Topics / Compute & Power
Compute & Power
The binding constraint on AI is quietly shifting from silicon to electricity. Training a frontier model is a one-time burst; serving it to hundreds of millions of people runs indefinitely on power, cooling and land near a substation. China generates more than twice the electricity of the United States and is adding new capacity many times faster, and it is increasingly wiring that capacity straight into the data centers that will drink it. This is the input the export-control regime did nothing to deny, and the one China produces better than anyone, and we track that contest with dates and sources.
The state of play
- How big the compute base actually is, in the only figures the state publishes. The Ministry of Industry and Information Technology put China’s intelligent computing capacity at 2,185 EFLOPS as of the end of June 2026, up 177 percent year on year, against the more than 1,590 EFLOPS and 42 ten-thousand-card clusters the same ministry reported in January 2026, and IDC’s count of 725.3 EFLOPS for 2024 and 416.7 for 2023. Read the growth rate against what sits behind it. MIIT’s chief engineer Wang Weiming put the overall rack-occupancy rate of national computing facilities at 71.4 percent in the same July briefing, so close to three tenths of built capacity was not yet in service, and that rate records whether equipment is installed rather than how hard it runs, which makes it a ceiling on utilisation and not a measure of it. Two more cautions belong on the number itself: EFLOPS totals sum across precisions and vendors and are not audited, so they are the state’s own accounting of scale rather than a like-for-like comparison with any single Western cluster, and the 2023 and 2024 points are IDC’s count while the later ones are the ministry’s, so the series is not measured by one hand throughout (MIIT, Xie Cun and Wang Weiming at the State Council Information Office briefing on first-half 2026 industry and IT development, 20 July 2026, via 中新网 and 新浪财经; MIIT via 新华社, 21 January 2026; IDC via 新华社).
- The scarce input is moving from chips to power, and one week in June 2026 made it legible. The IEA expects the electricity going to data centers worldwide to roughly double by 2030, from about 485 to about 950 terawatt-hours, growing near 15 percent a year, more than four times faster than all other demand. On 26 June China switched on Zhongwei, its first AI data center running entirely on green power, a dedicated 500-megawatt wind-and-solar project wired straight into the servers with no public grid in between; four days later, on 30 June, the US Department of Energy let the grid operator PJM order large data centers off the public grid onto backup generators on fifteen minutes’ notice, to keep home air-conditioning running through a record heat wave near a 166-gigawatt peak (IEA, “Energy and AI”; CGTN; US DOE via Utility Dive, via CAD).
- China’s generation and build-out lead is not marginal. It generated about 10,161 terawatt-hours of electricity in 2024 against the United States’ 4,393, more than the US, the EU and India combined, and in 2025 it added roughly 540 gigawatts of new capacity against about 63 in the United States. BloombergNEF expects China to add several times the US total again over the next five years. Abundant power does not fabricate a leading-edge chip China still cannot make at parity, so treat this as an edge in one input, not the whole race (Voronoi, citing Ember; CarbonCredits; BloombergNEF via Al Jazeera, via CAD).
- Compute-power coordination is becoming policy, not just endowment. China wrote 算电协同 (planning data centers and the power system as one machine) into its new five-year energy plan, which targets 80 percent renewable power for data centers by 2030; a Beijing storage firm and a Shanghai compute operator signed a joint compute-and-power deal in early July, and Shanghai’s development commission is letting large new compute loads pair directly with dedicated renewable supply. It is still more ambition than norm: Zhongwei is a demonstration, and Chinese data centers drew only about 11 percent of their power from renewables in 2023, so the 80 percent target is a long climb (Dialogue Earth; Sina Finance; via CAD).
- The deployment half of the race is where cheap power converts to advantage, and a Chinese firm has claimed it can now serve AI on domestic silicon at a profit. SenseTime’s compute arm told a WAIC infrastructure forum in mid-July that it had turned the gross margin on its domestic-chip inference business positive, with daily token throughput running from 400 billion at the start of 2026 toward a targeted 2.42 trillion by end-July and 10 trillion by year-end, roughly twenty-five times where it began, by pairing one scarce high-end chip with about thirty domestic ones. Read it carefully: a gross margin is not a net profit, the disclosure is SenseTime’s own and unaudited, it depends on a high-end-chip scarcity window that can close, and only six or seven of the twenty-plus domestic chips it adapted clear the bar. National demand is real behind it: the National Data Administration put China’s daily AI-token consumption above 30 trillion as of the end of June 2026, against roughly 100 billion a day in early 2024 (SenseTime, via 量子位 and IT之家; National Data Administration, via tmtpost, via CAD, “Year One” and “The Collateral”).
