Before Robots Sell, Chips Sell First: Following the Money in the Humanoid Race
Billions of dollars are moving before humanoid robots reach mass sales. Compute suppliers may see revenue first, but that does not guarantee stock returns.
Billions of dollars are moving before humanoid robots go on sale in volume. Robot makers pay up front to train their AI. Chip and AI infrastructure suppliers get a chance to book revenue from that development race. But revenue arriving first does not guarantee a stock return.
What $3.5 billion says about the order of money
In September 2026, humanoid developer Figure announced an initial $3.5 billion compute commitment with AI cloud company Nscale. It also said it intends to expand the deal to more than $6 billion. The potential plan involves up to 100,000 Nvidia GPUs, with an initial deployment target in the second half of 2027. Nscale announcement
The figure should not be read as $3.5 billion of revenue Nvidia has already secured. What Figure wants to buy is Nscale's compute service. Actual GPU orders, equipment supply, data center start-up and revenue recognition are separate steps. The "more than $6 billion" is an intention to expand, not a firm commitment.
An upfront cost for Tesla, an upfront opportunity for chips
Robot companies must collect data, train AI models, and go through repeated testing and manufacturing before they can cut real customers' costs. They have to spend before they earn.
Chip suppliers can find business in development compute demand even if robots are not yet mainstream. But chips bought by a cloud company do not automatically become permanent demand from a robot company. Shortfalls in development funding, delays in securing data center power, a switch to in-house chips and gains in training efficiency can all change that demand.
Optimus: hours worked, not units built
In its third-quarter 2026 materials, Tesla said it had started Optimus production on a high-volume line in Fremont. The stage the company cites is an early ramp. That does not mean it is selling in volume to outside customers at a profit. Tesla Q3 update
Tesla has its car factories as a test environment. The first question is whether robots can do useful work internally and lower labor costs, defects and work time. The next stage is purchases and reorders by outside companies. Unit output alone should not be counted as success at both stages.
Investors should check several measures together: hours of unsupervised continuous work, success rates, human supervision time, breakdown and repair costs, and total cost of ownership per robot.
Nvidia's market is more than the robot's body
Nvidia divides robotics compute into training, simulation and on-device inference. Data centers teach behavior, virtual environments let robots rehearse risky attempts repeatedly, and the chips inside the robot sense and decide in the real world. Nvidia robotics
While robot sales are small, training infrastructure may grow first. As real deployments increase, on-device inference chips, sensors and power-management parts could matter more. But the three markets should not be assumed to have the same growth rates or margins.
What Figure shows about production versus sales
In April 2026, Figure said it had built more than 350 Figure 03 units and raised its production pace to about one per hour. The company also explained that a larger robot fleet is meant to build up training data and operating experience. In other words, production figures are not paid external sales figures. Figure announcement
A humanoid is tested three times: can it be built, can it work reliably, and will customers pay repeatedly. Investors should keep these three questions separate.
Three compute markets in robot AI
| Market | Where the money goes | What investors should check |
|---|---|---|
| Training | Data center GPUs | Actual purchases and utilization |
| Simulation | Virtual testing infrastructure | Software usage and revenue |
| On-device inference | Chips inside robots | Units deployed and revenue per chip |
Insight Times Editorial Desk
