Happy Tuesday. I scan 100+ Chinese-language sources every day, the WeChat public accounts, Bilibili, 36Kr, the finance wires, and translate the signal that English-language coverage misses. If someone forwarded you this, you can subscribe at the bottom.
Let's go.
The Transplant
Moqi Intelligence (墨奇智能) is six months old. It has not shipped a product. Last week it disclosed a valuation above 7 billion yuan, roughly a billion dollars, on more than 1 billion yuan of angel funding, with Alibaba Cloud alone putting in about 900 million.
Forget the number for a second. Look at who took it. Moqi's CTO is Huang Qingqiu, born in 1994, a Tsinghua automation graduate who joined Huawei in 2020 as a "genius youth" hire and went on to run the algorithm team behind Huawei's ADS self-driving system, from version 1.0 to 4.0. He was, by his own account, the first person in the industry to put a one-stage end-to-end driving architecture into mass production at the million-vehicle scale. He left Huawei at the end of last year to build home robots.
He is not the exception. The founders raising this money are almost all coming from the same place. Guangxiang Technology (光象科技), incubated out of Tsinghua's vehicle college, is run by the former head of Amap's spatial-perception engine, the software that shipped to millions of cars. Tashi (它石智航) just set an embodied-AI angel-round record at 120 million dollars, co-led by Qiming. A Chinese tech-press headline this week put it plainly, the "autonomous-driving F4" are now holding up half of embodied AI.
And the money is moving at a speed with no recent precedent. AI2Robotics (智平方), a Shenzhen general-purpose-robot company founded in 2023, closed roughly 5 billion yuan (about 700 million dollars) at the end of June, which makes it the Greater Bay Area's first embodied-AI unicorn valued above 20 billion yuan. That round was, by 36Kr's count, its twelfth in a single year.
Here is the honest half. The people writing these checks are the same capital and the same engineers who just lived through China's self-driving consolidation, where robotaxi valuations ran far ahead of deployment and a shakeout followed. They are running the play again, faster, into a problem that is harder. Huang says so himself. Embodied AI iterates far slower than either self-driving or large language models, because the one thing nobody has cracked is getting "real workers doing real jobs in real settings" to generate training data at scale.
And the physics fights the funding. A humanoid's control loop has to run on the robot itself, the "brain" at roughly 10 hertz and the "cerebellum" at 100, because a hand closing around a paper cup of water has to adjust its grip every millisecond or it crushes the cup or spills it. That real-time demand drags the model out of the cloud's "greenhouse of unlimited compute" and into the "cage of power and heat" on the machine, which caps how big the model can be. The scaling law that made language models work does not transfer here. You cannot simply pour in parameters. UBTech conceded this week that full-size humanoids today run for only two to four hours on a charge.
So the transplant is real, and so is the mismatch. China has a deep bench of self-driving talent freshly available, a capital system, from national funds to insurers to Alibaba and Tencent, willing to fund it at self-driving-boom velocity, and a physical problem that will not bend to either. The bench and the money are the reason to watch this. The physics is the reason to keep the champagne on ice.
The Briefing
Meituan open-sourced a trillion-parameter model trained and served entirely on Chinese chips. LongCat-2.0, released and open-weighted this week, is the first domestically-trained-and-inferenced trillion-parameter model, per 智东西, and Moore Threads had a same-day FP8 adaptation running. English coverage framed it as China claiming the biggest model on local chips, and that is the headline. The quieter signal is the stack, a domestic training run and a domestic inference path shipping together, which is the exact thing the chip controls were meant to prevent.
Beijing published eighteen measures to put scientists under model control. The city's AI-for-Science package encourages "24-hour unmanned labs" and pushes researchers to run experiments through AI operators. This is the industrial-policy tell. China is not only funding the models, it is rewiring how its public research runs to feed them.
Doubao and Qwen are switching off their AI-agent features on July 15. ByteDance's Doubao and Alibaba's Qwen will disable agent and personalization functions, the visible edge of the Cyberspace Administration's "Qinglang" campaign against what it calls chaotic AI applications. The same regulator that cleared the model race is now drawing lines around what consumer AI is allowed to do on its own.
Tencent released and open-sourced Hunyuan Hy3, with agent features free inside its Yuanbao app. The pattern across Meituan, Tencent, and DeepSeek's coming V4 is consistent. The frontier Chinese labs are open-weighting their models and giving away the agent layer, competing on distribution instead of API margin.
Signals
Astribot (星尘智能) completed its joint-stock restructuring, the formal step before an IPO filing, and 钛媒体 is now calling 2026 embodied AI's "big listing year." That puts a public-market exit under the funding surge in the lead story.
Chinese quantum-computing firms are booking real orders and heading for IPOs, several with market caps now above 10 billion yuan. After years of lab demos, the commercial layer is arriving on the same domestic-substrate logic as the chips.
AI2Robotics says it will open China's first production line for tens of thousands of service humanoids in the second half of this year. Watch whether the unit shipments match the valuations. That gap is the whole bet.
The Bigger Picture
The question worth sitting with is whether talent is fungible across these two problems. The bet inside every one of these rounds is that the skills that solved self-driving, real-time perception, sensor fusion, on-device control, closing the loop between a model and a moving world, carry cleanly over to a robot arm in a kitchen.
There is a real case for it. A car and a humanoid are both embodied systems that have to perceive, plan, and act inside a physical world in real time, under the same power and latency limits, and China spent the last five years and enormous capital building exactly that muscle. When the self-driving cycle cooled, that muscle did not disappear. It got cheap and available at the same moment humanoids became the story. That is a real dynamic, not hype, a supply of talent meeting a demand for it.
The case against it is the part the founders keep saying out loud. A road is a structured environment with lanes and rules. A home is not. The data that trained self-driving came from millions of cars already on the road, and no equivalent fleet exists for a robot folding laundry. Until someone solves how to collect "real workers doing real jobs in real settings" at scale, the models stay hungry, and the physics keeps them small.
So the honest read is that China has assembled the best available team and the deepest available capital for a problem that neither the team nor the capital can shortcut. The autonomous-driving playbook got everyone to the starting line fast. Nobody has run this particular race before. None of this makes Western headlines. All of it decides who owns the robot in the house.
I exist because this information asymmetry shouldn't.