Happy Monday. I scan more than 100 Chinese-language sources every day, the WeChat accounts, the Bilibili channels, the finance wires, the policy feeds, and I write up the China AI stories that English-language coverage misses. No team, no wire service, just the reading. Let's go.
The Speed Limit
On Friday DeepSeek pushed a quiet update to its V4 models and, almost as an afterthought, open-sourced the part that matters. It is called DSpark, a paper written with a team from Peking University and signed by Liang Wenfeng himself. On live production traffic it makes DeepSeek V4-Flash generate text 60 to 85 percent faster for a single user. The larger V4-Pro got 57 to 78 percent. Same chips, same model weights, same answers. Just faster.
DSpark is not a new model. It is an inference technique called speculative decoding, and the plain-English version is that a small, cheap draft model guesses the next several words, then the big expensive model checks the whole batch in one forward pass instead of producing them one token at a time. The idea is old. What DeepSeek and Peking University fixed is the part where it falls apart in production. Parallel draft models are fast but sloppy, and the further into a guess they go, the more the big model rejects what they wrote, a problem the paper calls suffix decay. DSpark adds a light serial module that models the dependencies inside each block, and a scheduler that decides which guesses are even worth verifying when the servers are busy. Coding and math get checked aggressively because the answer path is narrow. Open-ended chat gets checked conservatively because it is not.
The reason this is the lead and not a footnote is the chip it runs on. DeepSeek cannot buy more Nvidia, and the export ceiling that started with the H100 has only tightened since. When you cannot add hardware, the only direction left is down, into the software stack, squeezing more tokens out of every GPU you already own. DeepSeek's own framing is that DSpark eases inference bottlenecks and chip strain, which is a polite way of saying it does the work the chips they are barred from buying would otherwise do.
So the win and the constraint are the same fact. A lab with unlimited access to compute does not spend its best researchers and a Peking University collaboration on shaving milliseconds off the decode loop. It buys another data center. DeepSeek wrote DSpark because shaving milliseconds is the lever it has, and then it open-sourced the result, which is the move that makes everyone else's GPUs go further too. Efficiency is what scarcity buys you, and right now China is the one being forced to get good at it.
The Briefing
The most valuable thing inside Baidu is no longer Baidu. The company is spinning off its Kunlun chip unit for a Hong Kong listing at a target valuation around 50 billion USD, which is larger than Baidu's entire market capitalization, per tmtpost's reporting. Alibaba's T-Head chip arm is rumored to be next. One detail tells you what kind of market this is. Kunlun's roadshow asked would-be IPO investors to first commit to buying chips worth three to seven times their share order. Want to be a shareholder, be a customer first. Tencent already is one. For years these in-house silicon teams were cost centers that existed to make training cheaper than renting Nvidia. The Agent era flipped the math, because the real compute drain is not training a model once, it is serving inference to it a billion times a day, and a domestic chip that is merely good enough to serve inference is now a business worth more than the search engine that funded it.
DeepSeek is about to start charging by the hour. The V4 release that carries DSpark also comes with a launch date, mid-July, and a new pricing rule. API calls during peak hours, roughly 9am to noon and 2 to 6pm Beijing time, will cost twice what they cost off-peak. This is the same scarcity problem seen from the billing side. DSpark stretches the supply of compute, peak-and-off-peak pricing rations the demand for it, and both exist because DeepSeek is trying to serve a frontier model on a hardware budget that cannot simply grow when traffic does. American labs price on value and capability. DeepSeek is pricing on congestion, the way a toll road does.
A robot ran a phone assembly line for 64 hours and missed almost nothing. AgiBot livestreamed its G2 humanoid working the tablet quality-inspection stage at a Longqi factory in Nanchang from June 23 to 28, 64,828 tasks at a 99.99 percent success rate. The robot pulled units off a moving line, talked to the test rig, sorted pass from fail, and put them back. The same week, AgiBot said its 15,000th G2 had come off its own production line. The number that matters is not the headcount, it is the 99.99 percent, because that is the threshold where a humanoid stops being a demo and starts being cheaper than the shift it replaces.
The labs are all raising at once. This was the week the funding cycle for Chinese frontier models turned frantic, with DeepSeek, Moonshot, and MiniMax all moving on capital inside a few days. The trigger people keep naming is competitive rather than financial. Anthropic's Mythos cybersecurity result is what reportedly convinced Liang Wenfeng that staying self-funded was no longer an option. I am holding the full version of this story, including who is raising how much and why DeepSeek breaking its no-outside-money rule is the real signal, for the paid Monday Brief.
Signals
China has its first national standards body for humanoid robots. The Ministry of Industry released a national standard system for humanoid robots and embodied intelligence, the bureaucratic step that usually comes right before procurement budgets do. Standards are boring and they are also how Beijing tells a supply chain that a sector is now official.
Momenta opened its book. The autonomous-driving company is marketing what it calls the first physical-AI IPO, with 14 anchor investors lined up and those base seats described as hard to get. After a year where Chinese AD listings fizzled, an oversubscribed anchor round is a small tell that the appetite is back for the ones with real revenue.
Embodied-AI valuations are compressing into a few names fast. The robot foundation-model startup Zibianliang closed four rounds in two months to cross 20 billion yuan, and Zhiping Square raised a fresh 5 billion to reach a similar mark. The capital is not spreading across the field, it is stacking onto a short list of presumed winners.
The Bigger Picture
There is a number that should bother anyone who thinks this race is decided by who owns the most chips. By Epoch AI's estimate, OpenAI ended 2025 with compute equivalent to roughly 1.7 million H100s, up seventeen-fold in two years. And yet, as an analysis circulating in Chinese tech media this week laid out, frontier labs may be running those GPUs at under 10 percent real utilization, with three quarters of enterprises stuck below 70 percent even at peak load. The piece reached for the right historical analogy. In the late 1990s American telecoms buried millions of miles of fiber, and when the bubble burst more than 95 percent of it was dark, never lit, never carrying a byte.
The American AI build-out is laying a lot of dark silicon. The constraint everyone talks about is the number of chips, but the constraint that actually binds is how much useful work comes out of each one, and on that measure the most expensive clusters on earth are leaving most of their capacity idle because nobody has to be efficient when they can just buy more.
DeepSeek has to be efficient. That is the whole point of DSpark, of the congestion pricing, of a chip unit worth more than its parent because it can serve inference at a price the market will pay. Export controls were supposed to cap how much compute China could field. What they may actually be doing is forcing China to become the part of the industry that knows how to use compute it already has, while the side with the bigger pile learns to leave most of it dark. A speed limit is only a disadvantage if you were planning to win on top speed.
I exist because this information asymmetry shouldn't. If a friend keeps up with AI but only reads English, forward them this issue.