Happy Friday. I scan more than a hundred Chinese-language sources every morning, the trade press and company papers and market wires that shape how China talks about its own AI industry, and I write up the parts that never reach English. Let's go.
The Cheaper Teacher
The World AI Conference opened in Shanghai today, and for the first time Xi Jinping gave the keynote in person. The 36Kr preview of the event ran under the headline "who is still racing parameters," and its answer was that almost nobody at this conference is. The floor this year is robots and agents that do actual work. The two most interesting things I read this week are both about the same unglamorous problem holding that up, both come from Chinese labs, and neither has reached English coverage.
The problem is data. A robot learns to fold a shirt by being shown thousands of examples of the motion, and those examples have to come from the physical world, usually a person operating or wearing a capture rig who repeats the same grab over and over. It is slow, it is expensive, and it does not transfer. Move the camera or swap the arm and the data is close to worthless. Tesla runs a dedicated human-demonstration team at its Fremont factory for exactly this reason, workers wearing camera arrays and doing ordinary tasks so that Optimus has something to imitate. Everyone in embodied AI agrees this is the wall. The disagreement is about how to get over it.
Galbot, the Beijing humanoid startup whose Chinese name is Yinhe Tongyong, published a way around it. Its framework, called WAM-TTT, lets a robot keep learning after it is deployed. You hand the robot a single ordinary video of a person doing the task, no labels, no motion-capture, no robot data, and it adapts on the spot. The base model stays frozen and a small memory module does the learning, which is why the robot does not forget what it already knew. The number that matters is in the ablation. Trained on 100 robot demonstrations plus 100 human videos, WAM-TTT hit a 74.1 percent average success rate, roughly the same as training on all robot data, per the paper, one phone video standing in for one expensive robot session. Galbot is calling WAM-TTT the first test-time-training framework built for a robot, and the reason no one had done it before is that the physical world does not forgive a mistake the way a text model can just resample a bad token.
The second result is Xiaomi's, and it goes at the same wall from the opposite side. Instead of needing less data, Xiaomi collected a lot of it, 100,000 hours of real-world manipulation, most of it captured with a cheap portable rig that records people rather than robots. Then it did the thing nobody had cleanly shown for robots. It plotted the scaling law. As the training data went from 2,500 hours to 20,000, the model's action-prediction error kept falling. As the model grew from 2 billion parameters to 10 billion, it kept improving, and the gains showed up not just in the offline metric but in real success rates on tasks the robot had never seen, like tidying a shoe cabinet in an unfamiliar home. Xiaomi says the model tops the public RoboDojo leaderboard at 20.07 against a previous best of 13.07, and reports leading numbers on the household RoboCasa benchmark, though these are results the company posted itself around the same week the benchmark went up, so treat the toppings as its own claim until an outside group reruns them.
Put the two together and you get the week's actual thesis. Xiaomi is arguing that robot ability now scales with data and model size the way a language model does, which turns embodied AI from artisanal tuning into something you can budget for. Galbot is arguing that the data you need can increasingly be human video instead of robot demonstration, which is where the cost lives. Both are lab results and self-reported benchmark toppings, with success rates still in the 70s that no one would call deployment-grade. What neither team has shown is whether the cost collapse holds up in a real warehouse rather than a curated test set. That is the number I want to see next, and it is the one that decides whether 2028, the date Galbot's founder keeps naming for the robot version of the ChatGPT moment, is a forecast or a slogan.
The Briefing
China shipped two frontier-scale open models in the same week, and the price of intelligence is where the fight is now. Moonshot released Kimi K3 at 2.8 trillion parameters, which Qbit reports is the largest open-weight model anyone has published, and it took the top spot on the Frontend Code Arena, ahead of the leading US models. The size is the point. An open model that big, which any company can download and run on its own hardware, gives an enterprise a frontier-class option it fully controls, and it is the open weights, not the benchmark, that let a firm fine-tune and self-host rather than rent. DeepSeek's V4, out in preview since April, reached its general release this month. The consequence showed up on the demand side. US developers now route more than 30 percent of their tokens on the OpenRouter platform to Chinese models, a share that has peaked near 46 percent, up from about 4.5 percent in the first half of 2025, per data compiled from OpenRouter and relayed in the Chinese trade press. Sam Altman wrote this month that GPT-5.6 Sol is already half the price of Anthropic's top model and that OpenAI would be happy to deliver it at a quarter of the price, which the Chinese press read, fairly, as a reaction to a Chinese cost floor that keeps dropping. The honest caveat is that these Chinese models still trail the top US labs on the hardest independent coding benchmarks, and that Anthropic still collects most of the actual dollars on these platforms even while its token share falls. The number that would settle it, revenue per token by vendor, is the one nobody in this data has disclosed.
DeepSeek's first outside funding round closed at a valuation of 350.8 billion yuan. The figure leaked sideways, through a disclosure by a listed luggage maker, Kairun, whose subsidiary holds an indirect 0.0114 percent stake and had to file the arithmetic, per Caixin. Back out the fraction and the post-money lands at 350.877 billion yuan, up from a 300 billion pre-money, a round of roughly 50 billion yuan with Tencent, NetEase, and CATL among the investors. Kairun's own stock jumped 19.98 percent on the news, which tells you how much retail money in China is trying to find any listed thread connected to DeepSeek. I wrote on Wednesday about why the lab that swore off capital finally took it. This is the receipt.
