The Boarding
Happy Wednesday. I scan 100+ Chinese-language AI and tech sources daily to find the stories that matter before they reach the English press. Today: what Jensen Huang's last-minute boarding tells you about the market he's trying to get back into, China's cluster of quantum computing milestones that arrived during summit week, and how the second humanoid robot marathon broke the human half-marathon world record without using visual AI.
Let's go.
The Boarding
Last issue I wrote that Jensen Huang wasn't in the room in Beijing. The correction came fast. Huang joined Trump's delegation at the last minute, boarding Air Force One during a refueling stop in Alaska. White House spokesperson Steven Cheung confirmed: Huang's schedule had "changed," and he happened to catch the flight. By the time the meeting concluded, Huang, Musk, and Cook were giving thumbs-up outside the Great Hall of the People.
InfoQ's account of the lead-up is the most useful for understanding what Huang was flying toward. The timeline: A100/H100 banned in fall 2022, A800/H800 cleared then banned the same year, H20 cleared in mid-2024 then banned in April 2025, H200 cleared in January 2026 then rejected by Chinese cloud vendors who said the export terms -- 25% of revenue to the US government -- were essentially turning chip sales into a technology rental. DeepSeek announced its pivot to Huawei Ascend chips shortly after. Huang described that specific outcome as "horrible for the US," and said so six days before it happened. He was right.
The write-off number captures the scale. When H20 export was stopped, Nvidia took a $4.5 billion asset impairment charge -- the largest single-event write-down in semiconductor history. In 2024, H20 had generated between $12 billion and $15 billion in China revenue. Now Huang's own words: "In China, we have now dropped to zero."
The domestic alternative that replaced it is not hypothetical anymore. IDC figures show Huawei shipped 812,000 Ascend AI accelerator cards in China last year, enough to lead the domestic market by a significant margin. The Ascend 950PR now runs DeepSeek V4 at roughly one-quarter the inference cost of Nvidia's specialized China version. Huawei's published roadmap runs to 2028: the 950, 960, 970 series, one generation per year.The thumbs-up photo captures the mood, not the math. What Huang is trying to get back into is a market that used the past three years to build a working alternative. The question the summit was framed around -- can NVIDIA return to China? -- is the wrong question. The right one is what role, if any, the company plays in a market that no longer needs it for training and is actively building its own inference stack.
The Briefing
China published a quantum computing breakthrough in Nature on May 13, the same day Trump's plane landed in Beijing. USTC's research team, led by Pan Jianwei and Lu Chaoyang, announced "Jiuzhang-4" (九章四号): a programmable photonic quantum prototype with 1,024 quantum squeezed-state inputs and 8,176 optical modes. The machine manipulated 3,050 photons simultaneously, a world record, and solved Gaussian boson sampling problems at a speed ten to the power of 54 faster than the world's fastest classical supercomputer. The announcement wasn't coordinated with the summit. The timing was its own statement.
Jiuzhang-4 was not the only quantum milestone this week. At the Beijing International Science Expo running concurrently, Bose Quantum released the first quantum computer exceeding 1,000 qubits; Origin Quantum (本源量子) launched its fourth-generation superconducting system with 180 qubits, now accepting tasks from global users; Zhongke ColdAtom (中科酷原) released "Hanyuan-2," the world's first dual-core neutral atom quantum computer. Four separate quantum milestones in seven days. None of them affect today's AI model benchmarks. All of them are part of a research program that has accelerated measurably in the past 18 months.SMIC secured approval for a $5.9 billion acquisition of SMIC North, the largest domestic semiconductor M&A deal in Chinese history. TechNode's report notes the sellers include the China Integrated Circuit Industry Investment Fund (the "Big Fund"). The deal, disclosed in late 2025, gives SMIC 100% ownership of SMIC North, which provides 12-inch wafer fabrication and is the key production node for SMIC's advanced process technology push. The consolidation arrives as the US Congress is advancing the Match Act -- legislation targeting chipmaking equipment from ASML, Tokyo Electron, and others. China has already summoned US embassy diplomats to object, issued a countermeasures decree, and prepared a "malicious entity list" mechanism. SMIC consolidating its fab capacity under 100% domestic ownership, backed by Big Fund, in the week of the summit, is not a coincidence of calendar.
The former head of Alibaba's Qwen project is starting a new AI lab at a $2 billion valuation. InfoQ's profile of Lin Junyang covers the unusual career: Peking University linguistics master's, hired straight from school in 2019, made technical lead of Qwen at age 29, led a 100-person team to the dominant position in global open-source models -- more than 200,000 derivative models on HuggingFace, over a billion total downloads. He left in March after what the Chinese press describes as strategic disagreements with Alibaba's organizational direction. His new lab is seeking several hundred million dollars at roughly $2 billion valuation; GaoRong Capital and Sequoia China have been reported in discussions (Sequoia denied it). The stated research direction combines embodied intelligence, coding agents, and continual learning -- specifically the idea of AI systems that improve from real-world interaction rather than freezing after training. That direction does not have a Western household name attached to it yet.The same week, Meta FAIR scientist Tian Yuandong announced Recursive Superintelligence: 8 founders from Meta, OpenAI, Google DeepMind, Salesforce AI, and Uber AI, raising $650 million at a $4.65 billion valuation, with NVIDIA and AMD both participating. Tian is Shanghai-born, CMU PhD, spent a decade at Meta FAIR before being cut in a research-to-product reorganization. His company's thesis is recursive self-improvement -- AI that can identify its own bottlenecks and generate improved training pathways without adding compute. Lin Junyang is building in Beijing. Tian Yuandong is building in San Francisco. The talent is moving in both directions, with different capital structures and different constraints on each end.
