Figure AI plans $3.5 billion humanoid supercomputer build
Figure AI plans to commit $3.5 billion to a computing project with British AI infrastructure company Nscale, with an option to expand the agreement to $6 billion. According to Mikekalil.com, the cluster could eventually include 100,000 NVIDIA GPUs dedicated to training Figure’s humanoid robots.
The first systems are targeted to come online at Nscale’s campus in Barstow, Texas, during the second half of 2027. The planned Figure AI supercomputer remains a future build rather than operating infrastructure, and neither company has disclosed an installation schedule for the full 100,000 GPU configuration.
Key facts
- Initial commitment: $3.5 billion
- Potential maximum: $6 billion
- Planned scale: Up to 100,000 NVIDIA GPUs
- First systems: Second half of 2027
- Location: Barstow, Texas
Compute for Figure’s Helix training pipeline
Figure says the cluster could use NVIDIA’s next generation Vera Rubin platform. NVIDIA reports that one Vera Rubin GPU can provide up to 35 petaflops for AI training, 50 petaflops for inference and 22 terabytes per second of memory bandwidth. Those are processor specifications, not measured performance for Figure’s proposed cluster.
The compute is intended to train Helix, Figure’s self developed system for connecting visual observations, actions and task objectives. Its training pipeline increasingly depends on Index, a smartphone based program that pays participants to record everyday tasks with head mounted cameras.
Figure reports that Index had 44,000 weekly active users across 108 countries before its formal launch. The company says participants are collecting the equivalent of almost five years of human activity each day, while its pipeline processes more than 35 minutes of new task video every second.
A five stage process filters low quality and near duplicate recordings, groups videos by task and adds text descriptions. Figure says the resulting material is used to teach Helix how human actions relate to goals that its robots may encounter in the real world.
These are company supplied measurements. The volume of video does not by itself demonstrate that Helix can transfer human behavior reliably to a humanoid body, particularly across unfamiliar environments and long horizon tasks.
A large commitment with financing unresolved
Figure has separately committed more than $1 billion to data and computing over the next 12 months. Including the initial Nscale agreement, Mikekalil.com calculates that its public commitments now total at least $4.5 billion and could exceed $7 billion if the infrastructure partnership reaches its maximum size.
The company has raised around $2 billion, including more than $1 billion in a 2025 Series C round that valued it at $39 billion. Figure has not explained how it will finance the full infrastructure plan.
The proposed cluster supports Figure 03, a 173 cm, 62 kg electric humanoid with five finger hands and a reported operating time of about five hours per charge. Figure has shown the robot performing manipulation, household and logistics tasks, but the article notes that many of the strongest performance claims in this sector still come directly from companies seeking substantial capital.
The immediate test is therefore narrower than the proposed 100,000 GPU scale: whether Figure and Nscale bring the first Barstow systems online in the second half of 2027 as scheduled.
Source: Mike Kalil
