Figure locks in 100k NVIDIA GPUs in a $3.5B physical-AI bet

100,000 GPUs, $3.5B, and a Texas address

On September 3, Figure announced a strategic partnership with cloud provider Nscale to lock in large-scale compute for training its Helix humanoid models. By the two companies' account, Figure can draw on up to 100,000 NVIDIA Vera Rubin GPUs, with an initial commitment of $3.5 billion of compute and intent to scale beyond $6 billion. For a humanoid company, that is less a purchase than a long-term contract signed years before the capacity is needed.

The time and place are just as concrete: initial deployment is targeted to start in the second half of 2027 in Barstow, Texas, with later expansion toward the intended scale. The compute will not arrive immediately — what is being locked in now is supply capability several years out. It helps to split this into three layers: an intent (beyond $6 billion), an initial commitment ($3.5 billion) and a start date (in the second half of 2027). What the announcement fixes is those sizes and that timing; whether the power, cooling and buildings exist to run the chips can only be verified during deployment. With compute persistently tight, getting it at all, and at a workable price, often matters more than how much one wants.

The deal runs in more than one direction. Beyond buying compute, Nscale is making a strategic investment in Figure, and the two will also explore scaling Nscale's own supply chain with humanoids. That clause is easy to overlook: the humanoid maker's customer list now includes a compute company that intends to put robots to work.

Four ends of the flywheel: data, models, compute, bodies

Figure describes its own position bluntly: the company has entered a phase where it is largely bound by the data and compute needed to train Helix. The Index dataset announced the week before is generating roughly 35 minutes of new human-behavior data every second, but data alone cannot solve the problem — scaling physical intelligence demands an immense amount of compute. In Figure's words, the partnership supplies the compute runway needed for the next generation of models.

That runway serves a flywheel already taking shape: the Index human-behavior dataset, Helix training and F.03 volume production each fuel the others. All three accelerated at once — on September 30, Figure announced it was retiring the F.02 fleet to align with F.03 production, noting that continuing to maintain F.02 would slow F.04 development. Retiring one generation to free capacity for production and the next generation is exactly what turning a flywheel looks like.

From an engineering standpoint, the 100,000-GPU scale is about far more than chips. Training a foundation model means dense gradient synchronization across thousands of accelerators, so interconnect bandwidth, power delivery, cooling, scheduling and failure recovery all matter — a weak link anywhere burns expensive compute for nothing. Building such a cluster means enormous capital spending and a long construction cycle, while renting from a cloud provider packages all of it as a purchasable service, which is precisely Nscale's role here.

NVIDIA founder and CEO Jensen Huang described the chain in three steps: training Figure's models on Vera Rubin through Nscale's AI cloud, validating them in NVIDIA Isaac Sim, and deploying them on NVIDIA GPUs inside Figure's robots — the robotics flywheel, activated. Simulation and real hardware, training and inference, all folded into one hardware story: a deal shape NVIDIA is happy to see.

Figure founder and CEO Brett Adcock kept it short: Helix becomes more capable the same way every learned system does, with more data and compute, and the partnership is how Figure gets what that requires. Nscale founder and CEO Josh Payne called physical intelligence AI's next frontier, pointing to the growth his company had already seen in inference and agentic AI and crediting Figure with pushing the field further.

Why humanoid makers lock in compute early

To understand the money, understand how pretraining is spent. In large language models, data volume, parameter scale and compute investment have long shown a broadly verified relationship with final capability; and before a team finds the right architecture, data mix and training recipe, it typically has to run large-scale experiments again and again, each costing tens or hundreds of millions of dollars. For a company whose main business is physical intelligence, certainty of compute supply is close to certainty of research cadence.

That explains why Figure is putting $3.5 billion into training before robots ship at scale: the ceiling on robot capability depends on how much human behavior the model has seen and how many training runs it has been through. It also explains why the contract sits in the company's narrative alongside the Index dataset and F.03 production — it is the fuel segment of the flywheel.

For the humanoid industry the signal is blunt: competition is shifting from who can build a body to who can stack up training compute and data at scale. For the past two years the debate centered on joints, dexterous hands, runtime and yield; today the most expensive contract in the field buys GPU time. As bodies increasingly become procurable parts, training scale becomes the new dividing line — data can be collected, models trained and joints sourced, but accelerator supply at the ten-thousand-to-hundred-thousand scale is finite and booked years ahead, so signing that scale first means laying out a runway before the next round of competition begins.

What the contract does not yet mean

The boundaries deserve stating plainly. Initial deployment does not begin until the second half of 2027, the $6 billion-plus figure is intent that must be scaled into gradually, and intent is not the same as money spent. Scaling Nscale's supply chain with humanoids is, for now, still framed as exploration. And for Figure itself, the promise of robots in homes remains separated from this compute contract by hardware yield, cost, model reliability and actual deliveries.

Training compute is also only one necessary condition. Work like Helix 2.5 shows that broad pretraining on the data side is an equally decisive variable, and what a robot can do depends on how fast real-world data is collected and how well the model uses it. The contract fills the fuel tank; whether the car reaches the finish depends on the whole chain.

None of that makes the shape of the deal any less notable: it is not an order for robots but a capital commitment signed years ahead. Competition in humanoids is turning into a race over data scale and compute reserves — whoever locks both down first gets to talk about the next generation of models.