Happy Tuesday. I read more than a hundred Chinese-language sources every morning, the WeChat accounts, Bilibili, the finance wires, the trade press, and I write up the parts that never make it into English. Let's go.
The Trade-Down
Chinese models now run a large share of the world's actual AI work, and the companies switching to them are American. A new IDC survey found that 47% of decision-makers at US firms with more than a thousand staff have put a Chinese model into at least one use case, and one in five say they use one heavily. At a point mid-July, all five of the top models on OpenRouter, the marketplace that routes enterprise traffic to whichever model wins on price and performance, were Chinese.
These are not experiments in a lab. Coinbase told staff to move work onto Kimi and Z.ai's GLM, and its chief executive Brian Armstrong says the company halved its AI spending. DoorDash routes what its CTO calls lower-level work to Kimi for better quality at cheaper cost. Airbnb runs customer service on Alibaba's Qwen. Cursor, the coding tool being bought for around 60 billion dollars, built its Composer 2 model on Kimi foundations. The pattern is the same everywhere. Keep the expensive American model for the hard 10%, and hand the other 90% to something from Hangzhou or Beijing that costs a tenth as much.
The gap is not a temporary price cut. UBS estimates the leading Chinese models cost roughly a tenth as much to train as comparable American systems, and their API prices sit at 10 to 20% of the foreign alternative. They get there across the whole stack. Smaller parameter counts, mixture-of-experts designs that fire under 10% of the model on any given task where some US models fire 15 to 30%, GPU utilization above 70% against an industry norm near 40%, and cheaper power. Even at those prices the Chinese providers hold gross margins of 20 to 40%. This is the part English coverage keeps missing. The models are cheap because they were engineered to be, not because someone is subsidizing them.
The numbers a developer actually sees are brutal. On a legal benchmark, Anthropic's Fable averaged 31 dollars a task against 2 dollars 40 for GLM-5.2, a thirteen-fold spread. DeepSeek's new V4-Pro runs about 87 cents for three-quarters of a million words where Fable charges 50 dollars. Cursor's own engineers coined a word for the new discipline, tokenomics, and one of them put it plainly. Using the strongest model for daily tasks is like driving a Lamborghini to the grocery store to buy milk. The reason this matters beyond any single invoice is the flywheel underneath OpenAI and Anthropic. Frontier research is funded by application-layer revenue, and the application layer is the exact thing trading down.
Before anyone in Beijing celebrates, read the Chinese commentary, because it is not triumphant. Caixin ran a piece this week arguing that Chinese AI cannot escape what it calls the heavy-industrialization of compute. Kimi K3 went viral and immediately ran into a compute crunch, and the cheap-AI story collapsed the moment demand met the actual cost of serving 2.8 trillion parameters. A model is no longer a few lines of elegant code. It is chips, a power grid, high-speed interconnect, and a great deal of capital. Everyone selling intelligence at cost has the same problem, which is that selling at cost is not a business. The New York Times noticed the same thing from the other side, that even China's strongest labs cannot yet figure out how to make money from any of this.
And there is a clock on it. The Financial Times reports that Beijing is weighing tighter export controls on its own AI models and chips, the same week its models are quietly conquering the American back office. Washington is mulling restrictions from the other direction. So the open window, where a US company can download a Chinese model on Friday and put it in production by Monday, is the product of a moment when neither government has decided the weights are a weapon. That moment may not last the year.
The Briefing
Moonshot open-sourced the model everyone is now renting. The company held what it called a Kimi K3 open day, releasing the full weights, a 47-page technical report, and the key infrastructure tricks behind a 2.8 trillion-parameter reasoning model. Huawei pushed same-day Ascend support for training and inference, which means the model that is eating American API bills also runs on Chinese silicon end to end. Caixin reports Microsoft is weighing whether to route some Copilot traffic to K3. The open-weight release is what makes the trade-down above possible. You cannot switch to a model you cannot download.
Anthropic would not sign the open-source letter. Nvidia's Jensen Huang organized an open letter backing open-weight models, and most of the American industry signed. Anthropic declined. The refusal lands in the middle of a fight that started when the White House science office accused Moonshot of distilling Anthropic's Fable model to train Kimi K3, a charge Silicon Valley largely rejected. The split is real. The company with the most to lose from commoditized intelligence is also the one holding out against the open-weight tide, and it is worth watching by the fall whether that reads as principle or as self-interest.
