Ex-Li Auto execs raise nine figures for a robot foundation model
Robot brains and AI infrastructure drew the money this hour: a Chinese embodied-model lab run by two of Li Auto's most senior AI executives closed a nine-figure angel round, and one of OpenAI's most visible former leaders resurfaced on the board of a data-center company heading for an IPO.
Xieyue Intelligence has raised several hundred million yuan in an angel-plus round, announced September 12, with Linear Capital, Junshan Capital, Hongyi Capital and Yinshan Capital participating. The Hangzhou company was founded in February 2026 and is building what it calls a general embodied model — a foundation model for the physical world — aimed at homes rather than factory floors. Its co-founders are Chen Wei, Li Auto's former chief AI scientist and head of its foundation-model division, and Zhang Xiao, the automaker's former product line president: the two people who respectively ran the model stack and the product roadmap of one of China's largest EV makers. Chen's argument for why the round matters is that embodied intelligence is nowhere near converged — data-collection methods differ, robot bodies differ, and the ways models get combined with hardware differ — so no single open model can cover the whole chain. Model capability itself, he says, is the moat. Chinese robot-training data has been drawing serious checks all month — we covered this in Mecka AI nears $500 million valuation as Sequoia bets on robot training data.
The technical bet is a paradigm Xieyue calls Duplex Reasoning. Today's robot systems run half-duplex: take a command, execute, report back, with success graded on task completion. Homes don't work that way — the request is vague, changes mid-sentence, and needs to be interrupted mid-task, so what a household actually grades the robot on is whether it stops the moment you say so. Xieyue wants interaction and action modeled together in the data, the architecture and the training run, with a vision-language-action backbone wired to world-model prediction: language compresses what the camera, microphone and environment are doing into transferable high-level semantics, while the world model forecasts the consequences of an action to constrain long-horizon tasks. The company is candid that the hard part is turning "interruptible at any moment" into stable model behavior instead of patching it with engineering rules.
Where it lands first is deliberately unglamorous. Xieyue plans to validate cross-space generalization, task success and unit economics in semi-structured settings — hotels and elder-care facilities — before going into homes, prioritizing high-frequency, long-horizon chores like laundry, tidying and cleaning. On the data side it has built its own egocentric capture hardware and a pipeline layered across first-person, body-agnostic and body-specific data, and says it intends to have the full collect-clean-label-train loop running inside 2026. On hardware it is staying with a single body form and a closed stack, to keep the hardware variable out of model development until the model earns the right to ship.
Fidji Simo, who stepped down as OpenAI's second-in-command in July for health reasons, is joining the board of Nscale. The UK data-center company is preparing an IPO this fall, and it framed the hire around scale: founder and CEO Josh Payne said Simo "understands what it means to scale products used by billions of people and the demands those products place on the underlying systems." Simo previously ran Instacart and led the Facebook app at Meta, and joins a board that already includes Sheryl Sandberg, Susan Decker and former UK deputy prime minister Nick Clegg. Nscale has been in the news as a buyer of compute and a seller of it — it was Nscale lines up $3.5B in pre-IPO cash with Nvidia at the table four days ago — and its IPO pitch rests on the same thesis as Simo's new board seat: the demand for AI compute is a consumer-scale operations problem now, not a hardware one. Anthropic is among its signed tenants, which makes the board list part of the disclosure story as much as the deal story.
What to watch: whether Xieyue's duplex claim survives real household data — the test is whether "interruptible" becomes trained behavior or stays a demo trick — and whether Nscale's board-heavy, revenue-light pitch clears IPO diligence in a market that has started asking who actually pays for the compute.
Does an embodied foundation model aimed at homes beat the robotics startups that have already shipped products into real households? Tell us in the comments.
Sources: Beijing News (贝壳财经) · Sootoo via 10jqka · Eastmoney · TechCrunch · Nscale press release · Bloomberg
