XDOF at $1.2B: the training data is the product

A company that does not build robots and does not train models is in talks to raise at a $1.2 billion valuation, three months after leaving stealth. XDOF collects robot training data — humans remotely driving robot arms through ordinary tasks so a model has something to learn from. The frontier labs building the robots are the customers. That is the whole story, and it says something specific about where the value in an AI stack actually sits.

What actually happened

TechCrunch reported September 4 that XDOF is in late-stage talks for a Series B led by 8VC at roughly a $1.2 billion valuation. The round is not closed. It would follow a $70 million Series A in June 2026 with Thrive Capital, Andreessen Horowitz, Lux, and Spark Capital participating.

XDOF was founded in 2024 by UC Berkeley researchers Philipp Wu (CEO) and Fred Shentu (CTO), who previously worked on GELLO, a low-cost teleoperation rig that lets a human drive a robot arm to generate training examples. TechCrunch reports annualized revenue approaching $50 million across about 20 customers, several of them frontier AI labs, and says investors approached the company rather than the other way around.

The company's pitch is that it runs the data pipelines, collection tooling, and annotation systems that labs cannot easily stand up themselves — an outsourced supply chain for the one input that has no shortcut.

Why proprietary training data matters for your business

You are not selling robot demonstrations. But run the logic backward and it lands on your desk.

Models are converging. Weights leak, prices fall, and this month's frontier model is next quarter's commodity endpoint — we have written that sentence about a dozen releases. What does not converge is data that only exists because someone did the work in the real world. XDOF is worth a billion dollars because a lab with unlimited GPUs still cannot conjure ten thousand hours of a human folding laundry.

Your version of that asset is already on your servers and you are probably not treating it as one. Five years of quotes with the ones that closed marked. Support threads with the resolution attached. Field notes, service histories, the reasons deals died. That is a labeled dataset describing your specific market, and no general model has it.

Two practical consequences. First, when you wire up an AI vendor, read what they may do with what you send — the training-rights clause is the part of the contract that matters most and gets read least. Second, start capturing outcomes now, even crudely. A ticket log with a "did this fix it" field is worth more in eighteen months than the same log without one, and adding the field costs an afternoon.

The models will keep getting cheaper. Your data will not get easier to collect retroactively.

Sources: TechCrunch.