94 cents a video to train your replacement

Agility Robotics is charging six figures a year for a robot while generating only $1.78 million in annual revenue and losing $138 million. If you think these numbers look bad, try finding out what their robots are actually capable of doing.

Agility’s September 4 Form S-4¹, filed for a merger that takes it public, is the first place a major US humanoid company has been forced to show its workings. The filing reports $1.78 million in 2025 revenue against a $138.1 million net loss. It also says management sees substantial doubt about the company's ability to keep operating, and that the merger itself does not resolve it.

Then there is the $300 million backlog Agility has been using to describe demand for Digit v5. The filing says it is one three-year rental agreement covering 1,000 robots, sold to a related party (a customer with another financial connection to Agility) and carrying 453,000 stock warrants that vest as robots are deployed. Agility says it is still working out how to account for the arrangement. The headline number is therefore real as a contractual commitment, but it is not remotely the same thing as $300 million already sitting in the bank.

Agility has at least done one thing the rest of the industry has avoided: it published a price. Aaron Prather laid out the figures on LinkedIn: approximately $8,500 a month to rent Digit, plus a $25,000 deployment fee, or roughly $200,000 to buy one, plus $20,000 for deployment and about $36,000 a year for software and maintenance. I ran the monthly figure against a normal full-time year: that is $102,000 before the deployment fee, electricity, downtime or anything else. Agility's margin calculation excludes its overheads and research spending — which is where most of the $138 million went.

Peter Corke, a robotics academic, compared that $8,500 monthly fee with Australia’s minimum wage and calculated that Digit costs roughly two to 8.4 times as much as a human, depending on how continuously the robot works. His point was not that Digit fails the test. It was that we still do not know the other half of the equation: how much useful work does it produce? Sean Dotson, an industrial automation founder, ran a similar calculation against US factory wages and came to a less charitable conclusion, arguing that “They are, unfortunately, doomed. Possibly before they even get to the SPAC”.

Putting a price on the robot changes the question. Nobody's asking whether Digit is impressive anymore. They're asking whether it earns its keep. Agility didn't choose to answer that question. The S-4 forced it out of them. Figure, Apptronik and 1X haven't had to answer it yet. I don't think that's a coincidence.

Figure has done something unusually candid: it has described the people training its humanoid robot as the people whose work the robot is supposed to eliminate.

Index is an app where Figure pays “Creators” to film themselves doing ordinary physical tasks such as cleaning, folding and tidying. Figure also offers in-person sessions, where these creators are booked to perform tasks while their movements are captured for training data. Both are ways of teaching its robots how people actually move through the physical world.

The scale is already impressive. In its own Index announcement, Figure said it had paid Creators $15 million. The same announcement said more than 16 million videos had been uploaded in the four months before launch. That is roughly 94 cents per uploaded video on average. It is only a rough average, because some of the $15 million went to people booked for in-person sessions rather than video uploads, and Figure does not describe it as a per-video rate.

Figure also published this sentence: “Today, you have people coming to help clean your house; eventually, a robot will do everything for you.” That is not an analyst projecting what automation might eventually mean or a critic interpreting a business model. It is Figure describing the sequence itself, on its own website.

This is in line with Adcock’s framing of their vision. In May, Figure livestreamed a ten-hour "Man vs. Machine" contest on its own channel, pitting a company intern against one of its warehouse robots. Aime Gerard, the intern, won by 192 packages. Adcock's response after the event: "the last time a human will ever win."

The mechanics are almost comically literal. Fold a shirt for the camera, and you are training the machine that will eventually fold it without you. Wipe a counter, upload the clip, and you have handed over one more example of a job a robot no longer needs a person to do.

I don't need to speculate about whether that is where the economics of humanoid robotics are heading. Figure has already supplied the roadmap. The people being paid today are generating the data that makes them redundant tomorrow.

For years this industry ran on demo footage, which nobody outside it could really check. Both of these stories run on numbers instead. A monthly rental rate, a payout per video, and a number can be checked by anyone with a calculator. Within a week of Agility's filing, two people who don't work there had already done it.