Machinepower Index / Methodology
How the Index is built.
Three layers, twelve measurements, one number — and an open account of how much that number depends on each choice behind it. If you plan to cite the Index, read this first.
The formula
Potential is the plain average of Watts, Weights and Will. Score combines the twelve measurements beneath them so that one weak number pulls the whole down rather than being averaged away. The gap between the two is what being lopsided costs: nothing if a nation is evenly built, more the further its strengths and weaknesses sit apart. In formal terms the Score is a CES mean of the twelve cells with an elasticity of substitution of one third — explained below.
Why not average
Every comparable index adds its dimensions up or averages them, which lets a strong score in one place cancel out a weak score in another. That is the wrong shape for sovereignty. A world-class model is no use on a grid that cannot power it, and an excellent grid is no use with no models to run on it. A nation needs all three at once.
Economists have a standard dial for how far one input can stand in for another: the elasticity of substitution. Set it to 1 and you get a geometric mean. Set it to 0 and you get a pure minimum — Liebig's law, where output equals the weakest input and nothing else counts. The Index sits at one third: the layers stand in for each other only weakly, without throwing away everything but the worst number. That puts it among the composite indexes built to stop strengths cancelling weaknesses; the Mazziotta–Pareto Index does the same job by a different route.
Why twelve cells rather than three axes
A chokepoint hides inside an average. Taiwan scores 10.0 on hardware and silicon and 3.0 on frontier models; average those into one Weights figure and both facts vanish, leaving a forgettable middle number. Across the Index, the nations playing Chokepoint Leverage look flatter than average at layer level and spikier one level down. Scoring on the layers would reward them for a balance that averaging invented.
The three layers are still the plain average of their four measurements, and still how the site describes every nation. Only the Score is worked out one level down.
Robustness
The obvious objection to any scoring rule is that the setting was picked to produce the answer we wanted. These tables are the reply. Every figure here is regenerated from the data at each quarterly refresh.
1. Every aggregation rule, from average to pure minimum
| Rule | σ | ρ vs average | Moves | Top five |
|---|---|---|---|---|
| Plain average (no penalty) | — | 1.000 | 20 | United States, China, South Korea, Germany, France |
| Geometric mean | 1.00 | 0.988 | 18 | United States, China, South Korea, France, Germany |
| CES | 0.50 | 0.959 | 14 | United States, China, South Korea, France, Japan |
| CES — published (current) | 0.33 | 0.921 | — | United States, China, South Korea, France, Japan |
| CES | 0.25 | 0.884 | 14 | United States, China, South Korea, France, Japan |
| CES | 0.14 | 0.850 | 15 | United States, China, South Korea, France, Japan |
| Pure minimum (Liebig) | 0.00 | 0.796 | 20 | United States, France, South Korea, China, United Kingdom |
ρ is the Spearman correlation of the resulting ranking against a plain average; lower means the rule genuinely reranks. "Moves" counts nations that would change rank against the published edition.
2. Sensitivity to the single weakest cell
Because a low measurement pulls hard, the Score depends more on a nation's weakest judgement than on any other. These are the ten most sensitive cases: what happens to a nation's rank if its lowest measurement were scored a full point higher or lower.
| Jurisdiction | Weakest cell | Rank | −1 | +1 | Swing |
|---|---|---|---|---|---|
| Germany | Frontier Model Output 3.0 | 8 | 19 | 5 | 11 |
| Singapore | Firm Power Capacity 3.5 | 11 | 19 | 8 | 8 |
| India | Frontier Model Output 4.0 | 9 | 16 | 8 | 7 |
| Taiwan | Frontier Model Output 3.0 | 12 | 19 | 6 | 7 |
| Israel | Firm Power Capacity 3.9 | 10 | 17 | 8 | 7 |
| United Kingdom | Frontier Model Output 4.2 | 6 | 12 | 6 | 6 |
| Italy | Frontier Model Output 3.0 | 17 | 19 | 11 | 6 |
| Netherlands | Frontier Model Output 3.0 | 15 | 19 | 9 | 6 |
| Canada | Frontier Model Output 4.2 | 7 | 12 | 6 | 5 |
| United Arab Emirates | Hardware & Silicon Access 4.0 | 13 | 18 | 10 | 5 |
Median swing across all 26 ranked nations is 5 places; the largest is 11. A full point is a large revision on a one-decimal rubric — most quarterly corrections are a tenth or two — so read this as an outer bound rather than an expected error.
