwidth to 22 terabytes per second, alongside NVLink 6 which doubles rack interconnect speeds to 260 terabytes per second. This high memory bandwidth is vital for processing high-frequency image and action tokens.
This partnership positions Nscale alongside major cloud providers like CoreWeave and Lambda as an official NVIDIA Cloud Partner for Vera Rubin. Nscale has also secured a 1.35-gigawatt deal with Microsoft in West Virginia and a European deployment in Portugal. By backing Figure, Nscale is betting heavily on physical AI. NVIDIA CEO Jensen Huang described this ecosystem as a "physical AI flywheel" where models are trained on Rubin, validated in Isaac Sim, and run on-robot, solidifying NVIDIA's end-to-end control of the robotics pipeline.
This is our own summary of reporting by AlphaSignal
