NVIDIA Eyes $1B Figure AI Stake at $38B Valuation

NVIDIA Deepens Physical AI Bet Through Figure AI

NVIDIA is reportedly in advanced discussions to commit an additional $1 billion to Figure AI, valuing the humanoid robotics company at approximately $38 billion pre-money. The Information first reported the talks on October 7, 2026, citing a person familiar with the matter. This potential round would follow NVIDIA's participation in Figure's Series C, which closed in September 2025 with more than $1 billion committed at a $39 billion post-money valuation. Neither company has confirmed the new discussions, and terms remain fluid.

The timing aligns with Figure's aggressive scaling of training infrastructure. In September 2026, Figure signed a multi-year agreement with Nscale for an initial $3.5 billion in AI cloud capacity using NVIDIA Vera Rubin GPUs, with plans to scale beyond $6 billion. The deployment targets up to 100,000 GPUs starting in the second half of 2027 at a site in Barstow, Texas. NVIDIA CEO Jensen Huang described the arrangement as a "physical AI flywheel" linking model training on Vera Rubin hardware, simulation in Isaac Sim, and onboard inference in Figure robots.

Figure's Helix foundation model relies on large-scale vision-language-action training to enable general-purpose manipulation. The additional NVIDIA capital would directly support continued expansion of this compute pipeline, which already underpins zero-shot generalization across factory and warehouse tasks.

Compute Commitments Drive VLA Model Scaling

Figure's $3.5 billion Nscale commitment represents one of the largest single-robotics compute deals disclosed to date. The infrastructure will train successive generations of Helix, Figure's end-to-end neural controller that maps vision and language inputs directly to joint torques and actions. Vera Rubin GPUs deliver up to 50 petaflops of NVFP4 inference performance per device, a fourfold training efficiency gain over Blackwell according to NVIDIA's January 2026 platform launch materials.

This scale matters for reinforcement learning loops in manipulation. High-fidelity simulation data from Isaac Sim must be paired with real-world rollouts to refine policies for contact-rich tasks such as connector insertion or deformable object handling. The Barstow cluster size—potentially 100,000 GPUs—allows parallel sampling of millions of trajectories per day, a requirement for stable policy improvement in high-dimensional action spaces.

NVIDIA's equity position in both the chip supplier and the end-user robotics company creates a vertically integrated feedback loop. Capital from NVIDIA accelerates Figure's data collection and model iteration, which in turn drives demand for more Vera Rubin capacity. Critics have labeled the structure a financing triangle, yet it mirrors NVIDIA's broader physical AI strategy of owning layers from silicon through cloud orchestration to embodied deployment.

Physical Intelligence Pi API Beta Expands Access to VLA Fine-Tuning

Parallel developments at Physical Intelligence highlight how the same class of models is moving from closed research to hosted services. On October 13, 2026, Physical Intelligence opened a beta of its Pi API, allowing external developers to upload robot demonstration data, fine-tune π0 and π0.5 vision-language-action models on managed infrastructure, and run inference through a hosted gateway.

The π0 architecture is a flow-based VLA transformer; π0.5 adds knowledge insulation training for improved open-world generalization. Early partners include Almond for dual-arm metal finishing and General Instinct for heavy-equipment operation. The service reduces the barrier for teams lacking their own multi-thousand-GPU clusters, letting them iterate policies in days rather than months.

This hosted approach complements the capital-intensive path taken by Figure. Where Figure secures dedicated racks for proprietary Helix iterations, the Pi API democratizes adaptation of open foundation models. Both trends point toward a future in which neural controllers are trained centrally and deployed with lightweight on-robot inference heads.

Valuation Context and Market Implications

At the reported $38 billion pre-money level, Figure would command a valuation exceeding Goldman Sachs' 2035 global humanoid market projection of $38 billion. The September 2025 Series C already lifted the company from a $2.6 billion Series B mark in February 2024 to $39 billion post-money. Secondary market marks as of mid-October 2026 show some softening to around $35.5 billion, reflecting typical post-round digestion rather than operational setbacks.

For investors tracking NVIDIA Figure AI investment activity, the follow-on round signals continued confidence in the humanoid thesis despite elevated pricing. The $1 billion check size would roughly double NVIDIA's prior commitment and reinforce its role as both technology provider and financial backer.

Technical Implications for Neural Controllers

Larger training budgets directly affect torque mapping precision and sample efficiency. Helix-style models must learn low-level impedance control alongside high-level task planning. Additional Vera Rubin capacity enables longer-horizon reinforcement learning with denser reward signals derived from force-torque sensors and tactile arrays on Figure 02 hands.

The Pi API beta similarly targets data-efficient fine-tuning. Physical Intelligence reports that partners achieved target behaviors with 5 percent of the adaptation data previously required, using the hosted pipeline. This efficiency gain matters for serviceability: field teams can upload short demonstration sets from a single robot and receive an updated policy without retraining the entire foundation model.

Unresolved questions remain around sim-to-real gap closure at 100,000-GPU scale and the long-term economics of hosted inference versus edge deployment. Figure's BMW pilot and ongoing deployments will provide real-world telemetry on policy robustness, whi