Happy Monday. I scan more than 100 Chinese-language sources every day, 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 Giveaway
The honeymoon on free AI is over, and you can watch it end inside the companies building the models. In early June a Tencent employee opened the internal dashboard and found the monthly token allowance had gone from about 2,000 dollars to 1,400 yuan, and burned through it in two days. Across departments Tencent now hands staff a fixed monthly token budget, 1,400 yuan for senior engineers, 700 for everyone else. Doubao started charging consumers. The word that lived in engineers' config files two years ago, token, is now a line item on everyone's bill.
That bill is the real constraint now, and it explains a move out of Beijing that should stop you. For two years the story of Chinese AI has been about one scarcity, the advanced chip, and the embargo that caps how many a company can buy. If you cannot get more chips, the only other lever is to make each chip do more work per token. In the last days of June, DeepSeek and Peking University open-sourced a framework called DSpark that makes DeepSeek's own V4 model generate 60 to 85 percent faster for a single user, released free under an MIT license. The paper's last author, the senior slot a lab reserves for the person who runs it, is Liang Wenfeng, the founder of DeepSeek.
Here is the part that should make you sit up. Almost the same fortnight, The Information reported that OpenAI's engineers had found a way to cut inference costs by more than half, enough that the fleet of chips serving its logged-out traffic briefly dropped to a few hundred. OpenAI kept the method in-house. Its gross margin moved from 33 to 39 percent in a quarter, aimed at 52 by year end. Two labs, the same discovery in the same weeks, that running a model cheaply is now the whole game. One published it and gave it away. The other treated it as a trade secret.
Both choices are rational, and the difference is the story. OpenAI is protecting a premium that open weights have already eroded. On coding, DeepSeek's V4 runs within a few points of GPT-5.6 at a fraction of the price, and OpenAI's best model is now sold to roughly 20 approved US institutions rather than the open market. When your high-margin buyers have narrowed to the government, you guard every edge you have. DeepSeek is doing the opposite, spreading its stack by making it cheap and legal for anyone in the world to run fast. One optimizes for margin, the other for adoption, and adoption is how a standard gets set.
The clearest statement of the strategy came from a new hire, not a press release. DeepSeek is hiring hard, and its V4 author list now includes Gu Yuxian, a Tsinghua PhD who just joined. His line reads like the whole plan. When hardware is constrained, he said, algorithm innovation becomes the way to break the compute bottleneck. That is the entire Chinese AI thesis for the embargo era in one sentence, and DeepSeek just shipped the proof as open source.
Read the asterisk, because it matters. DSpark is a speculative-decoding trick, it drafts tokens fast and checks them, which helps one user's latency and can even hurt throughput when a server is packed. It is not a chip, and it does not close the hardware gap that started all this. Open-sourcing your efficiency work also hands it to your rivals. But that is the wager China's best-funded lab just made in public. It would rather everyone run its models cheaply than keep the discount to itself, and in a market where the binding cost is now the token, the giveaway may be the shrewdest thing it could have done.
The Briefing
Alibaba is banning Claude Code inside the company, and the reason is a backdoor. From July 10 Alibaba bars employees from using Anthropic's Claude Code at work, having put it on a high-risk-software list after reports it can carry a backdoor, and is pushing its own Qoder instead, per an InfoQ industry roundup. This is not paranoia in a vacuum. Anthropic is separately backtracking on what The Information calls spyware targeting Chinese users, after a Reddit post showed Claude Code quietly tracking and transmitting information. DeepSeek, meanwhile, just absorbed an open-source coding agent called Deep Code, tuned to its V4 models and already past 1,500 GitHub stars. The tool layer is splitting along the same border as everything else.
Chinese platforms are switching off their AI agent features, on a deadline set by regulators. On July 4 ByteDance's Doubao and Alibaba's Qwen told users, almost in sync, that they will retire their agent-building features on July 15. That is the day China's new rule on anthropomorphic AI interaction takes effect, the same rule behind Friday's companion-chatbot shutdowns. The pattern from that issue holds. The state is drawing lines around persona and emotional AI faster than the platforms can build it, and the platforms are complying by subtraction.
