Figure AI Commits $3.5 Billion in Compute: Signs for 100,000 NVIDIA GPUs with Nscale, Humanoid Robot Compute Supercycle Officially Begins
Introduction: A Compute Commitment That Exceeds Total Fundraising
In September 2026, a shocking announcement emerged: humanoid robot company Figure AI signed a multi-year compute agreement worth $3.5 billion with AI cloud provider Nscale, committing to access up to 100,000 GPUs based on the NVIDIA Vera Rubin platform.
The figure is staggering: the $3.5 billion commitment far exceeds Figure AI's total historical fundraising of approximately $1.9 billion as of September 2026. This is not an immediate payment but a long-term compute reservation agreement that Figure plans to fulfill as it continues to raise capital. But the commitment itself clearly signals that the humanoid robot industry is entering a new era where compute is the core competitive factor.
Deal Structure: Compute, Equity, and Strategic Alliance
Core Agreement Terms
Based on public information, the Figure AI-Nscale agreement includes the following key terms:
- Commitment amount: $3.5 billion, with intent to scale beyond $6 billion
- Compute scale: Up to 100,000 NVIDIA Vera Rubin platform GPUs
- Deployment timeline: Planned to begin in the second half of 2027 at the Cedarvale campus in Barstow, Texas
- Equity arrangement: Nscale will receive equity in Figure AI as partnership consideration
- Preferred provider: Nscale becomes Figure's preferred compute provider
NVIDIA's Strategic Role
Notably, NVIDIA is a common strategic investor in both Figure AI and Nscale. This triangular relationship ensures stability throughout the compute supply chain: NVIDIA provides hardware, Nscale provides compute infrastructure, and Figure provides robotic application scenarios and training data.
Analysts note that NVIDIA's dual investor role effectively provides an "anchoring effect" for the entire humanoid robot compute ecosystem — ensuring infrastructure is built and compute is fully utilized.
Figure AI's Compute Flywheel Strategy
Why Do Humanoid Robots Need Such Massive Compute?
Figure AI's core AI system is Helix — a vision-language-action (VLA) model responsible for controlling robot perception, decision-making, and action. Helix's capability improvements directly depend on large-scale training data and compute.
Figure's compute needs come from three levels:
- Model training: Training more powerful Helix versions requires substantial GPU compute
- Data processing: Processing massive video and sensor data from robot field deployments
- Simulation training: Large-scale robot behavior simulation in virtual environments
Index Initiative: 16 Million Video Data Flywheel
In August 2026, Figure launched the Index platform — a system for capturing human work data via sensors, expected to cost $1 billion. By the time of the agreement signing, Index had collected over 16 million videos documenting human operations in various industrial scenarios.
This data will become the core material for training Helix models, forming a self-reinforcing flywheel:
- More robot deployments → More field data → Stronger Helix models → More robot deployments
The $3.5 billion compute commitment is designed to support this flywheel's high-speed operation.
Figure AI's Commercial Progress
Production Scaling: One Robot Per Hour
In early 2026, Figure AI completed the critical transition from prototype development to high-volume production. At its BotQ factory, Figure 03 production reached one robot per hour, achieving a 24x throughput increase within 120 days. By June 2026, the company's operating robot count (approximately 740) exceeded its human employee count.
Commercial Partnerships: BMW and Catalyst Brands
BMW partnership: In June 2026, BMW announced that Figure 03 robots would begin logistics and assembly work in Hall 52 of its Spartanburg plant. This is an important commercial milestone for humanoid robots in mainstream automotive manufacturing.
Catalyst Brands partnership: Figure signed an agreement with Catalyst Brands to scale humanoid robot deployment in its supply chain facilities in Reno, Nevada, for warehousing and logistics operations.
Market Context: GPU Compute Supply-Demand Tension
Structural Shortage in the Compute Market
In 2026, the market for high-end AI accelerators is severely imbalanced. Delivery times for data center accelerators like NVIDIA H100 and B200 have extended to 36-52 weeks, while broad availability of the next-generation Vera Rubin platform is not expected until 2027.
In this context, Figure's strategy of locking in 100,000 Vera Rubin GPUs in advance is self-evident — this is a "compute stockpiling" strategy in an era of compute scarcity, ensuring the company obtains sufficient training compute before competitors.
Humanoid Robot Industry Business Model Bifurcation
Figure's compute investment reflects a deep business model bifurcation occurring in the humanoid robot industry:
Hardware manufacturing model: Represented by traditional industrial robot manufacturers, with core competitiveness in mechanical design, manufacturing processes, and supply chain management — relatively stable margins but limited growth.
Intelligent model model: Represented by Figure AI, with core competitiveness in AI model (Helix) capabilities. Compute investment is the core cost, but once model capabilities break through, they can be rapidly scaled to large numbers of robots with decreasing marginal costs.
Figure has clearly chosen the latter — betting on exponential improvement in AI model capabilities through large-scale compute investment, ultimately achieving large-scale commercialization of humanoid robots.
Asia-Pacific Humanoid Robot Investment Opportunities
Figure AI's compute commitment also provides important market signals for investors and enterprises in the Asia-Pacific region:
Manufacturing applications: Manufacturing giants in Japan, South Korea, and China are actively evaluating the application potential of humanoid robots. Companies like Toyota, Hyundai, and Foxconn have begun negotiating partnerships with multiple humanoid robot companies.
Compute infrastructure: Data center operators in the Asia-Pacific region are accelerating expansion to meet compute demands for humanoid robot training. Japan's data center market is expected to nearly double to over $32 billion by 2028.
Local competitors: Chinese companies like Unitree and Fourier Intelligence are rapidly catching up, with some products already competitive in cost and performance.
Conclusion: Compute Is Competitiveness
Figure AI's $3.5 billion compute commitment is a landmark event marking the humanoid robot industry's entry into a "compute supercycle." In this new era, compute is no longer just a core resource for AI labs but has become a core competitive factor for robot companies.
Whoever can lock in sufficient training compute in an era of compute scarcity will maintain a lead in the race for AI model capabilities. Figure AI's strategic positioning may prove to be an important watershed in the competitive landscape of the humanoid robot industry.
