Figure committed $3.5 billion to compute while its robots fails 44% of trials

Index pretraining alone turned 9% into 56%

Figure says Helix 2.5 walked into 30 Bay Area homes it had never seen, ran three chores off one unmodified checkpoint, and passed 237 of 420 blind trials ¹²³. Grading allowed no partial credit: all 13 to 15 toys in the basket, every towel folded, the comforter smooth ¹. A twin policy trained from scratch on identical data passed 9%; the Index-pretrained twin passed 56%, and Figure says pretraining on its human-video dataset was the only experimental variable ¹³⁴. Every number traces to Figure's own writeup↗; no outside evaluator has rerun it ³.

A four-decimal loss forecast is what the money rests on

The sharper claim sits under the demo: Figure trained four models across an eightfold range of Index data and says it forecast its largest run's action-prediction loss to four decimal places before training began, a human-to-humanoid transfer scaling law it calls the first measured on a humanoid ¹⁴. If the curve holds, the payoff of the next doubling of human video can be priced before the run is paid for ². The Tesla Optimus rivalry is framed the same way: a contest of data moats, Tesla's fleet miles against Figure's paid video ⁵. Index adds roughly 35 minutes of new human experience every second, and its public counter↗ read 24,593,204 uploads on 24 September, against more than 16 million videos when the app launched on 25 August ¹⁴⁶⁸.

The other 44 percent

Under the same all-or-nothing rubric the robot failed 44% of its trials: toy tidying passed 40%, towels 62%, bed making 67% ²³. Sunday Robotics chief Tony Zhao answered within hours: "failing half the time is not 'doing real useful work'" ². His counter-number, 778 of 785 folds at 99.1%, counts single garments, not whole chores, so the two yardsticks do not compare head to head ². The eval's other payer was the 30 households that lent their rooms: short-term rentals staged with toys, decluttered by definition, no pets and no laundry piles, the conditions the 44% is graded under ².

The bill lands before the proof

Figure has raised $1.9 billion in its life, just over $1 billion of it last September at a $39 billion valuation, and has never disclosed revenue ⁸⁹. On 3 September it committed $3.5 billion↗ of Nscale compute, intent to scale past $6 billion, with up to 100,000 Nvidia GPUs landing no earlier than the second half of 2027; Nscale became a shareholder in the same deal ⁷⁸⁹.

Watch two things before a single GPU lands: whether the next doubling of Index moves the 56, and whether Figure, Sunday and 1X (the third humanoid camp in the benchmark argument) ever agree to run one neutral home-task benchmark ². Figure itself concedes the point is not that humanoid robotics is solved ¹. The bet says the coin flip has a scaling law.