UBTech is selling a humanoid for the living room, and 11,000 people have already ordered one. UBTech unveiled the U1, its first full-size consumer humanoid, under a new home brand called UWorld, with 88 joints, an emotional-interaction model, and encrypted data kept on the device by default. Prices run from 119,800 to 990,000 yuan, roughly 17,600 to 145,000 dollars, and pre-orders on JD since June 2 have passed 11,000 units, with first shipments September 16. A humanoid you can actually buy, priced like a car, is a very different thing from a demo reel.
China answered Washington's latest chip-and-defense listings, and the answer was deliberately small. On June 22 Beijing's commerce and finance ministries hit 56 US entities, 10 with export controls and 46 with procurement bans. The names are almost all defense and rare-earth firms, Lockheed Martin, Boeing Defense, Shield AI, Red Cat, in sectors where US-China trade was already close to zero. This came after the US Defense Department added Alibaba, Baidu, and Tencent to its list of Chinese military companies. China had every opening to escalate into consumer tech and chose not to. The restraint is the message.
A trillion-parameter model from a food-delivery company is now open source. Meituan released LongCat-2.0, 1.6 trillion parameters trained end to end on domestic chips, and put the weights out. I covered the model on Sunday, but the follow-through is the point. Meituan is now restricting its own staff from using ByteDance's Doubao internally and steering them to LongCat and DeepSeek. The same companies that buy each other's models are walling them off from their engineers, one procurement memo at a time.
Signals
A clever hack cut one model's token bill by up to 70 percent, using OCR. A user found that rendering long text, system prompts, tool docs, chat history, into a dense image and letting Fable 5 read it back through OCR costs far fewer tokens than sending the text straight. About 48,000 characters that would run 25,000 text tokens became roughly 2,700 image tokens, a bill 59 to 70 percent lower. When people are gaming the tokenizer to save money, the token bill has become the product.
Even Tesla is rationing AI. Alongside Tencent's per-employee token budgets, Tesla capped staff AI spending at 200 dollars a week, per the same InfoQ roundup. The cost of running models has become real enough that the companies closest to the technology are the ones metering it hardest.
The token factory loses money on every token it sells. SiliconFlow, the inference middleman whose Hong Kong IPO filing I wrote up last week, posted a public-cloud gross margin of minus 119 percent, losing 1.19 yuan for every yuan of that revenue. Everyone from the chipmaker to the reseller to the app is now fighting over the same thing, the cost of turning a chip into a token, and almost nobody has made it pay yet.
The Bigger Picture
For two years the frame on Chinese AI was scarcity of one input, the leading-edge chip, and everything got read through the embargo. This week the industry answered in a different register. If you cannot get more chips, you make each chip carry more work, and you make each token cost less.
You can see the pivot at every layer. A lab open-sources an inference framework and its founder signs it. A Tsinghua researcher joins that lab and says algorithm innovation is what beats the hardware limit. A delivery company trains a trillion-parameter model on domestic silicon and gives it away. Users render prompts into images to shave the tokenizer. A cloud giant rations its own model to its own staff by the yuan. None of these is a frontier-model headline. All of them are about the same quieter thing, the economics of running the models we already have, on the chips China is allowed to buy.
The contrast with the American approach is the part to hold onto. OpenAI found a way to halve its inference costs and, so far, kept it, while its most capable model is gated to a government whitelist. China's most-watched lab took its efficiency work and published it. Both moves protect something. OpenAI protects a margin that open weights have already thinned. DeepSeek protects nothing and bets that being the cheapest stack to run is worth more than owning the trick. That is a wager on standards over secrets, and it is a very different theory of how you win.
None of this makes Western headlines in the shape it actually has. All of it decides who pays what to run AI for the next three years.
I exist because this information asymmetry shouldn't. If a friend would find this useful, forward it, and if someone forwarded it to you, subscribe below.