Happy Sunday. 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 Windfall
On Thursday night a company called Longsys filed a profit warning that reads like a typo. First-half revenue of 22 to 25 billion yuan, up around 130 percent. Net profit of 9.2 to 11 billion yuan. Year on year, that profit line is up somewhere between 62,204 and 74,394 percent.
The filing is not wrong. A year ago Longsys earned 14.76 million yuan in the same half. This year it earned up to 743 times that. For context, its best-ever full year, 2025, brought in 1.42 billion yuan of net profit. Longsys just made roughly seven times its entire 2025 in six months, and it is probably the most profitable company on the whole A-share semiconductor board right now.
Here is the part that should make you sit up. Longsys is not a chip designer. In the Chinese trade it gets called a 二道贩子, a middleman, a reseller. It buys DRAM and NAND wafers from Samsung, SK Hynix, and Micron, wraps them in its own controller and firmware, and sells finished memory modules to phone and server makers. For years this was the least glamorous seat in the semiconductor supply chain. The seat nobody wanted just printed the year's most obscene earnings number.
What happened is a timing trade, executed at enormous scale. Longsys spent years quietly stockpiling memory, from 2.2 billion yuan of inventory in 2020 to 17.96 billion by the first quarter of this year, which is about 1.8 times a full quarter of sales. It bought cheap in the trough. Then AI turned the memory market violent. TrendForce clocked second-quarter contract prices up 58 to 63 percent for DRAM and 70 to 75 percent for NAND. Longsys is selling today's product at today's spiked prices while carrying inventory it booked at yesterday's cost, and the spread is the profit. To keep the goods flowing it levered up hard, doubling long-term borrowing and running prepayments up 440 percent, a balance-sheet bet that prices keep climbing.
It is not one company. Rival module maker Demingli posted a first-quarter profit up 4,943 percent. GigaDevice, a designer left for dead as a commodity NOR Flash outfit at 55 yuan a share in early 2024, now trades near 815 yuan, a market cap around 570 billion yuan, because the big three are walking away from the low-margin niche DRAM it makes and handing it the market.
Why walk away? Because AI eats wafers. Every gigabyte of high-bandwidth memory for AI accelerators consumes about three times the wafer area of ordinary DDR5, rising toward 4.3 times by 2028 on Morgan Stanley's numbers. Samsung, SK Hynix, and Micron are pouring their best capacity into HBM, which now leaves less for the plain memory in your laptop, your car, your router. Micron just reported a single quarter of 41.4 billion dollars at an 84.9 percent gross margin and signed 16 take-or-pay contracts, mostly running to 2030, locking in roughly 100 billion dollars. SK Hynix has a name for it, an "HBM-led memory super-cycle."
So enjoy the number, but read the asterisk, because it is the same story as the number. Longsys made 743 times its money on a one-time revaluation of cheap inventory into an expensive market. That is a windfall, not a business model, and windfalls end. TrendForce already sees third-quarter DRAM gains cooling to 13 to 18 percent as consumer buyers hit their price ceiling. Xiaomi, OPPO, and vivo have told suppliers to cut shipment targets by as much as 30 percent, and the decisive reason is the cost of memory. A cloud giant building an AI data center can absorb a higher memory bill. A 1,000-yuan phone cannot. The AI boom's appetite is being paid for, one layer down the stack, by the devices in everyone's pockets, and Longsys just showed you the invoice.
The Briefing
China's most advanced chip line cannot feed its own industry, and the rationing is now political. A sharp tmtpost breakdown puts hard numbers on it. SMIC's N+2 process, the only line in the country making leading-edge AI chips at scale, targets about 2.6 million AI chips this year against domestic demand near 4.2 million, a 40 percent gap. Huawei has locked 43 percent of that capacity on a five-year contract, having embedded its HiSilicon team beside the line and contributed most of the yield-ramp data. Cambricon holds 9 to 11 percent, secured less by chip quality than by its Beijing state backing. Everyone else, Biren, Moore Threads, Alibaba's T-Head, fights over the rest. You can measure how scarce the capacity has become from what Baidu's Kunlunxin unit is now doing. It is chasing a 50 billion dollar Hong Kong listing, higher than Baidu's own market value, with a clause reportedly asking investors to commit to buying three to seven times their subscription in chips. When you have to force your shareholders to become your customers, the constraint is not capital. It is a place in line at the fab.
Meituan open-sourced a trillion-parameter model trained entirely on domestic silicon. LongCat-2.0 has 1.6 trillion total parameters, activates around 48 billion, and was trained end to end on a 50,000-card Chinese compute cluster with more than 30 trillion tokens and native million-token context. A food-delivery company is not who you would expect to ship the largest model trained without Nvidia, which is exactly why it matters. The interesting China story this year is less about any single frontier model and more about how many capable players now train around the chip controls rather than through them. Reuters framed it as catch-up. The Chinese coverage reads it as proof the domestic stack is now good enough to be boring.
