Happy Friday. 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 Grid
For two years the China AI story has been told through chips. Can they buy them, can they smuggle them, can they build their own. This week the number that mattered was on the electricity meter, and it points the other way.
On June 26 China switched on its first AI data center running entirely on green power, in Zhongwei, out in the Ningxia desert. A dedicated 500-megawatt wind-and-solar project wires straight into the servers, with no public grid in between. It is one facility, and one built to make a point. But it is the physical form of a phrase that just entered China's new five-year energy plan, 算电协同, "compute-power coordination," the idea that data centers and the power system should be planned and scheduled as a single machine. On Thursday a Beijing energy-storage company and a Shanghai compute operator signed a deal to do exactly that, a joint compute-and-power operation built for cheaper, greener supply.
Four days after Zhongwei, on June 30, the United States did the opposite. With a record heat wave bearing down on the mid-Atlantic, Energy Secretary Chris Wright signed an emergency order letting PJM, the grid operator for 13 states and Washington, order large data centers off the public grid and onto their own backup generators, on 15 minutes notice, to keep the air conditioning running in people's homes. PJM was staring down a peak near 166 gigawatts, enough to break a record set in 2006. The country that makes the best chips on earth had to choose between cooling its houses and running its data centers, and it chose the houses.
That is the shift worth naming. The scarce input in AI is quietly moving from silicon to power. The IEA expects the electricity going to data centers worldwide to roughly double by 2030, from about 485 terawatt-hours to 950, growing close to 15% a year, more than four times faster than demand from everything else. Whoever has the most cheap, spare electricity to pour into compute wins the half of the race that comes after the model is trained, the half where you actually serve it to a billion people.
On that axis China is not a little ahead. It is not close. China generated 10,161 terawatt-hours in 2024 against the United States' 4,393, more than the US, the EU and India put together. In 2025 it added around 540 gigawatts of new capacity while the US added about 63, roughly eight times as much in a single year, and BloombergNEF expects China to add more than six times the US total over the next five years. China builds the rough equivalent of the entire American grid every few years, and it is increasingly building it next to the data centers that will drink it.
None of this fixes the chip. Abundant power does not fabricate a leading-edge processor China still cannot make at parity, and it does not close the gap on the best models. Cheap electricity makes Chinese AI cheaper to run, not automatically better. The Zhongwei plant is a demonstration, not the norm, Chinese data centers drew only about 11% of their power from renewables in 2023, and the plan's own 80%-by-2030 target is a long climb. But the direction is not in doubt. One country is rationing power between its homes and its servers. The other is generating more than it can use and wiring the surplus straight into compute. If the AI race comes down to who can feed the most electricity to the most silicon, that is the one contest where China starts in front.
The Briefing
UBTECH put a full-sized humanoid robot on sale for the price of a car, and 11,000 people have already ordered one. On Tuesday the Shenzhen company unveiled the U1 Series under a new consumer brand called UWorld, its first humanoid aimed at households rather than factories. Founder Zhou Jian said the U1 has taken more than 11,000 orders across all channels since pre-orders opened on JD's platform on June 2, ahead of the first shipments on September 16. The three versions run from 119,800 yuan (about 17,600 dollars) to 990,000 yuan (145,000 dollars), the robot carries 88 joints and an emotional-interaction model, and UBTECH says user data is encrypted and kept on the device by default. The order number is the part worth watching. Chinese embodied-AI startups raised 93.5 billion yuan in the first half of 2026, five times a year earlier, and most of those companies still ship nothing. UBTECH is one of the few putting a priced product in front of real buyers and finding out what home demand actually looks like.
DeepSeek will start charging more for its next model during working hours. Tencent Cloud announced this week that the official release of DeepSeek V4 lands in mid-July, and that pricing will move to a peak-and-off-peak structure following DeepSeek's own. A frontier lab metering its model by the clock is telling you where the constraint sits. The bottleneck is no longer whether the model is good enough. It is whether there is enough compute, and enough power behind the compute, to serve everyone who wants to run it between nine and noon. Surge pricing on tokens is what scarcity looks like once it reaches the meter, and it lands the same week the lead story does.
China's server makers are buying compute by the billion. Three separate Chinese firms disclosed AI-hardware commitments in a single day on the market wire. Shenhao Technology plans to buy up to 2 billion yuan of servers to rent out as compute capacity. Digital China's subsidiary won a 371-million-yuan bid to supply Huawei-based AI-compute servers to a large state bank, built on its own super-node hardware. And Pengding Holding filed to raise up to 9.6 billion yuan for an AI-server and high-speed optical-module plant, part of a 12.7-billion-yuan project. None of this is a model or a benchmark. It is the physical layer, the racks and the interconnect, being poured in as fast as the capital can be raised, and every rack of it needs a wall socket.
Alibaba is banning Claude Code inside the company and pushing its own tool instead. Chinese outlets reported on July 3 that Alibaba has classified Claude Code as high-risk software, citing a reported security backdoor, and will bar employees from using it in the office starting July 10, steering them to its in-house Qoder assistant. The backdoor claim is unverified and comes from anonymous sourcing, so treat it as an allegation, not a finding. The verifiable part is the move itself. A US coding agent that a lot of Chinese developers leaned on is being swapped for a domestic one at one of China's largest tech companies, and the stated reason is security. Decoupling in developer tools now runs in both directions.
Signals
Boson Quantum closed a several-hundred-million-yuan pre-IPO round, and a national project led by Maiwei Technology kicked off work on flip-chip bonding equipment for quantum chips. The quantum-as-compute-hedge trade I wrote about in The Hedge keeps drawing money, one funding round at a time.
China's transport ministry told its agencies to push AI deeper into roads, rail and ports at a July 3 party meeting, under an "AI plus transport" program tied to the 15th Five-Year Plan. It reads like boilerplate until you connect it to the lead. The state is wiring AI into the physical infrastructure it already owns, and that infrastructure runs on the grid it also owns.
OpenAI has found a way to cut its inference costs roughly in half, according to The Information. The efficiency race is not one-sided. Every gain a Chinese lab books on domestic chips, a US lab is booking too, on better ones. The open question this whole issue circles is which side runs out of a different input first.
The Bigger Picture
For two years the question everyone asked about Chinese AI was whether it could get the chips. That was the right question for the training era, when the game was assembling enough Nvidia hardware in one building to build a frontier model. It is the wrong question for the era starting now, which is about deployment, about serving models to hundreds of millions of people every day, cheaply, indefinitely.
Deployment does not run on scarce top-end chips. It runs on electricity, on cooling, on land near a substation, on the unglamorous physical plant of the compute economy. That is the layer China has spent 20 years overbuilding, first for factories, EVs and high-speed rail, and now for this. The export-control regime was built to deny China the one input it could not produce. It said almost nothing about the input China produces better than anyone on earth.
That is the caveat cutting the other way. Power abundance is an endowment, not a strategy, and endowments get squandered. China still has to make the chips, write the models, and show that 算电协同 is more than a line in a plan. But the US just spent a heat wave choosing between its air conditioners and its data centers, and China spent the same week wiring a desert full of solar panels straight into a server hall. Watch the meter, not just the fab.
I exist because this information asymmetry shouldn't. If someone forwarded you this, subscribe below, it's free, and it goes out every morning.