Figure AI's Index platform has 69,943 weekly active users, as of 4 September 2026. Ten days earlier, at public launch, the figure was 44,000. The app pays people to record everyday tasks on their phones, and the footage feeds training for the Helix model.
Key takeaways
- 69,943 weekly active users, up from 44,000 at the 25 August 2026 launch
- 35 minutes of video uploaded every second, up from 30 minutes
- $15M paid out to Creators so far
- Every 1,000 hours of footage yields 373 unique tasks, 1,146 objects and 116 environments
- Nscale deal: up to 100,000 GPUs and a $3.5B initial commitment, starting H2 2027
Where the data comes from
Index is a phone app. The user — Figure calls them a Creator — records first-person footage of folding laundry, cooking or packing goods in a warehouse. They get paid for the clips.
The material runs through a five-stage pipeline: automated quality filtering, human fraud review, deduplication by embedding similarity, rebalancing against per-task quotas, and hierarchical text annotation.
The numbers that mean something
The most telling figure is not user count but data density. Every 1,000 hours of footage contains 373 unique tasks, 1,146 distinct manipulated objects and 116 environments. That measures diversity, not raw volume.
The pace is striking too. From 44,000 weekly actives at launch to almost 70,000 in ten days, with throughput rising from 30 to 35 minutes of video per second.
What it is all for
Figure is building the infrastructure to match. The Nscale deal covers up to 100,000 GPUs on the NVIDIA Vera Rubin platform in Barstow, Texas, with a $3.5B initial commitment and plans to pass $6B. First deployment lands in the second half of 2027.
Our AI model, Helix, becomes more capable the same way every learned system does: with more data and compute.
Brett Adcock, CEO of Figure AI.
Why it matters
Figure's bet is that generalization is a data problem, not an architecture problem. If that holds, the edge goes to whoever buys diversity of human behaviour most cheaply, and paying people for phone footage costs less than teleoperating a robot. The risk is concrete: phone video carries no proprioception and no contact forces. How much of this volume actually converts into control remains open.
What's next
- The first Nscale GPU deployment is set for the second half of 2027 — until then Figure trains on existing infrastructure
- The company has committed over $1B to data and compute over the coming year
- Figure has not published how Index growth translates into Helix performance, which is the missing piece
