Machinepower Index / Theory and method

How the Index is built.And how far to trust it.

Six chapters, from the 1890 idea behind it to the limits we know about. Read this before you cite the Index.

01

After Mahan.Sea power, applied to AI.

Mahan’s test for sea power was simple: a navy is only as strong as everything behind it. Machinepower applies that test to AI, in three layers.

A navy with no harbours goes nowhere. AI with no power is the same.

Strength in one layer cannot make up for a gap in another. Each layer has four measures, each scored 0 to 10 from a published dataset or a stated rubric. How we combine them is the most important choice in the Index, so it comes next.

02

Why we don’t just average.A strength should not hide a gap.

Most indexes add their measures up or average them. That lets a strength hide a weakness.

A nation could turn out thousands of AI graduates and climb the table with no power or chips to put them to work. On an average, a high score in one place pays for a low one somewhere else. AI does not work like that, so in our Score the weakest measure pulls hardest.

One example nation. Eleven measures at 8 and one at 2: plenty of talent and models, almost no power. Not a real nation.

Firm Power Capacity, scored 2

Plain averageWhat most indexes do
7.5
Geometric meanUsed by the UN’s Human Development Index
7.1
Our ruleWeak scores count far more
5.3

On a plain average the missing power costs half a point. Under our rule it costs more than two.

03

The formula.Drag a measure and watch the Score.

Score1 ÷ √(average of 1 ÷ measure²)
Potentialaverage of Watts, Weights and Will
To gainPotential − Score
Score–
Potential–
To gain–
Rank–

Drag any measure. Lift a strong one and the Score barely moves; lift the weakest and it jumps. The four unpainted bars are the weakest links; the white line is the Score and the dashed line the Potential.

For the technically minded: the Score is a power mean with exponent −2, a CES aggregate with an elasticity of substitution of one third: a 5.2 counts more than three times as much as a 10.

04

Twelve measures.Four in each of three layers.

Showing the United Kingdom against the median of all 50

WattsThe material base

Planning & Permitting

5.2Weakest link

Median of 50: 5.9

How quickly a nation can get large AI sites and their grid connections approved and built.

Assessed Country evidence, checked against World Bank Business Ready 2025

Firm Power Capacity

5.4United Kingdom

Median of 50: 5.0

How much electricity a nation generates, and so how much compute it could power.

Ember and the Energy Institute Statistical Review

Sovereign Compute

6.7United Kingdom

Median of 50: 4.5

The supercomputing capacity a nation holds on its own soil.

TOP500, June 2026

Hardware & Silicon Access

5.8United Kingdom

Median of 50: 4.8

Whether a nation can make, or reliably buy, the chips that AI runs on.

Public fab, ASML and export-control records

WeightsThe models and the people

Frontier Model Output

5.2United Kingdom

Median of 50: 3.0

How many leading AI models a nation produces.

Stanford AI Index 2026 and Epoch AI

Alignment Capability

9.2United Kingdom

Median of 50: 5.8

How well a nation can test, evaluate and control advanced models.

Assessed Country evidence, checked against the CSET Research Almanac

Regulatory Standing

9.2United Kingdom

Median of 50: 8.0

The strength and influence of a nation’s AI rules and standards.

Oxford Insights, 2025

Talent Density

7.2United Kingdom

Median of 50: 5.0

How many of the world’s leading AI researchers and inventors work there.

Zeki Data via the Stanford AI Index 2026

WillThe politics

Policy Capacity

10.0United Kingdom

Median of 50: 8.5

Whether the state has the strategy, budget and institutions to act on AI.

Oxford Insights, 2025

Public Sector Adoption

6.9United Kingdom

Median of 50: 7.4

How far government itself puts AI to work.

Oxford Insights, 2025

Workforce Transition

8.6United Kingdom

Median of 50: 7.8

How ready a nation’s workers and labour policies are for AI.

IMF AI Preparedness Index, 2023 data, published 2024

Public Trust

4.5Weakest link

Median of 50: 6.5

Whether the public backs the build-out.

Ipsos AI Monitor 2026; KPMG and University of Melbourne, Trust in AI 2025

182 of the 600 scores in this edition (30%) are assessed against a written rubric rather than taken from a dataset. Each is flagged on its nation’s page and in the data download. A 10 means the best in the Index or full marks on the source’s own scale.

05

Does it change the ranking?The leaders hold. The rest move for a reason.

Weighting the weak link is a choice, and a fair question is whether that choice, rather than the data, decides who comes top. So we ran the whole Index all three ways.

Where the United Kingdom ranks each way

  1. Plain averageWhat most indexes do7th
  2. Geometric meanUsed by the UN’s Human Development Index7th
  3. Our ruleWeak scores count far more7th
17 of 18nations in today’s top 18 stay there all three ways
34 of 50nations move three places or fewer between the three
China and the USare the top two all three ways

Where it does matter, it is on purpose. Estonia moves most: 29th on a plain average, 43rd under ours, because its weakest measures sit far below its strongest.

Every nation, all three ways

Each bar runs from a nation’s best rank to its worst across the three. The dot is its rank under our rule. A short bar means its place holds however you combine the scores.

The ranking holds. Where a nation moves, it is because an average would have hidden its gaps.

Below the top ten, nations sit close together, so real progress shows up fast. Lift a nation’s weakest measure by a single point and, outside the top ten, it climbs 4.0 places on average. At the top, where the gaps are wider, it climbs 0.6.

06

Built on the public record.And always expanding.

SourceUsed forEdition
TOP500Sovereign ComputeJune 2026 list
Ember and the Energy InstituteFirm Power Capacity2025 data
Fab, ASML and export-control recordsHardware & Silicon AccessRubric, 2026
World Bank Business ReadyPlanning & Permitting (cross-check)2025
Stanford AI Index and Epoch AIFrontier Model Output2026
Zeki Data, via the Stanford AI IndexTalent Density2026
CSET Research AlmanacAlignment Capability (cross-check)2026
Oxford Insights, Government AI Readiness IndexRegulatory Standing, Policy Capacity, Public Sector Adoption2025
IMF AI Preparedness IndexWorkforce Transition2023 data, published 2024
Ipsos AI MonitorPublic Trust2026
KPMG and University of Melbourne, Trust in AIPublic Trust, where Ipsos does not survey2025

What’s next

The Index is rebuilt every quarter. These are the gaps we know about, and what we are doing to close them.

Richer compute data

TOP500 lists disclosed systems, so today’s compute figures are a conservative floor. Richer datasets on commercial capacity come into the Index in future editions.

A sharper read on Will

Surveys and readiness scores tend to cluster, so Will separates nations less than Watts or Weights. A procurement measure, built from what governments actually buy, is in development.

Where records are thin

Some nations publish little. Where the record runs short, as with China’s Public Sector Adoption, the score is assessed against our written rubric and flagged.

An advisory board

Assessed scores follow a written rubric, which we will publish. An advisory board is being formed to review and challenge them. The strategic plays are editorial and never scored.

Every source is the latest edition and no more than three years old. Spotted an error? Email elliot@machinepowerindex.org with the nation, the measure and the source; confirmed fixes land in the next edition.