Two Months Old, Worth $500M: Huawei's LLM Twin Stars Move into Physical AI

Physical-AI company XIRRA announced on September 23 that it had completed consecutive seed and angel rounds totaling several hundred million yuan within two months: led by Dunhong Asset, with Huakong Fund, Sanhua Holdings, Gingko Capital, Zeran Capital, Benjian Fund, Shanghai Angel Association, and Biaopu Investment participating; the company discloses a valuation of $500 million. Founded in July by Li Yin, former CTO of Huawei Cloud's large models, with chief scientist Zhang Hanwang — Huawei's former chief scientist for multimodal AI and a president's chair professor at NTU — it is a physics-foundation-model company that is two months old and worth $500 million.

The nut graf: two of Huawei's core large-model people quit to found a startup worth $500 million two months in, led by industrial capital — the financing news carries more information than the numbers. It signals that "physical AI" is moving from paper concept to a serious private-market category, and the talent supply for this cycle comes directly from the engineering bench built during the self-developed-LLM era. XIRRA's approach is also worth unpacking: it is not building a general multimodal model but a "physics foundation model" — native physical-AI pretraining at scale plus real physical-interaction data, pointed at embodied AI and world models — spending on data and models rather than stacking applications.

The people and the direction make the logic of this startup path clear. Li Yin previously built Huawei Cloud's LLM program from zero, running large-scale training, cluster management, and industry delivery; Zhang Hanwang led the multimodal track at Huawei. Together, founding a physics-foundation company merges LLM engineering capability with multimodal perception toward an embodied-AI goal. The financing cadence shows capital's acceptance: seed and angel rounds closed within two months, and the follow-on list includes manufacturing capital like Sanhua Holdings — for physical AI, industrial parties are both investors and potential deployment sites (robotics, industrial automation, simulation), a clearly different logic from pure-software AI investing.

Attribution and boundaries: the amounts (several hundred million yuan), the $500 million valuation, and the founders' résumés come from company disclosure and media reports, with no independent financial audit. The "physics foundation model" technical route and commercialization timeline are not public; model scale, data sources, and first deployment scenario have no details yet — a two-month-old company, however smooth its valuation story, still has to deliver products and customers. In the larger frame: big-model talent spilling out of giants, industrial capital entering, and embodied AI moving from demo to pretraining — these three lines converge in 2026, and XIRRA is just the first sample to be reported.

There is also a telling detail in who is not leading this round. The lead investor is Dunhong Asset, with manufacturing capital like Sanhua Holdings in the follow-on list — no big-model platform player is named as a major backer. That suggests the financing is being priced on the physical-AI thesis and the founders' track record, not on an existing distribution channel or dataset moat, which is both the opportunity and the risk: the company must build its data and model advantages from scratch, in a field where open-world physical interaction data does not yet have a settled source or standard. For anyone tracking where the next talent exits and capital flows in Chinese AI, XIRRA is a data point worth watching past the valuation headline.