Figure, the humanoid robotics startup that raised more than $1 billion at a $39 billion valuation in 2025, has signed a strategic partnership with AI cloud provider Nscale to deploy up to 100,000 of NVIDIA's Vera Rubin GPUs. The initial compute commitment stands at $3.5 billion, with ambitions to scale beyond $6 billion as the company races to train what it believes will be the breakthrough models in physical AI.
The announcement on September 3 positions Figure alongside OpenAI and CoreWeave as one of the largest buyers of cutting-edge AI infrastructure. OpenAI signed a letter of intent in September 2025 for 10 gigawatts of NVIDIA systems, while CoreWeave placed a $6.3 billion order the same month. What separates Figure's bet is its focus: the startup isn't building chatbots or image generators, but machines designed to walk, manipulate objects, and work in spaces built for humans.
"Our AI model, Helix, becomes more capable the same way every learned system does: with more data and compute," Brett Adcock, Figure's founder and CEO, said. That hunger for compute has intensified since the company unveiled its Index dataset program in late August. Figure claims the system now generates 35 minutes of robot training data every second—a torrent of movement patterns, object interactions, and environmental variables that the company believes will teach its robots to generalize across real-world tasks.
The infrastructure is slated to come online in the second half of 2027 at a facility in Barstow, Texas, according to the September 3, 2026 announcements from both companies. That timeline aligns with broader availability of NVIDIA's Vera Rubin platform, which entered full production in June after its unveiling at CES in January.
NVIDIA CEO Jensen Huang framed the partnership as evidence of a broader transformation. "Humanoid robots extend physical AI into the world designed for people—opening a major new industry," he said. Huang has described what he calls a "robotics flywheel": companies train models on NVIDIA Vera Rubin through cloud infrastructure, validate them in NVIDIA Isaac Sim simulation environments, and deploy them on NVIDIA GPUs embedded in the robots themselves.
At a March 2026 event, Huang told attendees that NVIDIA expects "at least" $1 trillion in revenue from its Blackwell and Rubin chips through 2027, according to Axios. Those projections hinge in part on infrastructure deals like Figure's.
A Platform Built for Scale
Nscale, a London-based AI cloud provider, is among the first to deploy the Rubin platform globally. "We're excited to partner with Figure as physical intelligence becomes AI's next frontier," Josh Payne, the company's CEO and founder, said.
The Rubin architecture represents a departure from prior generations. Each system pairs the Rubin GPU with the Vera CPU, NVLink 6 interconnect, Spectrum-6 Ethernet, and BlueField-4 DPU in a liquid-cooled configuration NVIDIA calls NVL72. Each rack holds 72 Rubin GPUs and 36 Vera CPUs, according to technical documentation. The architecture is designed for what NVIDIA calls "million-GPU AI factories"—clusters capable of training models at scales previously reserved for the largest tech companies.
Forbes noted on September 4 that the second-half 2027 timeline aligns with platform availability but raised questions about the relationship between GPU counts and dollar commitments. Multi-year infrastructure deals typically span capacity, power, cooling, and services well beyond chip purchases alone. Figure declined to disclose the initial GPU tranche size, power capacity, financing structure, or the conditions that would trigger the scale-up beyond $6 billion.
A Sector Drawing Billions in Bets
Figure's compute commitment suggests the startup believes general-purpose physical AI will demand training runs comparable to—or exceeding—those used for large language models. The company claimed on September 17 that its Helix 2.5 model achieved "zero-shot 30-home generalization" after pretraining on Index data, suggesting the robot could perform tasks in 30 different homes without additional training specific to those environments.
Whether that claim holds under scrutiny remains to be seen. But Figure is hardly alone in wagering that humanoid robotics will require both massive capital and massive compute.
Agility Robotics opened a 60,000-square-foot facility it calls a "Physical AI" plant in Fremont, California, in July 2026. The company said it has secured more than $300 million in multi-year orders for its Digit v5 humanoid, contingent on meeting milestones. Customers include logistics giant GXO, Toyota Canada, automotive supplier Schaeffler, and e-commerce platform Mercado Libre.
Apptronik closed a $520 million Series A-X round in February, bringing its total Series A to $935 million at a valuation above $5.5 billion, Bloomberg reported. Investors span venture firms, strategic corporates, and sovereign wealth: B Capital, Google, Mercedes-Benz, PEAK6, Qatar Investment Authority, AT&T Ventures, and John Deere.
1X Technologies, backed by OpenAI, opened a 58,000-square-foot factory in Hayward, California, in April. The company is targeting 10,000 home humanoids in year one and 100,000 units by the end of 2027, Bloomberg reported. Boston Dynamics, the robotics pioneer now owned by Hyundai, unveiled a productized electric Atlas humanoid for industrial applications at CES this year, with staged deployments to Hyundai factories.
NVIDIA has positioned itself as the infrastructure layer beneath all of these efforts. The company released open foundation models for humanoids through its Isaac GR00T program—versions N1, N1.6, and N1.7 across 2025 and this year—alongside Omniverse and Cosmos tools for generating synthetic training data and transferring learned behaviors from simulation to reality.
Market Projections and the Physical AI Bet
Goldman Sachs projected in 2024 that the humanoid robotics market would reach $38 billion by 2035. Secondary sources in September cited a new Goldman update lifting the 2035 forecast to approximately 6.5 million units and a $138 billion market, though those figures await primary confirmation from the firm.
If those projections prove accurate, they would represent a market roughly the size of the global semiconductor equipment industry today. Figure's willingness to commit billions in compute spending before shipping commercial robots at scale signals a belief that the path to generalization runs through data volume and model size.
The comparison to large language models is instructive. OpenAI, Anthropic, and Google have each spent hundreds of millions—perhaps billions—training their flagship models. Figure appears to be betting that teaching a robot to fold laundry, navigate stairs, and assemble components will demand similar or greater computational intensity.
What remains unclear is whether the startup's timeline and capital will prove sufficient. The gap between impressive demonstrations and reliable, economically viable robots has swallowed companies before. Figure's compute infrastructure won't be online until the second half of next year, and even then, translating training runs into robots that work consistently in messy, unstructured environments is a challenge no company has fully solved.
For now, Figure has secured its place in the infrastructure arms race. Whether that translates into robots that can actually do the work remains the harder question.