The honest conclusion: an exact rank in the middle of this table is not reliable, and we would rather say so than imply a precision the data cannot support. Twelve nations sit within half a point of each other, so changing one measurement moves several places at once. Treat the leading group and the tiers as findings; treat a mid-table nation's exact position as noise. It is also why the Index gives you a Score, a Potential and a strategic play rather than a bare league position.
3. What is judged rather than measured
97 of 324 measurements (30%) are our own judgement rather than a number taken straight from a published dataset. Each one is flagged on its country page. They cluster where no dataset measures the thing cleanly — how fast a country permits and builds, how good it is at alignment, how well it is retraining its workforce. This is the honest weak point of the Index, not a rounding detail.
Where the idea comes from
The Index takes its shape from Alfred Thayer Mahan’s argument, in The Influence of Sea Power upon History (1890), that maritime power had two separate sources: what a nation could build and crew for itself, and which positions it held that everyone else had to pass through. Neither reduces to the other. A country could have a large fleet and no positions, or hold a position and have almost no fleet.
That is why there are two measures here, kept apart. The Score covers the first: capability a nation controls itself. The chokepoint register covers the second: position others cannot route around. Adding them would bury the thing worth seeing. Ukraine is the clearest case — last of the twenty-six on capability, and holding a dataset its allies cannot reproduce.
Mahan’s positions came with a warning, and it applies here. A position is only a chokepoint for as long as it can be held. Until 2022 two Ukrainian plants supplied roughly half the world’s semiconductor-grade neon; that trade now runs mainly through China. Chokepoints are not permanent, which is one more reason not to fold them into a Score.
Chokepoints, and why they are not scored
The Score measures sovereign capacity: what a nation can build, control and run on its own. It rewards balance, because sovereignty requires all three layers at once. That is a measurement choice, not a strategy recommendation. For most nations, becoming indispensable in one layer is a better bet than being adequate in three — which is why every dossier carries a strategic play alongside its Score, and why a low Score under a well-played specialisation is not a failure.
A chokepoint is the other kind of power: not what you can do alone, but what others cannot do without you. It behaves differently from sovereignty, and it does not sit inside the twelve cells. Our cells measure a nation's access to a capability; leverage is its control over everyone else's access. The Netherlands has moderate access to silicon and total control of the machines that make it. Those are different quantities, and occasionally opposite ones.
So chokepoints are listed rather than scored. Building one AI chip takes six steps, and each one is controlled by a single country; the full chain is set out on the leverage page, and each holder carries the detail on its own dossier. There is no leverage number and no leverage ranking, because producing one honestly would need data on how critical each chokepoint is and how long it would take to build around — and that data does not exist per country today. Deriving it from the twelve cells instead produces nonsense: on a peak-cell reading, Indonesia ranks eighth and Taiwan nineteenth.
A chip is not the only thing that can be withheld, and not everything withheld is a chokepoint. The register distinguishes two kinds. A chokepoint is leverage so complete that others’ systems stop when it closes; eleven sit in the chain that produces an accelerator, and a twelfth sits outside it — access to American frontier models, a lever the Commerce Department pulled in June 2026 when it ordered a named model licensed for any foreign person and its maker switched it off worldwide. An exclusive asset is a step below: something nobody else holds that others pay to reach, but can live without. Ukraine’s Avengers Labs battlefield dataset is the register’s one listed asset — licensed under Cabinet of Ministers Resolution No. 310 and exercised on 24 August 2026 by admitting Britain first. Both rungs pass the same two-limb test: one holder, one working government lever. Confining the register to silicon — or to chokepoints alone — would have been a choice about where to look, not a finding about where power sits.
