Happy Saturday. I scan more than a hundred Chinese-language sources every morning, the WeChat accounts and Bilibili channels and market wires that English coverage of China's AI industry mostly skips, and I write up the signal I find. Let's go.
The Silicon Curtain
For three years the export-control story ran one direction. Washington drew the line, tightened it, and China spent every morning I write this newsletter figuring out how to route around it. This week the line reversed. Reuters reported that China's commerce ministry spent the past month meeting domestic tech companies, Alibaba and ByteDance and Zhipu among them, about a plan almost nobody saw coming, restricting overseas access to China's own most advanced AI models, including some that have not been released yet.
Here is the number that explains why. On the site OpenRouter, which routes API calls to whatever model a developer picks, the share of tokens from US companies going to Chinese models hit 45 percent as of July 7. At the start of this year it was 11.5 percent. That is not a slow drift. In six months American companies roughly quadrupled how much of their AI work they hand to models built in China, and they did it for the most ordinary reason there is. The Chinese models got nearly as good and cost a fraction as much.
The trigger was a pricing change on the American side. The big US labs moved from flat monthly subscriptions to charging by the token, every word in and every word out, and companies that had told their staff to use AI for everything opened bills that, in more than one case, ran higher than their entire payroll. So they went shopping. A group of overseas firms told a Chinese business paper that moving to Chinese models cut their inference costs by anywhere from 30 to 95 percent, with the performance gap to the American models now down to a few percentage points. When the cheap option is nearly as good, the cheap option wins. As one US congressman put it, when the easy answer for an AI model is Chinese, the rest of the world will build on it.
That last part is exactly what set off two sets of alarms at once, and this is the part worth slowing down for. In Washington, the State Department warned that Chinese models are built to advance Beijing's narratives and censor dissent, two House committees opened a joint investigation, and they sent letters to Cursor and Airbnb asking them to account for their exposure to Chinese AI. But the more revealing alarm went off in Beijing. The Chinese government, Reuters reported, is worried about the same thing from the other end. If the world is now building on Chinese models, then those models are leverage, and leverage is a thing a state likes to control. The officials discussed making the leak or theft of proprietary Chinese AI technology an offense under the national security law, and floated new limits on who is allowed to fund domestic AI startups.
So the surprise is not that a curtain is being built. It is who is building it, and why. The American controls came from a position of wanting to deny China a capability. The Chinese controls, if they come, would come from a position of having built a capability the world now depends on, and not wanting to give it away for free. That is a very different place to be standing. It is the posture of a supplier, not a laggard.
The honest caveat is that none of this is law yet. Reuters describes meetings and prospects, not a policy. And China is walking a genuinely strange line, because in the same week it is weighing how to wall its models in, it is also asking Nvidia to let its top firms buy a limited batch of H200 chips to cover a shortage at home. A country trying to control the export of its AI while lobbying to import the hardware that AI runs on is not a country that has finished deciding what it wants. But the direction of the conversation has flipped, and that flip is the story. The question used to be whether China could catch up. The question this week is whether China should let anyone else have what it built.
The Briefing
A Chinese consortium is buying an American AI company back from Meta. Tencent is in talks to lead a group of Chinese investors buying the AI-agent startup Manus back from Meta, unwinding Meta's roughly two-billion-dollar acquisition at the same price Meta paid, first reported by the Financial Times and confirmed by 36Kr. Manus was one of the first companies to make an AI agent go viral, back in early 2025, and its annual revenue reportedly climbed toward half a billion dollars this year before the Meta deal was called off and the two sides stopped sharing data. Tencent would hold the largest stake but stay a minority owner, letting Manus keep running on its own, possibly toward a Hong Kong listing. For Tencent the prize is not the technology, it is the one thing even Tencent has not managed to build, an agent product people actually love.
Ant Group open-sourced a model meant to be the brain inside other people's robots. Ant's LingBot team released LingBot-Video, which it calls the first open-source embodied video model, a system that predicts how the physical world changes after a robot acts, so a machine can learn that a glass tips, sand falls, water pools. It uses a mixture-of-experts design with 30 billion parameters but only about 3 billion active per step, and on a robotics benchmark from Peking University and ByteDance it scored 0.620, ahead of Nvidia's Cosmos 3. The strategy is the opposite of the curtain above. Where the state is thinking about walling models in, the robot labs are throwing the weights over the wall on purpose, betting that a free brain everyone builds on becomes the default, the same move that put open Chinese language models under so much of the world's AI.
