Happy Tuesday. 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 English-language coverage misses. One person reading the Chinese internet so you don't have to. Let's go.
The Harder Half
The company that crossed the line everyone said was uncrossable is a food-delivery app. On Tuesday Meituan open-sourced LongCat-2.0, a 1.6 trillion-parameter model with a million-token context window, roughly the scale of DeepSeek's flagship V4-Pro. The headline number is not the interesting part. The interesting part is where it was built. Meituan says LongCat-2.0 is the first trillion-parameter model to complete its entire training and inference run on a 50,000-card domestic computing cluster, with no foreign accelerators anywhere in the loop.
To see why that matters you have to split the chip problem in two. There is inference, the part where a finished model answers a query, and there is pre-training, the part where the model is built in the first place by grinding through a data set the size of the readable internet. Inference is the easy half. Chinese chips have been good enough to serve models for a while, and DeepSeek's own V4-Pro runs inference on domestic hardware today. Pre-training is the hard half. It needs tens of thousands of chips holding a single computation together for weeks, enormous memory bandwidth, and a distributed software stack that does not fall over when one node in fifty thousand does. The export-control thesis was never really betting that China couldn't run models. It was betting China couldn't train them at the frontier without Nvidia.
A month ago I wrote about the first crack in that bet, in an issue I called The Last Dependency. LongCat-2.0 is the crack widening. What changed is the claim is now about pre-training specifically, at 1.6 trillion parameters, on what Meituan describes as clusters of tens of thousands of "AI ASIC superpods." An ASIC is a chip built for one job rather than a general-purpose GPU, and in China that language points at the domestic accelerator lines, Huawei's Ascend and its peers, though Meituan pointedly does not name which silicon it used.
That coyness is the first asterisk, and there are a few. This is a claim, self-reported, not yet independently benchmarked by anyone outside Meituan. "Trained on domestic chips" is a proof of possibility, not a proof of parity. It tells you the run finished and the model came out competitive. It does not tell you the run was cheap, or efficient, or that it would have been anyone's first choice if the H100s were for sale. Burning a 50,000-card cluster for weeks is the kind of thing you do when you have no alternative, not when you have a better one.
But the narrow version of this story is bigger than the headline, not smaller. The specific capability export controls were designed to deny, frontier-scale pre-training, is the capability a company most people don't even think of as an AI lab just said it no longer needs to import. And the fact that it's Meituan is the tell. A food-delivery business runs on per-order margins measured in fractions of a cent and already operates one of the larger data-center footprints in the country to route couriers in real time. When training has to move onto domestic silicon, the company that squeezes compute for a living is as likely to get there first as the labs that buy their way out of every constraint. Scarcity rewards the cheap, and nobody in Chinese tech is cheaper than the people delivering your lunch.
The Briefing
China now has a trillion-yuan AI chip company, and it has barely shipped anything. Cambricon's stock jumped 8.84 percent on Tuesday to a market value of 1.01 trillion yuan, making it the first pure-play AI chip designer in China past the trillion mark and the first stock on the STAR Market to get there. The financials are real and growing, first-quarter revenue up 159 percent to 2.83 billion yuan, a quarter that alone matched 43 percent of all of last year. The shipments are the part to sit with. By IDC's count the entire Chinese market bought about 4 million AI accelerator cards in 2025, and Cambricon shipped roughly 116,000 of them, a 2.9 percent share, tied for fifth, while Nvidia held 55 percent. A trailing price-to-earnings ratio of 373 is not pricing the chips Cambricon sold last year. It is pricing the LongCat-shaped future where every frontier run in China has to happen on cards like these.
The carriers just put 40,000 servers on the table, and Huawei swept the big half. China Telecom closed bidding on a roughly 11.5 billion yuan tender for 40,000 high-performance servers, and the larger package, 28,000 ARM-architecture machines worth about 8.1 billion yuan, went entirely to six vendors that all build on Huawei's Kunpeng platform. China Mobile and China Unicom are running their own buildouts in parallel, and the procurement language is explicit that the gear has to come from China or a normal trading partner. This is the demand side of the LongCat story made concrete. Somebody has to own the 50,000-card clusters, and increasingly it is the state-owned carriers, buying domestic by mandate, with the order book tilting from training toward inference deployment.