The embodied-AI capital flood got louder, not quieter. In one week's roundup, humanoid-brain startup AI2Robotics closed nearly 5 billion yuan at a 20-billion valuation, X Square ran four rounds in two months past a 20-billion valuation, and home-robot maker Lexiang raised from Ant with orders past 30,000 units and first-half revenue up 600 percent. Kuaishou's video model Kling landed close to 3 billion dollars at an 18-billion valuation, the largest raise a video-AI company has ever done, with Tencent, Alibaba Cloud, and Baidu all in. Moonshot's Kimi is now marking a 31.5 billion dollar pre-money, disclosing annual recurring revenue past 300 million dollars with API sales above 70 percent of the total. The money is not waiting for profits. It is buying position in robotics and in the few labs showing real revenue curves.
A world model built in China beat Nvidia's on the field roboticists actually use. Wujie Power released MWA, a latent-space world model for embodied control, and it topped the RoboCasa GR1 TableTop benchmark, placing above Nvidia's GR00T-N1.6, Xpeng's DIAL, and others. The same week, Fei-Fei Li put her name on a new embodied paper arguing the field has the simulation problem backwards, that generating training worlds from a single video (Real2Sim) is cheaper and more plentiful than the expensive sim-to-real pipeline everyone chased. Chinese labs are converging on world models as the path to general robots, and they are no longer doing it a step behind.
An AI agent found four new superconductors, and they check out in the lab. Alibaba's DAMO Academy, with Renmin University and the Chinese Academy of Sciences, released Elements Claw, an agent that predicted 68,000 candidate superconducting materials. Four brand-new ones have now been synthesized and confirmed to actually superconduct, with the data released open. This is the quiet, unhyped version of AI-for-science, a model proposing real materials that a lab then verifies, and it is worth more than any chatbot demo.
Signals
A-share trading rules change Monday. From July 6, daily price limits on ST-flagged stocks widen to 10 percent and an after-hours session is added, per Caixin. Small mechanics, but they reshape how the retail-heavy market prices the AI and chip names that have been ripping all year.
Anthropic is designing its own chip and talking to Samsung to build it. It is following OpenAI in trying to escape Nvidia's margins, and it also lifted export controls on two frontier models while running Kimi K2.7 in the same core safety evaluation as its own Opus and GPT-5.5. Kimi flagged the same critical vulnerabilities. The parity keeps showing up in places that are hard to spin.
China's tax data quietly shows the robot economy is real revenue, not just rounds. Sales at embodied-AI firms rose 22.4 percent year on year in the first five months, according to the tax bureau's receipts. Funding announcements can be theater. Tax filings are harder to fake.
GitHub Copilot added its first open model, and it is Chinese. Kimi K2.7 Code is now selectable inside Microsoft's Copilot, the first open-weight model the tool has ever integrated. Meanwhile Alibaba is banning Claude Code internally from July 10 over a reported backdoor risk and pushing its own Qoder instead. The tooling layer is quietly picking sides.
The Bigger Picture
Step back from any single number this week and the same shape keeps appearing. The binding constraint on AI is no longer who has the cleverest model. It is who controls physical capacity, and where it is allowed to flow.
You can see it at every layer of the stack. At the top, HBM is swallowing the world's most advanced wafers, which is why Longsys and GigaDevice are having the strangest year of their lives one rung down, in the commodity memory the giants abandoned. In the middle, a single SMIC line decides which of a dozen Chinese chip champions lives, and the decision is made by national priority and provincial politics as much as by engineering. Korea's answer is to spend more than 500 billion dollars building four new fabs to double its memory capacity. Everyone has concluded the same thing at once, that the scarce resource is the factory, not the idea.
For China this reframes the whole import-substitution project. The famous fights are about the frontier, the leading-edge logic Nvidia and TSMC still gate. But the quieter, more winnable fight is one layer down, in mature-node memory and inference silicon, where the incumbents are voluntarily retreating toward higher-margin AI parts and leaving the field open. CXMT's fab capacity behind GigaDevice, LongCat trained on 50,000 domestic cards, Kunlunxin forcing demand through its IPO, these are all the same move, occupying the ground the leaders walked away from.
And the bill lands somewhere. When the world's memory pours into AI data centers, the shortage flows downhill to the cheapest devices, and the person subsidizing the boom turns out to be whoever is buying a budget phone this year. Longsys made 743 times its money on that transfer. The number is a windfall. The mechanism is not going away.
None of this makes Western headlines in the shape it actually has. All of it decides who wins 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.