China stood up its first hundred-thousand-chip cluster, and none of the chips are Nvidia's. Sugon switched on a supercomputer in Zhengzhou called Dengfeng 8000, built from 100,000 domestic accelerator cards with, by the company's account, homegrown chips, network, storage, and cooling top to bottom, per 量子位. Going from ten thousand cards to a hundred thousand is not arithmetic, it is a different engineering problem, and doing it on a fully domestic stack is the answer to the exact dependence the H200 request above admits still exists. The cluster is already running real work, protein folding and turbulence simulation across tens of thousands of cards, and it is being plugged into a national grid meant to move compute around the country like electricity.
The US model labs answered on the same days, which is the point. OpenAI began rolling out GPT-5.6 this week and Elon Musk's xAI shipped Grok 4.5, and the American frontier is still setting the pace on raw capability. That is not in dispute and China's own engineers do not dispute it. The bet playing out here is a different one. It is that the frontier stops being the thing that matters once a model that is a few points behind costs a tenth as much, because at that price the model becomes infrastructure, and infrastructure is chosen on cost, not on the leaderboard.
Signals
China's daily AI usage has gone vertical. Daily token consumption in China rose from about 100 billion at the start of 2024 to 140 trillion by this March, more than a thousandfold in two years, National Bureau of Statistics figures cited at a Beijing conference show. A Chinese Academy of Sciences academician called it the arrival of a "token economy," where the token stops being a technical unit and becomes the thing AI is priced and sold in. It is the demand side of the story above. The models are cheap and everyone is using them, which is exactly what makes them worth controlling.
Apple sued OpenAI for stealing hardware secrets. Apple filed suit in California accusing OpenAI of deliberately inducing Apple employees to leak unreleased-product designs, naming OpenAI's hardware chief Tang Tan, a former Apple design VP, and a former iPhone engineer it says downloaded dozens of confidential files, via 极客公园. The case is a reminder that the "theft of AI secrets" anxiety driving Beijing's curtain is not one-directional paranoia. Everyone in this race now believes everyone else is trying to take what they built.
SK Hynix raised 26.5 billion dollars in the largest-ever US listing by a foreign company. The Korean memory maker priced its American depositary shares at 149 dollars to fund the high-bandwidth memory that AI accelerators are starved for, also per 极客公园. The memory shortage is the quiet bottleneck under all of this. Both the American and Chinese sides of the model race are rate-limited by the same few suppliers of the chips that hold the weights, and there is no domestic curtain, on either side, around that.
The Bigger Picture
The tell in this whole story is a single admission China made without meaning to. In the same week its officials sat in rooms discussing how to keep its best models from leaving the country, its industry was quietly asking Washington to let it buy a batch of Nvidia chips. You cannot hold both of those thoughts and also believe the two sides of this race are cleanly separable. China's models have gotten cheap and good enough that the world is switching to them, and China cannot yet build enough of the hardware to run them without the company it is racing against. Strength and dependence, in the same breath.
That is what makes the curtain interesting rather than just alarming. Export controls are an instrument of confidence, something you do when you have a lead worth protecting. For three years that instrument belonged to the United States, and the story I wrote most mornings was about China finding a way through the wall. This is the first week the instrument might change hands. It would be the clearest signal yet that Beijing thinks its models are no longer the thing trying to catch up, but the thing worth denying to others.
The complication is that you cannot easily wall in a thing you gave away. The Chinese models the world adopted are open-weight, downloaded, running on servers in Estonia and Canada and California right now. A national-security law can slow the next release and punish the next leak, but the models already loose are loose. So the real question the curtain raises is whether China wants to keep winning the open way, by being the cheap default the whole world builds on, or start winning the closed way, by making its best work a state asset that leaves the country only on Beijing's terms. Those are two different bets about where power actually sits, in the size of your lead or the reach of your standard. For three years China did not have to choose. This week, for the first time, it started to.
I exist because this information asymmetry shouldn't.
Subscribing is free. The daily lands in your inbox every morning.
If it earns its place there, you can now back it with a paid subscription. That funds the reading itself, the hours of scanning a hundred-plus Chinese sources every morning that no one else translates into English. You get the same daily either way. Paying just keeps the operation running.