Brett Adcock’s compute bet: what Figure still has to prove

More training video gives Figure more material to learn from. Whether its robots improve depends on what they can do after training. The October update to Adcock’s data strategy belongs beside the earlier Helix evaluation and the future compute schedule: each answers a different part of that question.

Brett Adcock’s September compute argument is straightforward: better robots need more training data and more computing power. Figure’s September 3 agreement with Nscale puts a large infrastructure commitment behind that view. The company describes an initial $3.5 billion compute commitment and up to 100,000 GPUs, with initial deployment targeted for the second half of 2027. Those are future commitments and targets, not installed capacity today. Figure’s announcement

In the same announcement, Adcock connects Helix’s progress to more data and compute. That is the proposition to follow. The size of an agreement tells us something about the resources being sought; it does not measure the intelligence of the robots those resources may eventually train.

The earlier piece was access to the real world

A year earlier, Figure’s Brookfield partnership centered on collecting varied real-world training data and exploring deployment opportunities. Adcock described access to household environments as part of the path toward more general capabilities. September 2025 partnership

Read together, the announcements describe two inputs to the same strategy: varied experience for the model to learn from, and computing capacity to train on it. That is a coherent explanation of what Figure is trying to build. Those announcements alone do not measure how efficiently one input turns into a useful new skill.

October context: collection size and useful training data

The next useful comparison is how much of the added material survives selection and whether it reduces failures on the same tasks. A larger collection can add repeated footage as well as new situations. A repeatable evaluation can distinguish those possibilities; an upload total cannot.

The earlier performance test: Helix 2.5

Figure’s September 17 report supplies a different kind of evidence. In its Helix 2.5 comparison, Index pretraining raised task success from 9% to 56% across 30 unseen homes. The three behaviors were trained with task data collected elsewhere; zero-shot refers to new homes and objects, not learning the tasks without examples. These are Figure’s reported results. The evaluation and its scope

This strengthens the argument that broader experience can help transfer a learned behavior. It does not establish dependable household service, and it does not show that the future Nscale infrastructure has arrived. Model progress and the infrastructure delivery schedule remain separate things to track.

What the factory evidence can tell us

BMW provides a more concrete place to look. Its June account describes Figure 02 handling sheet-metal parts in production, followed by a planned Figure 03 logistics workflow. Those are defined tasks in a defined setting, which makes them more testable than a general promise about robots in homes. BMW’s account

But success at one task does not answer how much retraining another task needs. For the wider strategy, we would want to see a robot learn a new workflow with less task-specific preparation, retain earlier skills and recover when objects or surroundings change. The relevant improvement is what a customer can use, not simply how much data entered the training process.

The spending clock and the revenue clock

Infrastructure may need to be arranged well before commercial returns arrive. That creates an interval in which commitments become more visible than revenue. The investment case depends partly on what happens during that interval: development progress, customer acceptance and the cost of supplying and supporting the machines.

NVIDIA is explicitly named in the Nscale arrangement, so the connection is documented. That does not mean the entire announced compute commitment becomes NVIDIA revenue at once, nor that an announcement specifies when a supplier recognizes revenue. The underlying services, delivery schedules and contractual details matter.

Brookfield’s earlier involvement is also documented, but that is not a reason to treat every Brookfield-listed entity as an interchangeable proxy for Figure. A stock connection needs the particular entity, exposure and materiality established before it becomes useful to a reader.

What would make the next headline stronger?

A confirmed infrastructure delivery would answer an execution question. A customer reporting dependable operation would answer a deployment question. Evidence that each new task requires less setup would support Adcock’s broader learning argument. These developments could arrive at different times.

For the factory side of this comparison, follow what Figure 02 and Figure 03 actually did at BMW. Those tasks provide a separate check on whether broader learning translates into useful work.