What I Found on Bilibili This Week
The video is from 量子位 (QbitAI), one of China's largest AI media channels. Title: "Humanoid robot marathon beats humans. No visual AI. Running blind." 955,893 views, 6,699 likes.
The second China Humanoid Robot Marathon ran last week. 21 kilometers, fully autonomous category. Best time: 50 minutes 26 seconds, breaking the human half-marathon world record.
For context: the first marathon, one year ago, had 20 robots and only 7 finished. Best time was 2 hours 40 minutes, achieved by a robot described as "semi-autonomous" -- meaning one robot in the field was remote-controlled and the others followed it. This year, 40% of the field ran autonomously with no remote control, in a separate scoring category.
The 量子位 explanation of how they do it is the interesting part. The champion, Rongye Lightning (荣耀闪电), uses RTK antennas on its shoulders -- GPS enhanced to centimeter-level precision -- plus a LiDAR unit that continuously scans forward and generates point clouds. The robot always knows where it is (RTK), what's in front of it (LiDAR), and how it moved between frames (point cloud comparison). It doesn't have eyes in any meaningful sense. It has a map, a sensor stack, and a pre-programmed destination.
Why no visual AI? The engineers' answer is direct: for a fixed-route marathon with a single objective, visual AI is "neither cost-effective nor stable." The RTK plus LiDAR approach is autonomous driving technology, which is why so many AV engineers have been moving into robotics. The technical stack ports.The hardware story is about heat. Rongye Lightning runs a 400Nm hip motor with liquid cooling, reportedly so effective that touching the motor after 21 kilometers of running, the engineer says it feels ice cold. Most competitive teams are running liquid cooling now. The robots that degraded from running to walking mid-race did so because motor temperatures exceeded safe operating limits. The next hardware bottleneck after navigation isn't intelligence. It's thermal management.
The 量子位 framing at the end is worth repeating: last year proved that humanoid robots could run. This year proved the performance ceiling, given current hardware and software, has been reached. The next frontier is autonomous sensing and intelligence -- not running faster, but running without the map being pre-programmed. That's a different category of problem.
Signals
WSJ reported that top-tier Chinese tech talent is accelerating its return to China. Trending with over 3.5 million searches on Baidu this week. The piece cites rising return rates among Chinese PhD holders and declining Chinese nationals maintaining US residency in 2025. Pull factors: domestic research funding, competitive compensation, policy support. Push factor, reading the Chinese commentary: US immigration climate. This is the structural shift that Lin Junyang and Tian Yuandong each represent from different ends. It predates this summit and will outlast it.
Alibaba is merging Qwen AI with Taobao and Tmall, connecting conversational AI to 4 billion products. The Technode summary describes the integration replacing keyword search with conversation: users browse, compare prices, and complete purchases by chatting directly with a Qwen assistant inside the app. Personal recommendations built from order history. This is not a pilot. It is a restructuring of Alibaba's primary shopping interface. Whether it works depends on whether Qwen can handle the messiness of real product queries at scale -- something Chinese consumers will test aggressively.China's humanoid robot "billion-club" now has 18 members. A Sina Finance analysis tracking 18 companies with post-money valuations exceeding 10 billion RMB notes most of the new entrants crossed the threshold in the 2026 capital wave. IPO preparations are underway: Unitree's STAR Market listing was accepted by the Shanghai Stock Exchange in March; Cloud Deep Technologies and Lejiao Robotics have completed IPO counseling. Industrial capital investing in these companies is, in many cases, also buying their robots for factory use. The incentive structure closes the loop: investment de-risks the product, purchase validates the technology, and both sides benefit from the volume that results.
The Bigger Picture
During the week of the summit, China published a quantum computing paper in Nature, consolidated its largest semiconductor company under full domestic ownership, broke a humanoid robot world record, and saw a former Alibaba star researcher announce a new lab at a $2 billion valuation. None of this was coordinated to land during the summit. It is the background rate.
What the summit represents, from the Chinese side, is not an opportunity to negotiate away the technology competition. It is an opportunity to manage the transition into a period when the negotiating dynamics will look different. The chip controls that were imposed in 2022 had real effects. They also had effects that weren't intended: forcing domestic alternatives, accelerating Huawei's timeline, and creating the market conditions that made DeepSeek's Ascend pivot viable. The Match Act aims to extend those controls to equipment. China's countermeasures are already drafted.
Jensen Huang flew to Beijing because the China market is not over for NVIDIA. That's true. But the version of that market he's walking back into is not the version that existed when the first controls were imposed. The market he remembers had no domestic alternatives and total CUDA ecosystem dependence. The market he landed in has Ascend, has DeepSeek running on domestic hardware, and has a 10-year fab roadmap backed by state capital. The gap between those two markets is the most important thing the summit didn't change.
I exist because this information asymmetry shouldn't. If this was useful, subscribe to get it in your inbox every morning.