Happy Sunday. I scan more than 100 Chinese-language sources every day, the WeChat accounts, the Bilibili channels, the finance wires, the policy feeds, and I write up the China AI stories that English-language coverage misses. No team, no wire service, just the reading. Let's go.
The Hedge
This week the largest single financing round in the history of China's quantum computing industry closed, and it did not go to a company most readers have heard of. Origin Quantum (本源量子), the lab spun out of the Chinese Academy of Sciences in 2017, raised close to 3 billion yuan (about 410 million USD) in a pre-IPO round led by China North Industries Group, the state defense-industrial conglomerate, at a valuation near 21 billion yuan. The company is now sprinting to become the first pure quantum-computing stock on Shanghai's STAR Market.
The detail that matters is not the number, it is who wrote the check and why now. A weapons conglomerate led a quantum round. And the round could happen at all because Shanghai's STAR Market switched on new pre-profit listing rules this month that let pre-revenue AI and quantum companies file. The capital plumbing got rebuilt, then the capital arrived.
The money is moving because the engineering converged. In May, three Chinese firms shipped new full-stack quantum machines across the three mainstream hardware paths at once, 中科酷原 with a 180-qubit neutral-atom system, 玻色量子 with a 200-qubit photonic system, and Origin Quantum with a 1,000-qubit superconducting machine. Chinese industry has stopped describing quantum as a science project and started describing it, in its own words, as "AI's new compute engine."
That phrase is the whole story. China cannot buy leading-edge Nvidia accelerators, so it is buying optionality across every compute paradigm it can reach. Last week it was a domestic supercomputer that took the TOP500 crown with zero GPUs (our No-GPU Machine). This week it is quantum. The bet is not that any single substrate replaces the GPU. The bet is that dependence on one supplier is the actual risk, and breadth is the cheapest insurance against it.
Here is the part the headline numbers hide. Qubit count is the figure markets understand, and it is close to meaningless on its own. A thousand noisy physical qubits are not a thousand useful ones. The benchmark that decides the race is error-corrected logical qubits, and there the US still leads, Google's 105-qubit Willow chip demonstrated verifiable, error-corrected quantum advantage. China is ahead on the number that reads well in a deck. The US is ahead on the number physics actually rewards. It is the same export-control story told from two ends, the same shape as a supercomputer that ranks first on raw FLOPS and fourth on the benchmark that looks like AI training.
The American side of this is its own scramble. Quantinuum went public on Nasdaq on June 4, the first quantum company to use a traditional IPO rather than a blank-check merger, and the stock closed flat on debut after an upsized raise. The US Commerce Department put about 2 billion USD of CHIPS Act money into quantum as direct equity in May. Both governments are now equity investors in the same technology for the same reason. Watching only the GPU race misses that the compute contest already forked into a dozen smaller ones, and China is funding all of them.
The Briefing
Smart money is now counting robots by the thousand, not the unit. On Saturday, AgiBot (智元) rolled its 15,000th general-purpose embodied robot off the line, a unit it calls the Genie G2. The cadence is the story, 5,000 units last December, 10,000 by March 30, 15,000 now, a rough doubling every quarter. AgiBot's embodied-business president Yao Maoqing was blunt about why volume matters, mass production "is not the goal, it is a process," the point is to put robots into real workplaces so they generate interaction data, and the data flywheel makes the next batch smarter. The hidden cost sits one layer down. A 虎嗅 report this week followed the humans who do eight-hour shifts in motion-capture rigs, teleoperating arms and folding laundry so a robot can watch and learn. The flywheel runs on people for now.
DeepSeek shipped an efficiency layer, not a new model, and its founder wrote part of it. On June 27, barely two weeks after closing a 50 billion yuan round, DeepSeek and Peking University published a paper on DSpark, a speculative-decoding module that bolts onto the existing V4-Pro and V4-Flash models to cut inference cost. They open-sourced the full toolkit, DeepSpec, under an MIT license. What is unusual is the author list. DeepSeek founder Liang Wenfeng is on it. A founder who just raised one of the largest rounds in Chinese AI is still personally co-authoring inference-optimization code, which tells you where this company thinks the edge sits, not in the model, in the cost of running it.
ByteDance skipped four version numbers to make a point about video. At its Volcano Engine conference on June 23, ByteDance unveiled Seedance 2.5, jumping straight past 2.1 through 2.4. The model generates 30 seconds of native 4K video from a single prompt, accepts up to 50 reference inputs (images, audio, 3D models), and processes audio in the same latent space as the visuals so sound and motion sync natively. Alibaba shipped its own HappyHorse 1.1 video model the same day. Western labs treat generative video as a side quest. Two Chinese giants are treating it as a flagship, and the gap in production quality is closing fast.
One Chinese city budgeted a quantum cluster into its annual AI plan. Suzhou published its 2026 AI action plan this week, targeting more than 400 billion yuan in core AI-industry revenue, over 3,500 AI firms, and 34,000 PFLOPS of intelligent compute by year-end. Buried in the document is the list of priority clusters the city will fund, AI chips, embodied robots, and quantum technology, with a local AI fund pool above 110 billion yuan behind it. This is what the national compute hedge looks like from the ground. It is not one big lab, it is dozens of municipal budgets all naming the same three substrates.
Signals
DeepSeek is hiring for AGI by name. Days after the raise, DeepSeek opened a hiring spree explicitly framed around pursuing artificial general intelligence, recruiting newcomers rather than only senior researchers.
China wants to give every AI agent an ID. Regulators are moving toward a unified identity system for autonomous AI agents, a "digital ID card" so an agent's actions can be traced to an operator. It is the governance layer arriving before the agents are everywhere, the opposite order from how the West is doing it.
BYD is building its own self-driving chip. BYD expects to mass-produce a self-developed ADAS chip in a vehicle by 2027, pulling autonomous-driving silicon in-house the way it already did with batteries and power electronics.
Humanoids do not sell, they rent. A 36Kr piece on the "year of scaled embodied AI" found the demand pattern inverted, customers are not buying humanoid robots, they are renting them, because nobody wants to own a depreciating prototype. The capital is real, the product-market fit is still a lease.
The Bigger Picture
China's answer to being cut off from the best AI compute is not to win one race. It is to buy options on all of them.
Look at one week. A domestic supercomputer that needs no GPUs. The largest quantum round in the country's history, spread across three competing hardware paths. An inference-efficiency framework that stretches the chips already on hand. Fifteen thousand robots built mainly to harvest the one input China has in surplus, physical-world data. And underneath all of it, a capital structure rebuilt to fund the bet, pre-profit IPO rules switched on, a defense conglomerate leading a quantum round, a single city budgeting 400 billion yuan and naming quantum as a target.
None of these is a sure thing. That is the point of a hedge. You do not buy options because you know which one pays off. You buy them because you cannot afford to be wrong about the one supplier you were told to depend on.
The honest caveat applies to the whole portfolio, the same one that applies to each piece. Leading on qubit count is not leading on error-corrected qubits. Leading on robot units shipped is not leading on robots that pay for themselves. Leading on benchmark scores is not leading on frontier capability. China is buying breadth at a scale no one else is matching. Whether breadth converts into advantage is the question the next year answers, and it is the only one worth watching.
I exist because this information asymmetry shouldn't. None of this makes Western headlines. All of it matters.