Stealth startup XDOF raised $70 million for the data pipelines that would teach humanoid robots to do physical work, the kind of infrastructure that powered the LLM boom but barely exists for robots yet.
A person in a back room moves a robotic arm through a folding task while a camera records every joint angle. The demonstration, captured in a single hour, becomes one data point in the slow, expensive work of teaching a humanoid robot to fold laundry. That kind of physical-interaction data barely exists for the AI industry, and a stealth startup called XDOF is betting $70 million that building it will be the next great AI infrastructure business.
According to Tim Fernholz's TechCrunch report, XDOF, pronounced "ecks-doff," is emerging from stealth today with a $70 million round from Thrive Capital and Spark Capital and joined by Andreessen Horowitz, Lux Capital, and WndrCo. The company says it has 20 customers, including unnamed "frontier AI labs," for teleoperation-driven data collection and annotation pipelines aimed at what the industry calls "physical AI": AI systems that act on the physical world, not just in chat windows or image grids.
The bet is forming in the open next to a reported relaunch of OpenAI's robotics program, which the company had shuttered in 2021, as Fernholz reports. Frontier labs that spent the last two years training on scraped web text are now watching robotics teams try to bootstrap on demonstrations that, as Fernholz notes, "barely exists" — referring to the data gap itself rather than the demonstrations. The pattern is familiar: a foundational data gap, a wave of labs rushing to close it, and a third-party infrastructure layer emerging to monetize the work.
"The arc that Scale AI and Surge followed for LLM training data is now starting for physical AI," XDOF founder and CEO Philipp Wu told TechCrunch, characterizing the situation as one where "labs that were late to build language-model data infrastructure are now trying not to be late to this one." Wu added: "We've already seen some of the downfalls of falling a little bit behind."
That is the category thesis, and it is plausible. Scale AI became a significant annotation and labeling business for the GPT and Claude generations — a market-positioning comparison the company draws — though independent confirmation of the multi-billion valuation is not in the provided source. If the same logic holds for robots, the company that owns the data pipeline for physical AI could own the chokepoint the next generation of foundation models depend on. XDOF says it has 20 customers, but the names and contract values are not disclosed, and "several frontier AI labs" is a company claim that no one outside the company can verify.
The reason the bet is also fragile is the reason text-trained models were never vulnerable to it. Large language models were built on a corpus of public web text that was, in effect, free to scrape and near-infinite in scale. Robotic training data is neither. Every demonstration has to be physically recorded by a human operator, often in a controlled environment, and then annotated frame by frame. The cost is closer to medical imaging annotation than to web scraping, and the throughput is orders of magnitude lower.
That gap is exactly what XDOF is trying to industrialize. Wu and CTO Fred Shentu built GELLO, a low-cost teleoperation system, at UC Berkeley before spinning the company out. XDOF now sells that pipeline, including collection rigs, remote-operator tools, and annotation workflows, to labs that would otherwise have to build it themselves.
The skepticism is not whether the data is needed. It is whether the demand is real enough to justify the round. Teleoperation-derived data has not yet been shown to produce foundation-model-quality results at the scale LLMs enjoy. The category is forming faster than the proof.
For the moment, that is the news. The data infrastructure layer for physical AI is assembling, and the first visible instance just raised at a reported $70 million valuation. Whether it scales, and whether the labs paying for it can turn it into a working humanoid, are the questions the next eighteen months will answer.