Something is listed only where a nation dominates it and its government could actually shut it off. A company monopoly is not national power: ASML counts for the Netherlands because Dutch export licensing governs it, while Arm does not count for the United Kingdom, since Whitehall has no comparable lever and the company is Japanese-owned. Seven nations control at least one. Nineteen control nothing, which is itself a finding. Entries are graded: demonstrated where the state has restricted supply and it worked, structural where dominance is clear but the lever is untested or legally unavailable. A July 2026 audit dropped three entries that failed the dominance test and merged a fourth.
Sources
Every source below is the latest edition available and no more than three years old, and its date is shown. 50 separate sources sit behind this edition, and every country page shows which ones its numbers came from. The main ones are TOP500, Epoch AI, the Stanford AI Index, WIPO, Oxford Insights, OECD.AI, JLL and the Energy Institute.
AI Singapore · ASML · AUKUS (Defence) · CSET Emerging Technology Observatory · CSIS Reports · Council of the EU · DIGITIMES · Data Centre Dynamics · Dept. of Industry, Science and Resources · Ember · Energy Institute Statistical Review · Epoch AI · EuroHPC JU · EuroHPC Joint Undertaking · European Commission (InvestAI) · Garante per la protezione dei dati personali · GitHub Octoverse · IEA · IMF · ISED Canada · IndiaAI Mission · Ipsos AI Monitor 2025 · Ipsos AI Monitor 2026 · Malaysia Budget 2026 · Ministry of Defence of Ukraine · NIST / IMDA · NVIDIA · NVIDIA / MSIT Korea · OECD.AI Policy Observatory · OpenAI / G42 · Oxford Insights · Oxford Insights / Brave1 · Oxford Insights / Diia.City · Oxford Insights GAIRI 2025 · Public Investment Fund · Rapidus · SDAIA · SaluteDevices (Sber) · Stanford AI Index 2026 · TOP500.org · TSMC · Taiwan Energy Administration (MOEA) · TenneT (Grid Operator) · The White House · UK Compute Roadmap · UK Government / MOFA Japan · US State Department · WIPO Technology SPARK · World Bank · YTL Power / NVIDIA
Known limitations
TOP500 undercounts commercial capacity. It measures disclosed research systems, so Sovereign Compute should be read as a floor rather than a ceiling — particularly for nations whose capacity sits inside hyperscaler estates.
Will discriminates far less than the other two layers, and the numbers say so. Watts and Weights rest on hard counts — terawatt-hours, petaflops, notable models. Will rests on surveys and readiness scores. The result is measurable: across the 26 ranked nations, Will has a standard deviation of 0.56 and spans 6.3–8.8, against 1.35 and 3.6–9.2 for Watts and 1.15 and 4.4–9.2 for Weights. Roughly half the spread, sitting a full point higher. So Will compresses differences rather than revealing them, and it flatters nations with thin infrastructure. Under a rule this weighted towards the minimum, that means Will rarely decides a Score except at the top of the table, where everything else is already high. Harder proxies — procurement actually spent rather than committed, technical headcount inside the civil service, counts of live deployments — would discriminate better. None of them yet exists on a comparable basis across all 26 nations, which is why this layer is the most likely to change.
Documentation is not adoption. Oxford Insights scores internationally documented evidence. China documents little, so its Public Sector Adoption cell likely reflects opacity rather than the actual reach of Chinese state deployment, which is among the most extensive anywhere.
One pair of hands. Every judged measurement is the work of one author, against a published set of rules. The strategic plays are editorial throughout and are never scored at all.
Q3 2026 is the first measured quarter. Country pages carry a direction-only outlook. Numeric quarter-on-quarter movement publishes at the Q4 2026 refresh.