Figure is buying its data moat by the minute, and it just published the price

Put the two throughput numbers side by side and they do not describe the same operation. Thirty minutes of video ingested every second is 1,800 hours per wall-clock hour, or 302,400 hours a week ^[3][2]. Spread across the 44,000 weekly active users Figure claims, that is about 6.9 hours of recorded footage per contributor per week ^[3]. Now price it: the $15 million paid during four months of stealth is roughly $862,000 a week, about $20 per weekly active user, which against 6.9 hours of filming is around $2.85 an hour ^[2][4]. Contributors are paid by the minute ^[9], and nobody is filming their own laundry at that rate. The likelier read is that current ingest sits far above the average that produced the $15 million, in which case the $15 million understates what this network now costs to run, and Figure has published the cheap end of its own price.

The number a competitor can actually copy is the unit one: about $0.94 per upload across 16 million uploads ^[1]. That is the retail rate for a minute or so of first-person household footage in 108 countries ^[2], device logistics, fraud review and annotation not included ^[9][11]. It is now the benchmark. Anyone building a humanoid manipulation stack either matches roughly $3.75 million a month in contributor payments ^[5] or goes back to managed properties and business partners, which is what Figure itself did with Project Go-Big and Brookfield in September 2025 before widening the pool ^[13].

The 100x claim is where the budget stops working. Adcock says Index is on a path to grow one hundredfold, against a commitment of more than $1 billion for data and compute over twelve months ^[6]. Scale the current payment run rate by one hundred and you get about $375 million a month, or $4.5 billion a year, roughly four and a half times the entire stated budget, which also has to cover training compute ^[5][7]. Something has to give: the per-minute rate, the share of uploaded minutes that get paid, or the meaning of 100x. For context, that $1 billion is about the size of Figure's whole Series C, announced in September 2025 at a $39 billion post-money valuation ^[8].

Which brings up the number Figure did not publish. Every submission goes through quality filters, fraud review, deduplication, rebalancing and annotation, and clips judged too similar to accepted footage are discarded ^[11]. Figure describes retained data by density instead of volume: per 1,000 hours, 373 distinct tasks, 1,146 manipulated objects, 116 environments ^[4]. At the stated ingest rate, 1,000 hours of raw video arrives every 33 minutes ^[6], so the discard ratio decides whether this is a training corpus or a very large pile of near-duplicate bedmaking. There is no standardized total of accepted training hours in the announcement, and every scale claim in it is Figure's own ^[5]. Uploads and downloads describe distribution. The moat is whatever survives the filter, and that quantity remains unpriced.