Figure AI will build a supercomputer for humanoid robots for $3.5 billion

Figure AI will invest $3.5 billion in a computing project with British infrastructure company Nscale. The agreement provides for an option to expand to 6 billion. In the future, the cluster may include 100 thousand NVIDIA GPUs for training Figure humanoid robots.

Key figures

Initial commitments: $3.5 billion

Potential maximum: $6 billion

Planned scale: up to 100,000 NVIDIA GPUs

First systems: second half of 2027

Location: Barstow, Texas

The first systems should be operational at the Nscale site in the second half of 2027. The planned supercomputer still remains a project of the future, and not an existing infrastructure. Neither company has disclosed a schedule for installing the full configuration on 100 thousand GPUs.

Calculations for the Helix training pipeline

Figure says the cluster could use NVIDIA's next-generation Vera Rubin platform. According to NVIDIA, a single Vera Rubin GPU is capable of delivering up to 35 petaflops of AI training, 50 petaflops of inference, and 22 terabytes per second of memory bandwidth. These are processor characteristics, and not the measured performance of the future cluster.

The computing power is dedicated to training Helix, Figure's proprietary design that connects visual observations, actions and goal-oriented tasks. Its training pipeline increasingly relies on Index, a smartphone program that pays participants to record everyday activities using head-mounted cameras.

Figure reports that before the official launch, Index had 44 thousand weekly active users from 108 countries. The company claims that participants collect material equivalent to nearly five years of human activity every day. The pipeline processes over 35 minutes of new task video every second.

The five-step process weeds out low-quality and nearly duplicate posts, groups videos by task, and adds text descriptions. According to Figure, the resulting material is used to teach Helix how human actions relate to goals that robots might encounter in the real world.

This is data provided by the company itself. The length of the video alone does not prove that Helix is ​​capable of reliably transferring human behavior to a humanoid body, especially in unfamiliar environments and during long-term tasks.

Large commitments with no clear funding

Separately, Figure has made more than a billion dollars in data and computing commitments over the next 12 months. Including the initial agreement with Nscale, the company's public liabilities now stand at a minimum of $4.5 billion. They could exceed 7 billion if the infrastructure partnership reaches its maximum size.

The company has raised about $2 billion, including more than a billion in a Series C round in 2025 that valued it at $39 billion. The company did not explain how it plans to finance the entire infrastructure plan.

The proposed cluster supports Figure 03, a 173 cm tall, 62 kg electric humanoid with five-fingered hands and a claimed battery life of about five hours on a single charge. Figure demonstrated the robot performing manipulation, household and logistics tasks. The article notes that many of the boldest performance claims in the sector still come directly from companies seeking significant capital.

The next test, therefore, is no longer about the scale of 100 thousand GPUs. The question is whether Figure and Nscale will launch the first systems in Barstow in the second half of 2027 as planned.