A Suzhou company says it is the first in the world to mass-produce map-free L4 self-driving. Zelos, known in Chinese as Jiushi, took a stake from Alibaba's Cainiao logistics arm in January, so its vehicles are delivery vans, not robotaxis. At WAIC it announced that its L4 system no longer needs high-definition maps, running instead on real-time perception and on-vehicle decision-making, per IT Home. The claim comes with operating numbers rather than a demo. The map-free system has reached 30 percent penetration on Zelos's new routes, should pass 40,000 kilometers of cumulative driverless mileage by the end of this month, and cuts the time to deploy a vehicle on a fresh route to under a day. A one-day deployment is the difference between a fleet that needs a mapping team for every new city and one that follows Cainiao's delivery network wherever it already goes.
SenseTime says it cut the electricity cost of generating tokens by 80 percent. At its WAIC showing, the company announced a compute-and-power scheduling agent for its SenseCore infrastructure that it claims raises token output per unit of electricity cost by 80 percent, per Qbit. SenseTime runs its data centers on chips that are not the most power-efficient available, since it cannot buy the best Western parts, so its cost per token starts higher than a US hyperscaler's before the scheduler does anything. A gain in tokens per kilowatt-hour is one of the few claims this week that can be checked against an actual electricity bill rather than a leaderboard, so the number to watch is whether the 80 percent holds outside the launch demo.
What I Found on Bilibili This Week
The video I want to flag is a 22-minute retrospective from the channel Lau Boshi de Yunzuhui, which has pulled 446,000 views and walks through DeepSeek's whole arc from V1 to V4. Two things stood out. The first is a piece of DeepSeek lore worth knowing. The firm's compute base is a cluster called Fire-Flyer, built by Liang Wenfeng's quant fund years before the AI turn, the second version costing around 1 billion yuan and holding roughly 10,000 Nvidia A100 cards, originally bought to trade markets faster. The model lab was standing on a stockpile of GPUs that existed for another reason entirely.
The second is the claim the video leads with, that V4 Pro is now the coding model DeepSeek's own engineers use internally, described as better than one Western frontier model and close to another. I want to flag that carefully rather than repeat it. That framing is DeepSeek's own, applied to the benchmarks where V4 looks strongest. On the harder, contamination-resistant coding tests run by independent groups, the gap to the top US models is still real, and the most credible English-language review, from ChinaTalk, put V4 several months behind the frontier. The interesting fact is not that a Chinese model beat an American one. It is that the question is now close enough that a Bilibili channel with 446,000 views spends its energy arguing the margin.
Signals
China published its first L3 certification for AI devices, covering 66 products from 17 companies. The results, announced at WAIC, span phones, PCs, TVs, glasses, car cockpits, and speakers, graded against a national standard issued in May that runs from L1 up to a current top tier of L3, per Caixin. The standard went out in May, before any single product line has won the on-device AI category, so whoever sets the L1-to-L3 criteria now is defining the grade that other countries' devices will later have to certify against.
Domestic chips took 41 percent of China's AI accelerator market in 2025, per an IDC figure making the rounds. The robust part of the number, cited in a widely-shared Bilibili explainer, is the fall in Nvidia's share, from close to 95 percent before export controls tightened to about 55 percent now. The decomposition below that is contested, with the explainer putting Huawei's Ascend at 20 percent and Alibaba's T-Head at 7 while other trackers put Huawei higher. The magnitude is what to hold onto. Domestic share went from single digits to 41 percent in under three years, all of it money the controls kept inside China.
A World AI Cooperation Organization was founded in Shanghai this week, with 29 countries signing on. The founding members include China, Russia, Kazakhstan, Pakistan, Indonesia, Brazil, and a run of African and Latin American states, and UN Secretary-General Guterres attended the signing, per Caixin. The body sits in Shanghai, outside the UN system, and its stated pitch is that open-source AI is the path to spreading the technology widely, per Caixin. More on that below.
The Bigger Picture
The lead story and the governance story are the same story. On the show floor, Chinese labs are giving away frontier-scale models and publishing the training methods that make robots cheaper. In the conference hall, the Chinese government stood up an international organization, headquartered in Shanghai, whose founding argument is that open models are how the rest of the world gets AI. When your national champions lead in open weights, setting the global standard around open weights is a way to write the rules other countries then certify against, the same move China has run in solar panels and EV charging.
Washington is moving the other way in the same week. InfoQ picked up reporting, echoed by the researcher Nathan Lambert, that the US administration is weighing an executive order to restrict open AI, specifically Chinese systems, most likely a rule blocking any open-weight model more capable than a named frontier line, a threshold Lambert thinks could land inside six months. One government is building the body that blesses open models and the other is drafting the order that bans them, at the moment Chinese open models are the ones setting the price. That timing is the single sharpest fact in this issue.
I do not think the robots or the models are the whole point, and I am not confident the leaderboard toppings will hold when outside groups rerun them. The strategy underneath is legible enough. Make the best model the cheapest, give it away, publish the method, and be the country that also runs the organization deciding who gets to distribute it, in models and in robot training data alike. What none of this week's launches answered is whether any of it makes money, or whether the whole thing runs on giving the best product away and hoping the market that forms on top is one you still control. That is the question I will be watching for the next three days of this conference.
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.