Happy Wednesday. I scan more than a hundred Chinese-language sources every morning, the ones no one translates into English, and write up what actually matters. Let's go.
The Tell
The most-watched AI company in China spent two years telling everyone it did not need better chips. It needed fewer of them, used better. That was the whole DeepSeek story. R1 trained on a few million dollars of compute, a mixture-of-experts design that only lights up a fraction of its parameters per token, an efficiency edge so sharp it wiped a trillion dollars off Nvidia in a single week last year. So the news that broke in Beijing late Monday night reads like a contradiction. DeepSeek is secretly building its own chip.
Here is the part that tells you what it means. The chip is for inference, not training. DeepSeek is not trying to build a rival to the Nvidia clusters that train frontier models, the hardest and most capital-hungry problem in the field. It is going after the calculation that happens after training, every time a user asks a question and the model writes an answer. That sounds smaller. It is not. Inference runs for the entire life of a model, and by most estimates it is 80 to 90 percent of what a model costs to run over its lifetime. A general-purpose Nvidia GPU carries a lot of circuitry it does not need for that job. A chip built for one company's own models, its own attention mechanism, its own low-precision math, can strip all of it out.
That is the tell. You do not build an inference chip to out-Nvidia Nvidia. You build one because you already know exactly what your models look like and you want to own the part of the compute bill that actually recurs. DeepSeek has been laying this track in plain sight. Its DSpark release two weeks ago, which we covered here, sped up inference on its V4 models by 60 to 85 percent in software. The UE8M0 FP8 number format it shipped in V3.1 was written, by its own engineers' account, to match the hardware traits of next-generation Chinese chips. The algorithm team was thinking about silicon while it wrote the model. Now it is designing the silicon.
The details are the giveaway on how serious this is. According to three people briefed on it, the project started roughly a year ago and DeepSeek is already talking to chip-design houses, a foundry, and memory suppliers. It has been hiring chip engineers for months without posting a single job publicly, entirely through private channels. And in June the company took its first outside money ever, about 51 billion yuan, roughly 7.4 billion dollars, after years of refusing investors on principle. The stated uses were domestic-chip compute centers, a self-designed AI chip, and talent. The raise and the chip are the same decision.
The honest part is that none of this is close. A competitive AI chip takes years and enormous money to move from design to a working part to volume, and this one is early. DeepSeek still trains on other people's hardware, Nvidia H800s at the start and Huawei Ascend since, and it will for a long time. The chip does not replace Nvidia tomorrow, or Huawei. But the direction is not ambiguous, and the market read it fast. Nvidia slipped about 1.6 percent before the bell on the report, on top of a slide that has taken roughly a trillion dollars off its value in under two months. One analyst put the trend bluntly, that Nvidia's share of the China market is drifting toward zero and staying there.
The Briefing
DeepSeek is not the only lab that decided this week to become a chip company. The same night, The Information reported that Zhipu, the Beijing lab behind the GLM open models, is weighing its own custom chip as demand outruns the compute it can legally buy. The trigger is concrete. Token usage of GLM-5.2 on the developer platform Vercel jumped 27 times in a single week after the model shipped. Zhipu cannot buy Nvidia's best parts, so a surge that any US lab would meet by swiping a credit card becomes, in China, a hardware problem you have to go solve in a fab. Two of the country's strongest labs reaching the same conclusion in the same week is the signal, not either one alone.
Zhipu also aimed straight at Anthropic's core business. It released a coding harness for GLM-5.2 and rolled out promotions to pull developers off Claude, the same week Anthropic was answering claims it had covertly tracked Chinese users. The pattern across both stories is a Chinese lab pushing into the exact seat, custom silicon and agentic coding, that the leading US labs treat as their moat.
The domestic chip these labs would lean on is already shipping in volume. MetaX, one of China's homegrown GPU makers, said this week its order book runs into next year and beyond and its MXC600 flagship, a training-and-inference part built end to end on a domestic supply chain, from design through manufacturing and packaging to the software stack, is now in mass shipment. It cleared the national security-and-reliability review in May. Huawei still holds about half of the roughly 50-billion-dollar China AI chip market, but the field behind it, MetaX, Alibaba's and Baidu's in-house parts, and now the labs themselves, is filling in fast.
A chip startup came out of stealth betting on a way around the rules rather than through them. Dongfang Suanxin, run by Wei Shaojun, a vice-president of the China Semiconductor Industry Association, exited stealth with a design that leans on 3D stacking to get performance the export controls deny China at the transistor level. Locked out of TSMC's most advanced nodes, Chinese designers are trying to win back at the packaging and architecture layer what they cannot buy at the process layer.
American companies are quietly running Chinese models to escape their own vendors' prices. As OpenAI and Anthropic raise costs, US firms are moving real workloads onto DeepSeek, Qwen, and GLM, cheap, open, and good enough that the savings win the argument. US lawmakers have started probing the trend. That adoption is the backdrop for everything above, and for the next section.
Signals
China is thinking about walling off the AI it just spent two years opening up. Chinese officials met Alibaba, ByteDance, and Zhipu over the past month to discuss limiting overseas access to the country's top models, open and proprietary, including ones not yet released, and possibly restricting who can fund Chinese AI startups. The country that was on the receiving end of export controls is now sketching its own.
China stood up its first national standards body for humanoids and embodied AI. The committee was formally established in Beijing's Yizhuang district alongside a 2026 standards framework, the kind of quiet plumbing that comes before an industry scales, not after.
The memory cycle behind all of this compute keeps running hot. Samsung's profit this year is on track to top its previous 40 years combined, JP Morgan is reiterating overweight on the stock, and SK Hynix is set to list on the Nasdaq this Friday. Every inference chip these labs design still has to be fed by high-bandwidth memory, and that market is in the tightest stretch it has seen.
DeepSeek V4 is due next week. Chinese coverage points to a trillion-parameter model with native multimodal support arriving in days. If the chip is the long game, V4 is the near one, and both are aimed at the same target, running frontier models cheaply on hardware China can actually get.
The Bigger Picture
Step back and the week tells one story from three angles. China's models got cheap and good enough that US companies now run them to save money. That success created two reflexes at once. Beijing started drawing up controls to keep its best AI from flowing out freely, and the labs started building their own silicon to keep their compute from depending on hardware they cannot reliably buy. Openness pulled the world in, and the pull is now producing walls, one on the model, one on the chip.
The chip reflex is the more durable of the two. Export controls can be loosened or dodged. A working inference chip, co-designed with the models it runs, is a structural change in who controls the cost of AI. For fifteen years the assumption was that the frontier belonged to whoever owned the most advanced fabrication, and that China, cut off from the top nodes, could not close that gap. DeepSeek and Zhipu are betting on a different route, that if you know your own model well enough, you can design a narrow chip that beats a general one on the only workload that matters at scale, and have it built in a domestic fab that never touches a US export list.
If that bet lands, the interesting number stops being how many Nvidia chips China can smuggle or make. It becomes how cheaply China can run inference at all, on hardware entirely of its own. That is a harder thing to sanction, because there is nothing at the border to stop. It is a company, its models, and its own chips, on its own soil, serving the world. The tell was that they started with inference. It means they are not chasing Nvidia. They are routing around it.
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.