- China can now stand up sovereign compute at frontier scale without a single Western chip, and it is the physical form of a market that has flipped. In July 2026 it switched on its first fully domestic 100,000-accelerator AI supercluster, Sugon’s 8000, nicknamed Dengfeng, at the Zhengzhou core node of the national compute grid; the National Development and Reform Commission confirmed it is running, already carrying more than 300 jobs across 26 scientific fields from new materials to drug discovery, every layer of it Chinese down to the immersion coolant, with no Nvidia or AMD parts. A hundred thousand accelerators is roughly the scale xAI and OpenAI train frontier models at, and it lands as domestic AI chips are widely reported to have climbed from near zero at the controls’ start toward a rough 41 percent of China’s market in 2025 while Nvidia’s fell from about 95 to 55 percent, though that precise split traces to a single estimate circulating on Chinese channels and should be read as directional, not audited. Two asterisks matter more than the milestone. Sugon calls the design “super-intelligence fusion,” one machine running FP64 double precision down to INT8, so today’s workload leans scientific simulation, not language-model training the way xAI’s Colossus does; and per chip, domestic parts still trail Nvidia’s top silicon and the CUDA stack everything is written against, so standing up 100,000 of them is a scale-and-supply-chain feat, not one-for-one parity with an H200. The same limit showed in June, when the CPU-only LineShine topped the global TOP500 at 2.198 exaflops on roughly 47,000 domestic CPUs and no GPUs yet placed only fourth on HPL-MxP, the benchmark that tracks the low-precision math AI training actually uses. Domestic power and compute are closing the deployment gap, not yet the frontier-training one (State TV and NDRC; 量子位 and IT之家; IDC estimate via Bilibili; TOP500 and Tom’s Hardware, via CAD, “Full Domestic”).
Best issues on this topic
- Camera Glasses Lost 78 Percent of Their Revenue as AR Glasses Grew 70
RUNTO (洛图科技), which tracks smart-glasses retail across JD.com, Tmall and Douyin, published July 2026 figures showing the category fall for the first time: 74,000 units, down 15.3 percent year on year, for 130 million yuan, down 25.8 percent, at an average price of 1,751 yuan, down 12.4 percent. RUNTO attributes the fall to the calendar, because June was the 618 shopping festival and last July was the second-highest month in thirteen. The composition underneath it is the part that is not seasonal. AR glasses, the ones with a display in the lens, sold 33,000 units for 98.28 million yuan, up 72.7 percent by volume and 70.0 percent by revenue, at an average price of 2,990 yuan that did not move all year. Audio-only glasses sold 13,000 units for 8.7 million yuan, down 33.5 percent by revenue at 656 yuan. Camera glasses, the no-display design Meta made familiar with Ray-Ban, sold 28,000 units for 23.1 million yuan, down 46.6 percent by volume and 77.9 percent by revenue, with the average price down 58.5 percent to 819 yuan. The three segments sum to 130.08 million yuan, matching the reported total, and the prior-year segments sum to 175.41 million, which returns the 25.8 percent decline to the digit, so the shift is not a rounding artefact. AR went from 33.0 percent of category revenue to 75.6 percent in twelve months while camera glasses went from 59.6 percent to 17.8 percent. The two leaderboards have come apart: the largest volume share in July, 17.0 percent, went to 清野, a cheap camera-glasses newcomer, while by revenue the order is Rokid, RayNeo, XREAL, Qwen and Huawei, with the top five taking 65.8 percent. Against that, Alibaba previewed a displayless pair called the Qwen N1 at the Bund Summit, with iris recognition at a claimed false-accept rate below one in a million and eye tracking meant to authorise payment by looking at a terminal. In the Briefing: Bank of China put at least 300 billion yuan behind the compute industry at the 2026 China Computing Power Conference, with China Telecom, China Mobile, China Unicom, PICC, China Orient Asset Management and CDB Financial Leasing, targeting at least 100 intelligent computing centre projects and at least 3,000 compute-consuming firms across the 15th Five-Year Plan, while ICBC took its compute-token loan national; China Telecom’s research institute forecast ten quintillion tokens consumed this year rising to more than 3,500亿亿 by 2030, two endpoints its own stated compound rate does not fit, and put inference at 80 percent of China’s compute market by 2029 with 2026 AI capex at leading Chinese tech firms near 600 billion yuan; XPeng’s humanoid production line began operating and the first IRON walked off it under its own power, carrying 76 degrees of freedom and three Turing chips for a claimed 2,250 TOPS; and 25 Fields medallists including Terence Tao and Yu Deng signed a statement arguing that using mathematical problems as capability benchmarks damages the field.
