XDOF nears a $1.2B valuation three months out of stealth
The money chasing physical AI has moved past the model builders and into the supply chain underneath them. Three stories this hour: a robot-data startup doubling its price tag in a quarter, Hong Kong's listed model labs being asked where the profit is, and a Chinese province standing up a state operator for the thing all of it runs on.
XDOF is in talks for a Series B that would value the robot-training-data startup at $1.2 billion — roughly four months after it raised a $70 million Series A, and only three after it left stealth. TechCrunch reported the talks on Thursday, noting the total raise, whether the figure includes new money, and the terms themselves are all still unsettled; XDOF and the reported lead investor, 8VC, did not respond to requests for comment. Secondary coverage puts the company's annualized revenue run rate near $50 million, a number XDOF has not confirmed.
The pitch is a gap nobody has closed. Language models were trained on a public internet that already existed; robots have no equivalent corpus, because the data that matters — how a hand closes on a fabric, how a box flattens — has to be physically performed to be recorded. XDOF sells the pipelines, collection hardware, and annotation tooling that frontier labs and robotics companies would otherwise have to build themselves, which is why investors keep describing it as the Scale AI or Mercor of physical robotics. The company came out of GELLO, a cheap teleoperation rig co-founders Philipp Wu and Fred Shentu built at Berkeley to generate training data, and it now combines remote robot operation with collectors wearing body sensors to record everyday tasks. It says it already works with 20 customers, including several frontier labs, and is preparing a large public robot-data release with Berkeley's AI Research lab.
The interesting part is what the valuation is actually pricing. A data-vending business is easy to undercut and hard to defend — Wu has said as much, which is why the company is pushing into cleaning, tooling, and annotation rather than raw collection. At $1.2 billion, investors are betting that physical AI's bottleneck stays at the data layer long enough for the plumbing to become the moat. The counter-case writes itself: if synthetic data or simulation closes even part of the gap, a business built on armies of human operators carries cost structure that no amount of software margin fixes.
Hong Kong's two listed model labs reported interim results with growth intact and profit still missing — and analysts have stopped asking about revenue. Zhipu's revenue rose 399.7% year over year while its overall gross margin fell from 50.0% to 26.4%, even as its API business swung from roughly break-even to a 24.6% margin. MiniMax grew revenue 283.1% and lifted gross margin from 12.1% to 17.9%, but still posted an adjusted net loss of $290 million. Both filed in late August.
The two are running opposite plays. Zhipu leans on cloud API sales — 86.5% of revenue — and has been raising prices, with average API selling price up about 101% since the start of the year while token call volume grew more than 40-fold. MiniMax is buying scale: its open platform and enterprise services brought in $73.9 million, 63.4% of revenue, with AI-native products at $42.6 million, and it says July token consumption was 20 times January's. Coverage of the filings cites JPMorgan watching whether Zhipu's models can defend that pricing, HSBC moving its break-even estimate from 2028 to 2027, and Jefferies flagging margin pressure, heavy compute spend, and the API transition as the risks. Usage is no longer the question. Whether 40x call volume turns into durable cash flow is.
Zhejiang has launched a provincial state-owned data company with 2 billion yuan in initial registered capital, wholly owned by Hangzhou Steel Group and advised by Wang Jian — the engineer behind Alibaba's cloud. The Zhejiang Data Group was unveiled in Hangzhou on August 26 with a remit covering public-data authorization, governance, compute supply, model training, and commercialization, backed by a blockchain layer and trusted data spaces. At the signing ceremony it took on six national AI application pilot bases spanning petrochemicals, culture and tourism, geoscience, healthcare, spatial data, and embodied intelligence.
It is one of more than twenty comparable provincial data groups now standing up across China, with Shanghai's registered at 5 billion yuan in 2022 and Shandong's unveiled the week before. What makes Zhejiang's worth watching is the operator: Hangzhou Steel shut its steelworks in 2015 and spent the following decade building cloud data centers, computing hardware, and a government-data business, which means the province didn't have to assemble a new team to run this. The read-through is straightforward — as data becomes the input that models compete on, provincial governments are deciding it should be held and sold by a state champion rather than left to the market.
What to watch: whether XDOF's Series B closes at the reported number, and whether Zhipu's next filing shows pricing power holding at 40x call volume.
If robot data is the new oil, should the refineries be private contractors or public utilities? Tell us in the comments.
