Machinepower Index / Dossiers / China
Rank 1–2 across the scoring rules tested. How this is tested
China
Total Sovereignty — For the two nations able to attempt every layer alone: own the whole stack, from energy and chips to frontier models and the rules, while depending on each other somewhere.
How the Score, Potential and bottleneck cost are worked out: methodology.
The three layers
Energy, grids, data centres and silicon — the physical capacity to run models at scale.
Compute tracked by Epoch AI: China controls 430,000 H100-equivalents, wherever the sites are. Hosted compute is shown, not scored.
Models, alignment capability, standards influence and the talent that produces them.
The capacity to wire AI into the state and the economy while holding public consent.
Where the score comes from
| Layer | Sub-dimension | Basis | Score |
|---|---|---|---|
| Watts | Planning & Permitting | Assessed | 9.8 |
| Firm Power Capacity | Dataset | 9.8 | |
| Sovereign Compute | Dataset | 8.7 | |
| Hardware & Silicon Access | Dataset | 6.2 | |
| Weights | Frontier Model Output | Dataset | 9.1 |
| Alignment Capability | Assessed | 8.0 | |
| Regulatory Standing | Dataset | 9.1 | |
| Talent Density | Assessed | 8.6 | |
| Will | Policy Capacity | Dataset | 9.3 |
| Public Sector Adoption | Assessed | 7.0 | |
| Workforce Transition | Dataset | 7.2 | |
| Public Trust | Dataset Sources differ | 9.0 |
Assessed marks a measurement we judged rather than took from a dataset: 4 of the 12 here.
Sources differ marks a cell where two sources disagree by more than a point. Public Trust: Ipsos and Melbourne/KPMG differ by more than a point (9.5 against 8.4); the cell is their average.
Sources
Share of adults agreeing that products and services using AI have more benefits than drawbacks; the Index scores public trust as 1 + share ÷ 10. Survey of 23,532 adults across 32 countries, March to April 2026.
View source →CnAISDA, China's counterpart to an AI safety institute, was launched in February 2025 by Tsinghua, Peking University, Shanghai AI Lab and others, and Shanghai AI Lab's frontier-risk evaluations now converge with US and UK methods. Analysts judge frontier safety a secondary priority to development, and China produces about 57% of US AI-safety research output with fewer citations per paper.
View source →Oxford Insights scores China's public-sector adoption at 49 of 100 and notes that limited transparency and restricted access to deployment data mean its measure likely understates China. Reported deployment is broad: at least 72 local governments adopted DeepSeek within two months of its release, and Shenzhen deployed it across every government department.
View source →LineShine (National Supercomputing Centre Shenzhen), a CPU-only indigenous system, debuted at #1 on the June 2026 TOP500 at 2.198 exaflops, ending El Capitan's reign.
View source →WIPO Technology SPARK (July 2026) records China publishing over 43,000 of the 56,000+ generative AI patent families filed in 2024-25, with the field growing at a 64% compound annual rate.
View source →Epoch's database shows Chinese open-weight models (DeepSeek, Alibaba Qwen) trailing the US frontier by ~7 months on average, with the top-model performance gap narrowed to 2.7% per the 2026 AI Index.
View source →Used for the workforce measurement. The IMF combines schooling, pay, labour mobility and active labour-market policies; the Index rescales it so the weakest of 173 economies scores 0 and the strongest 10.
View source →A cross-check for alignment, which stays assessed. The Research Almanac counts English-language research only, so it understates Japanese, Korean and Chinese work.
View source →Fossil, nuclear, hydro and other dispatchable plant; averaged with generation for firm power.
View source →Second source for public trust, averaged with Ipsos.
View source →Capital is shown beside the Score, not scored.
View source →On the register — what the world cannot get anywhere else
China holds two of the register's chokepoints: positions nobody else can route around, which its government could actually shut off. See the whole chain.
China processes the large majority of separated rare earths, the inputs for the magnets in manufacturing equipment, robotics, drones and defence hardware. Beijing has already restricted their export, which converted a commercial position into an instrument of state.
US Geological Survey →Both are needed for advanced semiconductors and optoelectronics, and both went under Chinese export licensing in 2023. A second chokepoint, and one Beijing has already shown it is willing to close.
US Geological Survey →Bottlenecks
These hold China below its Potential. Its two lowest scores cost it the most: hardware & silicon access at 6.2 (part of Watts) and public sector adoption at 7.0 (part of Will).
Assessment
China builds infrastructure at a scale no one else can match and its frontier models now rank close to America’s, so its real constraint is silicon: access to leading-edge chips and the tools to make them, which every new export control tightens.
The art of the possible
China's path runs through silicon: each indigenous lithography and accelerator breakthrough converts its unmatched build-speed and state diffusion into genuine full-stack sovereignty. If domestic chips close on the frontier, its bottleneck largely dissolves.
Strategic play — Total Sovereignty
Close the silicon gap before the rest of the chain closes on it: put the state's build speed behind indigenous lithography and accelerators, keep the materials beneath everyone else's chain Chinese, and remember that every gallium restriction also pays for the substitutes that would end it.
Leverage — the nearest position. Holds the floor of the chain — gallium, germanium and rare-earth processing — and has restricted gallium and germanium since 2023.
Open China in the live Index → — every source behind every number, and the comparison tool.