- China’s Compute Grew 177 Percent. Its 2030 Plan Asks for 40.
MIIT issued its five-year plan for the information and communications industry on 7 September, thirteen headline indicators across twenty-six tasks and six areas, and set intelligent computing capacity at 9,800 EFLOPS by the end of 2030. The base it grows from is 2,185 EFLOPS at the end of June 2026, measured at FP16 and up 177 percent year on year, so moving to 9,800 over four and a half years works out at roughly 40 percent a year compounded against a trailing twelve months that ran at 177. The plan asks for the orderly deployment of ten-thousand-card and hundred-thousand-card clusters, and MIIT put occupancy across national computing facilities at 71.4 percent at the same July briefing, so close to three in ten of the capacity already built is not in service. Three of the thirteen indicators are efficiency or carbon measures, including the power usage effectiveness of newly built large and hyperscale facilities. The money side puts cumulative information infrastructure investment at 3.8 trillion yuan across 2026 to 2030 and industry revenue at 4.1 trillion by 2030; at the onshore rate of 6.7105 yuan to the dollar on 8 September that investment figure is about 566 billion dollars, where English coverage rendered it as 532 billion, implying a rate near 7.14 that nothing traded at in that week. Chinese banks spent August building collateral for it: Agricultural Bank put a Token loan scheme into Shanghai and Zhejiang, Bank of Chengdu ran a pure credit product capped at 5 million yuan per borrower and backed by compute vouchers, and Zhangjiagang Rural Commercial Bank set aside a 2 billion yuan facility, underwritten on token consumption, compute service contract value and receivables from compute work.
- The Paying Tier
Chinese cloud vendors are reselling competitors’ models at roughly 30 percent of list price, below their own cost, because in AI coding the customer names the model. Leiphone reported the discount ladder on 21 August from named cloud sales sources: 60 to 80 percent of list six months ago, 48 percent now common, roughly 30 percent at the low end. The same day, Approaching.AI vice-president Guan Jiawei told the South China Morning Post that domestic chips can only serve the low-quality inference tier, the one with weak demand and weak monetisation, and that high-quality tokens, especially coding, still depend on Nvidia. iFlytek said the same week that a flagship model trained entirely on domestic compute would reach the domestic first tier on code capability.
- The Second Gap
Loongson chairman Hu Weiwu told the company’s 25th-anniversary meeting in Beijing on 19 August that its taped-out 3B6600 processor runs on a 1Xnm domestic process and performs at the level of mainstream x86 CPUs built on 7nm or better. The claim is Loongson’s own, made at Loongson’s own event, with no independent benchmark published and no workload named. Its audited filings show FY2025 revenue of 635.3 million yuan, about 89 million dollars, against a 455.1 million yuan net loss attributable to shareholders, a third straight nine-figure annual loss. Hu said the main contradiction has moved from research to sales, and that 2025 to 2027 is a transition from the policy market to the open market.
- The Collateral
A Guangzhou district launched Token贷, a Bank of China loan that underwrites AI companies on tokens burned rather than property owned, with lines to 30 million yuan on four non-property inputs. The trial tranche is about 28 million yuan, but CITIC Bank and Bank of Guangzhou have followed, and Beijing’s E-Town and Anhui have written token lending into policy. The National Data Administration puts national token consumption above 30 trillion a day as of end-June.
