Figure Bets $6B on NVIDIA Vera Rubin GPUs to Scale Humanoid Robots

Humanoid robot startup Figure has signed one of the largest compute contracts in robotics history, betting that the bottleneck to putting a robot in every home is training compute, not hardware. The company is partnering with UK-based AI cloud provider Nscale to deploy up to 100,000 NVIDIA Vera Rubin GPUs, with initial deployment starting in the second half of 2027 at a facility in Barstow, Texas.

The financial commitment is striking for a robotics company. The deal opens at $3.5 billion and is expected to scale beyond $6 billion, with Nscale taking a strategic equity stake in Figure as part of the agreement. The two companies also plan to explore using Figure's humanoids inside Nscale's own data center supply chain.

Why compute, and why now

Figure's CEO Brett Adcock has been increasingly explicit that the company is compute-bound. Last week Figure announced Index, a data collection effort generating 35 minutes of humanoid training data every second. That firehose of teleoperation and real-world video is worthless without the GPUs to train models on it.

Helix, Figure's vision-language-action model, follows the same scaling recipe as frontier language models: more data plus more compute yields more capable behavior. Figure is applying the LLM playbook to physical intelligence, and the price tag now resembles what frontier labs spend on text models.

What Vera Rubin actually buys you

Vera Rubin is NVIDIA's successor to Blackwell, pairing the new Vera CPU with Rubin GPUs in a rack-scale system designed for agentic AI, advanced reasoning, and large mixture-of-experts inference. NVIDIA claims up to 10x lower cost per token than Blackwell and 4x fewer GPUs to train MoE models. HBM4 triples per-GPU memory bandwidth to 22 TB/s, and NVLink 6 doubles rack interconnect to 260 TB/s.

For robotics, memory bandwidth matters more than raw FLOPS. Vision-language-action models chew through image tokens, proprioceptive state, and action tokens at high frequency, and inference latency directly limits how fast a robot can react in the physical world.

Nscale's very busy year

Nscale has quietly become one of the most aggressive builders in the AI infrastructure boom. The Barstow site is leased from Bitcoin miner Ionic Digital, offering roughly 240MW with phased delivery starting in Q3 2026 and plans to scale to 1.2GW over time. Other recent deals include:

• A letter of intent with Microsoft to provide 1.35 gigawatts of AI compute capacity at the West Virginia Monarch AI campus, as a flagship Vera Rubin NVL72 deployment.
• A European deployment through Start Campus in Portugal targeting frontier AI workloads on Vera Rubin NVL72.
• Official NVIDIA Cloud Partner status for Vera Rubin, alongside CoreWeave, Lambda, and Nebius, in addition to the four hyperscalers.

Adding Figure signals that Nscale wants exposure beyond hyperscaler contracts and into the physical AI category NVIDIA has been actively evangelizing.

NVIDIA's robotics flywheel takes shape

Jensen Huang used the announcement to sketch out what he called the physical AI flywheel: train models on Vera Rubin through Nscale's cloud, validate them in Isaac Sim, and deploy them on NVIDIA GPUs embedded in the robots themselves. Every layer runs on NVIDIA silicon, which is exactly the lock-in NVIDIA has been building toward as it courts humanoid startups.

What this means in practice

1. Humanoid robotics is now a compute-scaling game. If the field still seemed gated on actuators or hands, the biggest player in the space is signaling the constraint has moved.
2. Helix is likely headed toward much larger parameter counts and longer training runs, which means the behaviors it can generalize should expand meaningfully over the next 18 to 24 months.
3. The economics of humanoids are being underwritten by a bet that data-plus-compute scaling laws hold in the physical world, something unproven at the scale demonstrated for text.

Figure is committing multiples of its last known cash position to compute that will not come online until late 2027, and scaling laws for vision-language-action models remain an active research question. If Helix scales the way Adcock is betting it will, this deal will look cheap compared to what latecomers pay for Rubin capacity.