100,000 GPUs for Humanoids: Why Figure Is Making a $3.5…

A humanoid robot may stand less than two metres tall, but the computing system proposed to train its intelligence could fill an industrial data centre. Figure and AI cloud company Nscale have announced a multi-year partnership that could provide access to as many as 100,000 GPUs on NVIDIA’s Vera Rubin platform.

The agreement describes an initial compute commitment of $3.5 billion, with an intention to scale beyond $6 billion. Initial deployment is targeted for the second half of 2027 at Nscale’s site in Barstow, Texas. The figures are remarkable, but they require careful reading: “up to” is not the same as installed capacity, and the companies have not published the complete payment schedule or utilization guarantees.

What you need to know

Figure selected Nscale as a preferred computing provider for future versions of its Helix robot AI.
The partnership could scale to 100,000 NVIDIA Vera Rubin GPUs.
The initial commitment is stated at $3.5 billion, with an intention to exceed $6 billion.
Nscale will also become a Figure shareholder.
The first GPUs are targeted for deployment from the second half of 2027, not immediately.

Why a robot company needs an AI factory

Traditional industrial robots follow carefully programmed routines. A general-purpose humanoid is expected to interpret language, understand video, plan movements and adapt when an object or workplace changes. Training such models involves enormous numbers of examples and repeated evaluation.

Figure says its Helix models are constrained primarily by data and computing power. The company recently introduced Index, an effort to collect varied humanoid training data, and claims that the system is generating 35 minutes of data every second. Processing those demonstrations, testing new models and running simulations creates a workload closer to frontier AI research than to conventional factory automation.

From demonstration to simulation and back

The proposed infrastructure is intended to support a loop. Robots and human operators collect examples from physical tasks. Models learn from those examples in a data centre. Simulated environments test many variations without risking hardware or people. Improved models then return to robots for real-world evaluation.

NVIDIA calls this a physical-AI flywheel. The attractive part is speed: one physical fleet can generate data while far more virtual robots attempt related tasks in simulation. The limitation is reality itself. Contact, friction, deformable objects and unexpected human behaviour remain difficult to reproduce perfectly. A model that succeeds in simulation still has to prove itself on real hardware.

What 100,000 GPUs does—and does not—mean

The maximum figure would place Figure’s ambitions among very large AI infrastructure projects. It does not mean that 100,000 GPUs are already operating for humanoid training. Deployment begins in the future, depends on data-centre construction and supply, and may grow in stages.

More compute can support larger models, more experiments and faster iteration. It cannot automatically fix weak training data, mechanical failures or unsafe behaviour. A humanoid still depends on batteries, actuators, sensors, thermal limits and maintenance. Intelligence trained in a data centre must run with limited power and latency on the robot itself.

The economics are as important as the engineering

eWeek notes that the initial compute commitment is larger than the capital Figure has publicly raised to date. That comparison does not reveal when payments are due or how the agreement is financed, but it shows the scale of the bet.

The relevant question is not simply how many GPUs Figure can access. It is how much useful robot capability each additional unit of computing produces. Investors and customers should watch task-success rates, interventions, generalization to new environments and the cost of training each deployable skill.

The agreement also links two supply chains. Nscale will provide computing infrastructure and invest in Figure, while the companies say they will explore using humanoids in Nscale’s own operations. That could create a customer as well as a supplier, although no deployment volume for that use case has been disclosed.

What Figure must prove next

Figure now has an unusually large compute roadmap. The next evidence should come from robots: longer autonomous operation, repeatable performance outside a prepared demonstration, safe recovery from mistakes and clear productivity against existing automation.

Disclosure will matter too. Without utilization, energy and performance data, GPU counts are better understood as capacity ambitions than as evidence of progress. The most convincing milestone would be a direct line from more training compute to a measurable improvement in real work.

Training compute is not the same as robot intelligence

A large GPU allocation gives Figure more room to run experiments, train larger models and process its growing stream of humanoid data. It does not automatically produce a capable robot. Physical intelligence still depends on the quality of demonstrations, the design of the learning objective and a testing regime that exposes brittle behavior. Compute amplifies the pipeline it is given; it can accelerate a good learning loop, but it can also make an expensive mistake faster.

The distinction matters because the announced capacity is a ceiling rather than a report of hardware already running. Figure and Nscale target the first deployment for the second half of 2027. Between now and then, chip delivery, construction, power availability, networking and software efficiency can all change the useful capacity and its cost.

Why the data center becomes part of the robot

A humanoid’s onboard computer must respond quickly and operate within strict limits on power, heat and weight. The largest training workloads happen elsewhere. Data centers can replay demonstrations, generate simulated variations and compare many model versions before a compressed policy is deployed to the machine. Experience from deployed robots can then return to the training system, creating a loop between the fleet and the cloud.

This architecture makes infrastructure choices strategically important. Slow data movement, unreliable storage or inefficient scheduling can leave costly accelerators idle. A provider that combines power, networking, storage and orchestration may therefore matter almost as much as the GPU model itself. Nscale’s role as preferred compute provider places it close to Figure’s development cycle rather than treating capacity as a one-off purchase.

The energy and construction test

At this scale, the practical bottleneck moves beyond buying chips. A campus needs grid connections, transformers, cooling equipment, water or alternative heat-management systems and high-speed links between servers. Deployment in Barstow, Texas, also turns local permitting, construction schedules and power contracts into robotics variables. None of those details is visible in a humanoid demonstration, but all can determine how quickly a new model reaches the fleet.

The announced dollar figures should therefore be read as a long-term infrastructure commitment, not as a conventional hardware invoice. The initial $3.5 billion commitment and intention to exceed $6 billion span services and future capacity whose exact commercial terms have not been disclosed publicly.

Bottom line

The humanoid race is no longer only a contest of motors, hands and balance. It is becoming a competition for data centres, advanced chips, power and the ability to turn simulation into reliable physical behaviour.

Figure’s $3.5 billion commitment shows how expensive that transition could become. Whether the investment creates broadly useful robots will be decided not by the number of GPUs announced, but by the work the machines can repeatedly complete.