- Full Domestic
China switched on its first fully domestic 100,000-chip AI supercluster, and the National Development and Reform Commission confirmed it is running at the Zhengzhou node of the national compute grid. Sugon’s 8000, nicknamed Dengfeng, is Chinese at every layer, the accelerators, the interconnect and the immersion cooling, with no restricted imports; the honest caveats are that its workload today leans on double-precision science rather than model training and that per chip it still trails Nvidia on raw performance and the CUDA stack.
- The Grid
The US pulled data centers off its grid to keep home AC running through a heat wave. The same week, China wired a desert of solar straight into a server hall. The scarce input is shifting from chips to power.
- Year One
SenseTime told a WAIC forum it had turned the gross margin on its domestic-chip inference business positive, with daily token throughput up roughly 25x through 2026. The same issue tracked Apollo Go’s expansion into Kazakhstan.
- The No-GPU Machine
A Chinese system topped the global TOP500 at 2.198 exaflops on 47,000 domestic CPUs and zero GPUs. On the AI-relevant benchmark, it came fourth. The asterisk is the whole export-control story.
- 60,000 Chips, Zero From Nvidia
China’s national computing hub in Zhengzhou runs 60,000 accelerators made by Sugon, a firm tied to the Chinese Academy of Sciences, with zero chips from Nvidia, AMD or any Western supplier.
- The Interconnect
Zhongji Innolight, the world’s largest optical-transceiver maker and supplier of more than half of Nvidia’s optical modules, pulled off Hong Kong’s largest IPO in seven years at about 6.8 billion dollars, then fell 8 percent on debut as the AI-infrastructure trade repriced. The picks and shovels of the American AI boom, it turns out, ship from Suzhou.
These link to the full issues on the newsletter. New pieces on this topic go out in the daily first.
Common questions
How big is China’s computing power market?
The headline number China’s government publishes is capacity, not revenue. The Ministry of Industry and Information Technology put national intelligent computing capacity at 2,185 EFLOPS as of the end of June 2026, up 177 percent year on year, having reported more than 1,590 EFLOPS and 42 ten-thousand-card clusters in January 2026; IDC counted 725.3 EFLOPS for 2024 and 416.7 for 2023, so the series is not measured by one hand throughout. Two caveats matter as much as the level. MIIT’s chief engineer put the overall rack-occupancy rate of national computing facilities at 71.4 percent in the same July 2026 briefing, so a meaningful share of built capacity was not yet in service, and that rate records whether equipment is installed rather than how hard it runs. And EFLOPS totals sum across precisions and vendors and are not audited. A revenue figure in yuan for the market as a whole is not something we can trace to a named, current official release, so we do not publish one.
Why is electricity becoming the binding constraint on AI?
Because the race is moving from training to deployment. Training a model is a one-time burst of compute, but serving it to hundreds of millions of people runs indefinitely on electricity, cooling and land near a substation. The IEA expects data-center electricity worldwide to roughly double by 2030, growing about 15 percent a year, more than four times faster than all other demand. Whoever has the most cheap, spare power to pour into compute wins the half of the race that comes after the model is built.
Does China really have more electricity than the United States?
Yes, by a wide margin. China generated about 10,161 terawatt-hours in 2024 against the United States’ 4,393, more than the US, the EU and India combined, and in 2025 it added roughly 540 gigawatts of new capacity against about 63 in the US, increasingly next to the data centers that will use it. The caveat cuts the other way: abundant power does not fabricate a leading-edge chip China still cannot make at parity, and it does not close the gap on the best models. It makes Chinese AI cheaper to run, not automatically better.
What is 算电协同 (compute-power coordination)?
It is the idea, now written into China’s five-year energy plan, that data centers and the power grid should be planned and scheduled as a single system rather than built separately. In practice that means wiring generation directly to compute, as at the Zhongwei desert data center where a dedicated 500-megawatt wind-and-solar project feeds the servers with no public grid in between, and letting large new compute loads pair with their own renewable supply. It is still more ambition than norm, since Chinese data centers drew only about 11 percent of their power from renewables in 2023.
Does cheap power mean China is winning the AI race?
Not on its own. Power abundance is a real edge in one input, the deployment and serving layer, and it is the input the US export-control regime did nothing to deny. But it is an endowment, not a strategy: China still has to make the advanced chips, write the frontier models, and prove that all-domestic clusters pay outside a chip-scarcity window. The honest read is that China starts ahead on electricity and behind on leading-edge silicon, and the contest turns on which constraint binds first.