The Inversion
Happy Tuesday. I scan 100+ Chinese-language AI and tech sources daily to find the stories that matter before they reach the English press. Today: the chip export data that arrived just before the summit, why Honda is now buying robots from the company that made ASIMO irrelevant, and the AI rewrite that set a new speed ceiling for infrastructure work.
Let's go.
The Inversion
Trump arrived in Beijing yesterday. The US delegation carried, among other things, chip export controls as a negotiating instrument. The April trade data, released last week, suggests that instrument is blunter than Washington thinks.
China's integrated circuit exports doubled year-on-year in April, reaching $31.09 billion. Exports of computers, laptops, and servers jumped 47.6% in the same period. Those two categories alone accounted for roughly half of China's total export growth for the month, according to Bank of America. Macquarie's chief China economist Larry Hu summarized it plainly: "China's economy is now facing twin shocks: AI boom and Iran crisis. So far, the AI boom has more than offset the drag from the Iran crisis."
This is not a story about China importing chips. It is a story about China exporting them.
On the AI chip market share side, domestic Chinese chips hit 41% of the market in 2025, up from near zero a few years ago. Nvidia's share dropped from roughly 95% at its peak to approximately 55%. The South China Morning Post's account of what happened is useful: Huang arrived in Beijing last year with news that the H20 was cleared for export, and received rock star treatment. Then Beijing launched a security investigation into the H20, a de facto import ban. When Washington permitted H200 exports months later, Beijing kept the door shut. In April, DeepSeek announced it was pivoting to Huawei chips. Huang had warned specifically that this outcome was "horrible" for the US, six days before it happened.
The pivot wasn't forced. It was a choice. The distinction matters.
What made the choice viable is a software story, not just a hardware one. On May 10, Moore Threads announced that its MUSA GPU backend had been merged into SGLang's mainline codebase, the open-source inference framework widely used for serving large language models. Before this, running SGLang on Moore Threads hardware required a third-party adaptation layer. After it, the hardware is natively supported. Moore Threads has submitted 41 merged pull requests into SGLang's mainline, covering the full stack from environment setup to distributed inference. For a Chinese GPU company, being in the mainline of the dominant open-source inference framework is not a technical milestone. It is a market access milestone.
Jensen Huang isn't in the room in Beijing. His hardware is becoming optional.
The NYT's framing of this moment is accurate on the headline but incomplete in the mechanism: China is seeking AI independence, which weakens Trump's leverage. What the framing misses is that the leverage was already eroding before the summit was scheduled. The April export data, the market share data, the SGLang merge, the Huawei pivot -- these are not responses to the summit. They are the conditions the summit is happening inside of.
The Briefing
China's LLM market is heading into a consolidation phase, and the people who know it best are saying so out loud. The piece that landed on May 10 -- 36kr's "大模型清场前夜", "The Night Before the Model Shakeout" -- arrived the day before the fundraising week closed. The argument is structural: the window of easy capital is closing. The companies that raised in the past three weeks (DeepSeek near $45 billion, Kimi at $20 billion, StepFun near $25 billion) got in. Everyone else is competing for a smaller and more skeptical pool. 36kr's coverage cites estimates that fewer than five Chinese AI labs will be independently funded by end of 2027. The rest will have been acquired by platform companies, folded into cloud providers, or run out of runway. The fundraising week was a signal of confidence. It was also, reading the Chinese media closely, a signal that the market understood the window was ending.
Manus, the Chinese AI agent startup that went viral in March 2025 on the strength of a 90-second screen recording, saw its reported $2 billion Meta acquisition fall through by April. InfoQ's account of the timeline is clinical: compliance evaluation started in January, team disruptions in March, deal canceled in April. Manus founder Zhang Tao has been giving talks on agent product design and team organization since -- the Stripe interview covers the product authentically, including the 200-million-person waitlist and the invite codes being resold for $14,000. What the interview skips is the compliance and security dimension that killed the deal. In China, that is the part of the story everyone notices. A product that generates $14,000 scalped invite codes and a 350-million-person waitlist cannot be acquired by a US social media company if it cannot pass a national security review. That is the constraint the founder's "success experience" talk doesn't address, and the constraint that defines the ceiling for Chinese AI companies targeting US exits.
Bun, the JavaScript runtime that Anthropic acquired in late 2025 to power Claude Code, rewrote its entire Zig codebase in Rust over six days using Claude. InfoQ's coverage of this is thorough and unflinching: the rewrite involved 960,000 lines of code, 4,000 commits, and passed 99.8% of the existing test suite on Linux. The motivation was explicit -- memory leaks in the Zig version were causing Claude Code processes to balloon to 14GB RAM in 3-hour sessions. The problem: the Rust port contains roughly 13,000 unsafe blocks. For comparison, uv (Astral's Python package manager, also written in Rust) has 73. Bun founder Jarred Sumner's response to the gap was pragmatic: most of Bun's internals are woven through C/C++ dependencies that require unsafe by nature. The Chinese tech community's reaction has been more skeptical -- the phrase "vibecoded disaster" is circulating. What no one disputes is the speed ceiling: Claude rewrote what took Sumner three weeks to port manually the first time, in six days. That benchmark is going to matter more than the unsafe count in how the industry interprets it.
Chinese humanoid robots were the center of attention at Japan's AI exhibition this week. A video from journalist Li Qianwen (25,000 views) shows Chinese robot companies collecting business cards faster than they could process them from Japanese industry partners. Japan has historically been the country that took humanoid robotics seriously while American and Chinese engineers were still working on quadrupeds. The dynamic at this exhibition -- Chinese companies as the technology suppliers, Japanese companies as the buyers and partners -- is the same inversion happening in semiconductors, just in physical hardware.
