Industry context suggests that scaling data and compute could boost Helix’s performance, but the link between training scale and real-world robot capability remains uncertain and depends on transferring from human demonstrations to robot actions.

Figure strikes a multi-year deal with UK compute provider Nscale to access up to 100,000 GPUs on Nvidia’s Vera Rubin platform to train Helix, with initial deployment planned in Barstow, Texas, in late 2027.

Figure identifies data as a parallel constraint to compute and highlights its Index program, which pays people to record household and workplace tasks to gather training videos; as of August, Index had over 16 million uploaded videos, 264,000 app downloads, participation in 108 countries, and $15 million paid to contributors.

Milestones beyond chip delivery include proving larger training runs yield meaningful capability gains, showing improvements in factory and household tasks, and validating reliable, safe execution on actual hardware rather than cloud simulations.

The compute agreement starts with an initial commitment of $3.5 billion and aims to scale beyond $6 billion, though exact GPU tranche sizes and deployment schedules remain undisclosed.

Overall, the piece presents a high-stakes bet on accelerating humanoid robotics through massive compute and data collection, while acknowledging long lead times and significant uncertainty about practical gains.

Nscale becomes Figure’s preferred compute provider and shareholder, with plans to explore deploying humanoids within Nscale’s supply chain, though no customer deployments were announced yet.

Index processes data through filtering, fraud review, deduplication, rebalancing, and annotation to turn raw footage into usable training material, with ongoing evaluation of generalized results.