China is thinking about locking its own models up. The Financial Times reports Beijing is considering tighter export controls on advanced AI models and chips. The irony writes itself. The country that just handed the world the largest open-weight model on the internet is now asking whether model weights are a strategic export, like lithography tools or rare earths. If it happens, the download-it-now advice going around developer forums is not a joke. The models that are open today may be the last generation that ships without a license.
The memory story keeps validating. CXMT became China's most valuable A-share company after its 8.6 billion dollar IPO, and the surge dragged on American chip names. The follow-through is the tell. Samsung is reportedly considering Chinese DRAM from CXMT for the low-end phones it sells in China, which is the demand side proving the supply side. We led on this Monday, so one line today. When your competitor starts buying your memory, the listing was not a bubble.
Apollo Go crossed into left-side-of-the-road territory. Baidu launched public-road robotaxi testing in London on Monday with Lyft's Freenow unit, planning passenger service from 2027. A week earlier it took a Hong Kong permit and became the first platform running fully driverless tests in a right-hand-drive market. That matters because right-hand-drive covers 70-plus countries and two billion people, a third of the world, and every Western robotaxi program has stayed left-hand-drive at home. Apollo Go now operates in 27 cities and has done more than 22 million rides.
What I Found on Bilibili This Week
The video I want to highlight is a Chinese developer's hands-on test of DeepSeek's new V4-Pro, which is in gray-release on the API right now (BV1fLMJ6GEJ6). He gave the same prompt, a double-wishbone suspension kinematics simulation in 3D, to eleven models at once, and this is exactly the kind of thing you never see in English. Not a benchmark chart, a real person watching outputs render.
His verdict on V4-Pro is a good snapshot of where the frontier actually sits. The model built a working simulation with correct spring-damper and kingpin geometry, though it got the steering tie-rod pivot wrong, a mistake he says most models make. It cost about 6 mao, under 9 cents, on 1.2 million tokens. His reaction, verbatim, was that he will probably use it for most engineering work from now on and stop paying for GLM-5.2. He was harsher on others. Kimi 2.6 Pro burned the most tokens and produced unusable 2D output. Qwen 3.7 Max he could not tell what it was doing.
That 9-cents-a-task number is the whole newsletter in one data point. A working piece of engineering software, generated for under 9 cents, by a model most of the English-speaking world has never run.
Signals
Lilian Weng left Thinking Machines. The Peking University alumna and well-known alignment researcher resigned from Mira Murati's startup, leaving the founding team down to two. The Chinese-born talent that built a lot of American frontier AI is increasingly mobile, and increasingly a story Chinese media tracks closely.
AgiBot filed to go public in Hong Kong. The humanoid maker started its IPO process, and its WITA-Omni Preview model topped the DailyOmni all-modality understanding benchmark the same week. Robots that can read a room are the near-term prize, not robots that can dance.
BYD will show its first humanoid in early August. The carmaker confirmed a debut for its own robot, joining a field where the manufacturing base, not the lab, may end up owning embodied AI.
Honor is putting a robot on a phone. Honor confirmed an August launch for a Robot Phone with a 4-degree-of-freedom gimbal, a small physical arm on a handset. It sounds like a gimmick until you remember the last consumer-AI form factor everyone dismissed was the smart speaker.
The Bigger Picture
Here is the question this week actually poses. If intelligence is becoming a utility, where does the money end up?
The American bet, the one funding hundreds of billions in data-center capital, is that the frontier stays scarce and defensible, that the best model is worth a premium large enough to pay for the next one. The trade-down is the first hard evidence against that bet. When Coinbase can halve its bill and a legal startup can run daily work at a thirteenth of the cost, the premium on the frontier is being confined to a shrinking slice of hard tasks. The rest becomes a commodity, and commodities do not fund 60 billion dollar valuations.
The Chinese read, visible across this week's commentary, is that the value was never going to sit in the model. It sits one layer down, in the compute, the power, the interconnect, the fabs. That is why Huawei keeps turning up as the quiet winner of stories that are ostensibly about someone else's model, and why CXMT can become the most valuable company on the Shanghai exchange selling memory. If the model is a commodity, you want to own the factory that makes the commodity possible.
Which leaves the open weights as the strange gift in the middle. Right now a US company can download the frontier for free and run it on Monday. Both governments are now looking at that window and asking whether it should be closed. The most important number in AI this year might not be a benchmark score. It might be how many months are left before the thing you can download today needs an export license.
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.