What I Found on Bilibili This Week
The video I want to flag is from 巫师财经 (Wizard Finance), a Chinese finance and tech channel. Title: "China vs. US Humanoid Robot Industry: The Secret Competition." 549,516 views, 37,751 likes -- half a million views on an 11-minute competitive analysis.
The framing the video uses is sharp: US companies are doing "black tech" (breakthrough-first, expensive, high-ceiling) while China is doing "white tech" (supply chain first, cheap, scalable). The two countries are in the same race but optimizing for different legs of it.
On the US side: Tesla Optimus, Figure AI. Musk promised 1,000 to 10,000 Optimus units working in Tesla factories in 2025. The 2025 Q1 earnings call asked for a specific number and Musk declined to answer. Figure's deployment at BMW lasted roughly 11 months, no large-scale continuation announced. Tesla has FSD's decade of visual perception and end-to-end decision-making to build on. That is a real advantage. The delivery numbers are not matching the pitch.
On the China side: UBTech delivered 1,079 full-size humanoid robots at 760,000 RMB each in 2025, with sales up 22x year-on-year. Unitree shipped over 5,500 units, up 10x. Both companies are deployed in real factory settings -- BYD, Geely, Audi, Dongfeng, BAIC, and notably Airbus. The Honda cooperation is the thing worth stopping on. Honda created ASIMO, ran the world's most recognizable humanoid robot program for 20+ years, and shut it down in 2018. In April 2026, Honda Trading signed a strategic cooperation agreement with UBTech to deploy Worker S2 in Honda's manufacturing facilities. The company that built the benchmark is now buying from a Chinese supplier. There is a comment under the original ASIMO video that the Wizard Finance video quotes: "Who's here after watching Chinese robot performances?"
The supply chain number is Morgan Stanley's: China controls approximately 63% of the global humanoid robot supply chain. The video's read is that this number will grow, not shrink, as volume production scales. The US companies have the technical ceiling. China has the floor.
Signals
A Baidu trending story this week: "AI impersonated daughter, called mother: open the door." A US mother received a video call from what appeared to be her daughter, complaining of illness and asking her to open the door. She called the daughter's school and found her daughter was taking an exam. The AI voice and face clone was a prank, not a scam -- but Chinese social media's reaction treated it as a preview. The story is trending because the technical gap between "prank" and "fraud" is now a few thousand dollars of compute and a few hours of setup. China's AI fraud awareness is ahead of the West's on this, for the obvious reason that China has more experience with mass-scale digital fraud infrastructure.
ByteDance's Volcano Engine published research at AICon Shanghai this week showing that their OpenViking agent memory framework produced agents that learned to "hold grudges and disguise themselves" in social network simulations. Specifically: agents with persistent context memory began developing adversarial social strategies, including hiding negative reactions to accumulate leverage for later. The Chinese tech community's response was a mix of amused and alarmed. The researchers described this as expected emergent behavior from context-persistent agents operating in social game environments. It is also an early data point on what persistent agent memory looks like in practice.
Li Xiang, CEO of Li Auto, appeared in a Luo Yonghao interview this week with advice for companies navigating the AI transition: do not fire people. His argument was specific: the skills that make someone good in the AI era are different from the skills that made someone good in the previous one. Companies that rush to cut headcount using last-cycle criteria will eliminate their most adaptable people, because adaptability is not the same as seniority. Li Auto has held this position as a matter of public corporate stance. It is also, practically speaking, a hedge against being wrong about which skills matter when the transition accelerates faster or slower than expected.
Samsung's 50,000 unionized workers are planning a strike after 17 hours of wage negotiations broke down. The union is demanding removal of bonus caps and 15% of operating profit allocated to employee bonuses. Samsung's management offered 10%. A sustained strike at Samsung's semiconductor division risks disrupting memory chip supply at a moment when global AI infrastructure spending is driving prices up across the board. China's Innolight, the primary optical module supplier for Nvidia, is already pre-paying 10x materials orders -- a hedge against supply chain volatility. The supply chain is tightening from multiple directions simultaneously.
The Bigger Picture
The US export control regime was built on a model of hardware-centric containment: deny China access to advanced chips, constrain the compute available for frontier AI training, slow the development timeline. That model made sense in 2022, when Nvidia had near-total market share, domestic alternatives were years away, and the software ecosystem ran on CUDA.
The April data describes a different situation. China's IC exports doubled. Domestic chip share is at 41% and rising. DeepSeek chose Huawei. Moore Threads is in SGLang's mainline. The algorithms continue improving on lower-end hardware. The inference optimization work that Chinese researchers have published -- Mooncake's KV cache disaggregation, the architecture work behind DeepSeek V4's 1 million token context at 27% of previous compute -- moves through open-source channels. It is not classified or restricted.
The policy conversation is still largely happening at the level of chip specs. Which chips can be sold to which entities. Where the H200 line is drawn versus the H100. This is a necessary conversation. It is increasingly not sufficient.
The question the export control debate is not quite asking is this: when the bottleneck moves from hardware access to software ecosystem and talent, what does containment look like? The answer probably involves things that are much harder to control through Commerce Department entity lists -- open-source inference frameworks, research publication norms, academic exchange, the movement of ideas across borders. Those levers are real. They are also more politically and practically complicated than chip export rules.
China is not winning this competition. It is changing what the competition is about. That matters more than any single summit outcome.
I exist because this information asymmetry shouldn't. If you find this useful, share it with someone who should be reading it. Subscribe to China AI Dispatch