The one input nobody can route around is memory, and it is on fire. DRAM contract prices rose 93 to 98 percent in the first quarter and another 58 to 63 percent in the second, two record quarters in a row, while a single DDR5 spot part is up 627 percent over the year, per tmtpost's accounting. The cause is the same AI buildout, and the effect is a structural shortage rather than the usual six-month inventory cycle. The market is sorting winners from losers in real time. Micron jumped more than 15 percent the same week Apple fell 6 percent and shed 263 billion dollars, and Micron, Samsung, and SK Hynix are now facing a class-action suit alleging they coordinated the price moves. A model can be trained on a domestic GPU. The high-bandwidth memory stacked next to it is the dependency with no domestic answer yet.
A second domestic chip architecture just shipped, and it is not a GPU either. Zhonghao Xinying released its new TPU accelerator, called Xuyu, the follow-on to Chana, the chip the company says was China's first home-grown TPU when it taped out in 2023. A TPU is Google's bet that AI workloads want a tensor-specific design rather than a repurposed graphics chip, and Xuyu is tuned for exactly the long-context, huge-model traffic LongCat represents. The point is not this one part. It is that the "ASIC superpods" Meituan trained on are turning into a field with more than one credible name in it, which is what an ecosystem looks like before it looks like a supply chain.
Signals
A humanoid robot is now a 119,800-yuan consumer product with 11,000 orders. UBTech launched its UWORLD U1 line on Tuesday, a full-size bionic humanoid priced from 119,800 yuan for the Lite up to 990,000 for the Ultra, aimed at home companionship and trained on a Huawei Ascend stack. Pre-orders already crossed 11,000 units. For scale, UBTech sold 1,079 full-size humanoids in all of 2025.
BYD is bringing its self-designed driving chip in-house for 2027. LatePost reports BYD plans to put its own Xuanji A3 chip in a production Denza next year, a 4-nanometer part rated above 700 TOPS, three of them above 2,100, supporting L3 and L4 driving. Chairman Wang Chuanfu's framing is the quotable bit. The first half of the EV race was about the battery, he said, the second half is about the chip.
The efficiency race DeepSeek started last week already has a second entrant. Days after DeepSeek open-sourced its DSpark decoding framework, StepFun and collaborators released JetSpec, which claims to speed model decoding up to ten times. Two frontier labs racing to give away inference-acceleration code in a single week is what it looks like when compute is the binding constraint and everyone is reaching for the same lever at once.
The Bigger Picture
There is a number in an American utility filing that reframes this entire race. In the first quarter of 2026, at least 75 US data-center projects were stalled or delayed, representing 130 billion dollars of investment, and the reason was not money and not chips. It was electricity. Microsoft's Satya Nadella put it plainly on a podcast, that the worry is no longer a shortage of compute but a pile of chips sitting in a warehouse with nowhere to plug them in. A data center takes about two years to build. The high-voltage transmission line to feed it takes five to ten.
This is the part of the story the chip headlines keep missing. Export controls were a bet about silicon, and silicon turns out to be the constraint that moves fastest. The constraint that moves slowest is the grid, and on the grid the map flips. China wrote a phrase into its 2026 government work report, suan-dian xie-tong, computing-power and electricity working in concert, the idea that you schedule compute to follow cheap power instead of dragging power to follow compute. The country built 42 ten-thousand-card clusters in 2025, its data-center electricity use is climbing more than 45 percent a year, and its transformer exports rose 76 percent in a single quarter into a world that suddenly cannot build enough of them. CATL's chairman told Davos this summer that China's grid is mature enough to absorb the load, and that AI cut his own factories' power bills by 30 percent.
So put the two halves together. A country that can pre-train a trillion-parameter model on its own chips, the thing LongCat just claimed, and can also power the cluster that does it, the thing the grid numbers describe, has closed both ends of the constraint at once. The side with the larger pile of chips is the side now rationing them against a transmission queue measured in years. Efficiency will not save it, because the cheaper compute gets the more of it people use, an old rule that held for steam engines and holds here. Export controls were supposed to be the ceiling. They may turn out to have been the floor, the thing that forced one side to get good at training on what it has and powering what it builds, while the other learns that the most advanced chip in the world is worth nothing until you can turn it on.
I exist because this information asymmetry shouldn't. If a friend keeps up with AI but only reads English, forward them this issue.