Figure AI Locked In $3.5 Billion of Compute Before It Has a Product to Sell

At IROS 2026 in Pittsburgh, which ran from September 27 to October 2, one announcement stood out from the usual run of research demos. Figure AI said it had signed a multi-year partnership with the AI cloud provider Nscale, committing an initial $3.5 billion to compute, with plans to scale past $6 billion and to deploy as many as 100,000 Nvidia GPUs starting in the second half of 2027.

The build is planned for Barstow, Texas. Training would run on Nvidia's Vera Rubin platform, validation would happen in Isaac Sim, and the same GPUs would eventually run inside the robots themselves.

A bet placed before the proof

Humanoid companies have raised a lot of money, and most of them have spent it on hardware and pilots. Committing this much to compute before a robot is doing paid work in large numbers is a different kind of bet. It says the company believes the binding constraint on humanoids will be training compute and model quality, not mechanical parts.

That belief is testable, and the evidence so far is mixed. The most-watched humanoid programme in the world, Tesla's Optimus, raised output at its Fremont line to several hundred units a week in late September, roughly ten times its second-quarter rate, and suppliers were reportedly asked for parts for about 15,000 units across 2026. The reported blocker is not shells per week. It is hands and the inability of the AI to generalise across tasks. A factory that can build the body has not solved the problem of making the body useful.

Figure's compute commitment is, in effect, a bet that throwing much more training capacity at that generalisation problem is the right move. If that is true, the company that secures the compute first has an advantage. If it is not, the money buys a large bill and a lot of idle GPUs.

Why the 100,000 GPUs number is the story

Deploying GPUs inside robots is the part worth watching. Training compute is a known quantity. On-board compute that runs a large model in real time, within a power and thermal budget that a moving machine can carry, is a harder engineering target. If Figure is planning to put Nvidia GPUs into the robots at scale, it is signalling that the models it wants to run are large enough that edge accelerators will not do.

That has consequences well beyond Figure. It sets a requirement on power, cooling and cost per unit that changes what a humanoid is worth to a buyer. A robot that needs a serious GPU to operate is a different product from one running a small model on a phone-class chip.

Where the rest of the field is

The IROS floor gave a useful comparison set. Agility Robotics unveiled Digit 5, the next version of its bipedal robot, and cited more than 65,000 operational hours for Digit 4 across North American customer sites including Amazon, GXO, Schaeffler and Toyota Motor Manufacturing Canada. Agility also announced a partnership with FORT Robotics aimed at functional safety for humanoids. Those two things, hours in service and a safety case, are what a factory buyer asks for before signing.

Sanctuary AI reported a success rate above 99.5 percent at a 2.54 second cycle time on a wire-plugging task, validated against a tier-one auto supplier's live benchmarks. That is a narrow task, and narrow tasks are exactly where the technology is ready. Unitree announced H2 Plus as an Isaac GR00T reference platform for academia, extending a fast release cadence. Astribot used the conference to debut its T1 humanoid in North America at roughly $18,000, positioned as a research platform rather than a worker.

Read together, the field is splitting. Some companies are racing to prove general capability and spending heavily on compute. Others are shipping narrow, measurable tasks and building the safety and service records that customers actually pay for.

What the conference as a whole signalled

IROS is a research gathering, and the technical tracks reflected where the attention is. The workshops focused on agile loco-manipulation, on evaluating policies for vision-language-action controllers in the real world, and on getting simulated training to transfer onto physical bipedal machines. That last topic is the one that decides whether compute spent in simulation turns into behaviour on a factory floor.

The host city was Pittsburgh, which put Carnegie Mellon's robotics bench strength on display. For a field that has spent years producing impressive videos of robots doing parkour, the shift in emphasis toward contact-rich manipulation and real-world evaluation is a meaningful change of direction. It is easier to film a robot jumping than to certify one that picks up an unfamiliar object safely.

The compute argument, stated plainly

Figure's bet rests on a specific claim: that the gap between humanoids that demo well and humanoids that work is mostly a modelling problem, and modelling problems are solved with data and compute. That claim is not crazy. The last few years showed that scaling up training on larger and more varied data keeps improving generalisation in other domains, and there is no strong reason to think manipulation is exempt.

The counterargument is that robotics has a data problem that language models did not. Text is abundant on the web. High-quality manipulation data, of the kind that teaches a hand to grasp a new object in a new position, is expensive to collect and hard to simulate accurately. More compute helps if the data pipeline is there. If it is not, a large training cluster mostly sits.

What buyers should take from it

Two metrics matter more than robot counts. The first is manipulation reliability, meaning whether a hand can do varied grasps without failing. The second is generalisation, meaning whether a behaviour trained on one set of objects transfers to new ones, under new lighting, in a new layout. A weekly production figure says nothing about either.

Compute commitments like Figure's are aimed squarely at the second metric. Whether that works is an open question, and the answer will not come from a press release. It will come from whether a robot trained on Barstow capacity can pick up an object it has never seen and put it down correctly, on a line, without a supervisor watching.

The cost of being early

There is also a straightforward financial risk. Compute deals are multi-year obligations, and the GPU generations move fast. A commitment sized for today's training runs can look oversized or undersized in eighteen months, depending on which way the field moves. Figure has bet that it will need more, not less.

The bet is defensible. The frontier in robotics is moving toward larger models, and larger models need more compute. The question is timing. If generalisation arrives on the schedule Figure is implying, the early commitment looks prescient. If it takes longer, the company is carrying a large fixed cost while it waits for the problem to be solved.