SI Index: Methodology (v1.1)
Reference date: 30 September 2026. Calculated: 18 November 2026. 2026 is year zero. Version 1.1 changes no score, rank or status: it adds the evidence behind coded scores, the status of every zero, source vintages and clearer definitions.
What the index measures
The SI Index measures how ready each nation is for the age of superintelligence. It combines eight pillars, each weighted equally (12.5%). Scores run from 0 to 100.
| Pillar | Indicators | What they capture |
|---|---|---|
| Compute | CMP-B | AI compute commercially available in the cloud (OECD.AI availability zones, weighted by tier) |
| Energy | NRG-A, NRG-B | Electricity generation per person; growth in generating capacity (Ember) |
| Talent | TAL-B | Highly cited AI research per million people (OpenAlex, top 10% most cited, 2021-2023) |
| Education | EDU-A, EDU-B | Share of top performers in PISA 2025; how far AI is built into schools (coded) |
| Capital | CAP-A (60%), CAP-B (40%) | Private AI investment as a share of GDP (Stanford AI Index); depth of capital markets and private credit (World Bank) |
| Adoption | ADO-A | Share of the working-age population using AI (Microsoft AI Diffusion Report) |
| Institutions | INS-A, INS-B | Regulatory quality and government effectiveness (Worldwide Governance Indicators 2025) |
| Sovereignty | SOV-A, SOV-B | Notable AI models developed (Epoch AI, last three years); control over the critical semiconductor supply chain (coded) |
Within a pillar, indicators are weighted equally, except Capital.
Scope of the Compute pillar. CMP-B measures AI compute that is commercially available in the cloud: the availability zones listed by OECD.AI, weighted by tier. It does not measure all the compute installed in a country, such as national supercomputers, research clusters or private data centers that are not sold as cloud services. A Compute score of 0 means that no listed availability zone was found, not that the country has no GPUs or no data center able to run AI. 131 of the 164 economies score 0 on this pillar. A second indicator measuring AI compute physically installed on the territory (CMP-A) is planned for a v1.x version: it will complete the Compute pillar without changing the eight pillars, and past editions will be recalculated.
How scores are built
- Each indicator is normalised to 0-100 against fixed bounds (goalposts), set before any calculation. Some indicators use a logarithmic scale. Values outside the bounds are clipped to 0 or 100; on a logarithmic scale, a value at or below the lower bound (including 0) scores 0.
- A pillar score is the average of its available indicators. The global score is the average of the eight pillars.
- Missing data is never counted as zero. A missing pillar is shown as “Not available”.
Indicators built from two measures
- NRG-B combines capacity growth over five years (%, linear scale, 0 to 100) and capacity added per person (GW per million people, logarithmic scale, 0.01 to 2). Each measure is normalised separately and rounded to one decimal; NRG-B is the average of the two. When only one is available, NRG-B is that one.
- CAP-B combines stock market capitalisation and private credit, both as a share of GDP (World Bank). Both use the same bounds (10% to 300%, logarithmic scale). Each is normalised separately and rounded to one decimal; CAP-B is the average of the two, or the one available.
- CMP-B counts cloud availability zones with a weight per tier (tier 1 = 1, tier 2 = 0.5, tier 3 = 0.25), in absolute terms, not per person. The weighted count is normalised on a log(1 + x) scale from 0 to 100.
- INS-A and INS-B use the published Worldwide Governance Indicators scores (0 to 100) as they are.
Every published score can be recomputed from the raw values and the goalposts below; the site’s automated tests do this for every economy and indicator.
Coverage and status
The index covers 164 economies: every economy in World Bank data with at least one million inhabitants, plus Taiwan, plus the members of the OECD or the European Union with fewer than one million inhabitants (Iceland, Luxembourg and Malta), whose statistical coverage is complete. Economies follow World Bank data (plus Taiwan), without prejudice to their political status.
| Status | Rule | Display |
|---|---|---|
| Experimental score | 8 of 8 pillars | Scored and ranked |
| Indicative profile | 7 of 8 pillars | Scored, not ranked |
| Insufficient data | 6 pillars or fewer | No global score |
| Not measured | Under one million inhabitants, outside the OECD and the EU | Not measured |
Very small economies can show extreme per-person figures (electricity in Iceland, finance in Luxembourg). Goalposts cap every indicator at 100, so these figures cannot inflate a score beyond the scale.
Coded indicators
Two indicators are coded by hand from public sources, using written grids fixed before coding.
EDU-B: AI in schools. Four components: curriculum (35%), teacher training (30%), deployment of official AI tools for students (25%) and monitoring (10%). Each is scored 0, 25, 50, 75 or 100. Every score cites a dated source; when evidence is incomplete, the lower proven level applies.
- 41 countries are coded in detail.
- 75 countries are coded from desk research.
- 48 countries are not yet coded.
SOV-B: semiconductor sovereignty. Design capability (up to 25 points), control of critical links in chip manufacturing, from EUV lithography to advanced packaging (up to 25), and resilience of access to advanced chips (up to 20). The total is converted to a score out of 100.
- Critical links are ranked in two tiers. A tier A link has a single or near-single supplier for leading-edge AI chips, not replaceable within five years (20 points). A tier B link is dominant but replaceable within two to five years (10 points). An asset that is both controlled and located in the country counts in full; an asset only located there, or only controlled from there, counts for half.
- Resilience starts from the country’s access regime to advanced chips at the reference date: 14 for free access, 12 with strengthened anti-diversion controls, 10 under licence, 4 when leading-edge chips are banned, 0 under full embargo. Up to 6 points are added for production on the territory.
- The 13 pilot countries are coded in detail; the others follow written rules.
The status of every coded component. Each component carries one of four statuses, shown on the country page:
| Status | Meaning | Score |
|---|---|---|
| Evidence | The level is established by a dated source | The level coded |
| Confirmed absence | A source documents that the measure does not exist | 0 |
| Presumed absence | Desk research found no trace of the measure | 0, provisional and low-confidence |
| Not coded | The component could not be determined | Missing, never 0 |
A presumed absence is a provisional, low-confidence score. The absence of accessible documentation is not proof that no policy exists: a country that publishes little may score lower than a country with similar practices that communicates more. These scores will be revised as second sources are checked.
Each country page shows what its Education pillar is based on (PISA and AI-in-schools coding, PISA only, or coding only), the level and status of every coded component, and the dated evidence behind it.
Deintelligence, the signal and relative divergence
Deintelligence is the gradual loss of a nation’s collective capacity to understand the world, adapt to it, and build its own future. The SI Index does not claim to capture the whole phenomenon. It tracks two measurable signs of it, which it keeps apart.
- Deintelligence signal: a country’s global score, or one of its pillars, falls from one edition to the next, measured with the same fixed goalposts and the same method. Because the goalposts never move, a lower score means lower measured capacity, not just a lower rank. A decline is a warning sign after one edition; it is a confirmed signal when it exceeds the uncertainty of the sources and persists over two consecutive editions.
- Relative divergence: the change in a country’s gap to the frontier between two editions. The frontier is the average global score of the ten best-ranked countries in each edition. A positive value means the gap is widening: the country is losing ground, even if its own capacity is growing. Falling behind is not, by itself, Deintelligence.
- First measurement: the 2027 edition (reference date 30 September 2027), compared with year zero. An interim update of the fast-moving indicators (compute, AI models, investment, adoption) is planned for March 2027; it shows how these indicators move, without a new global score.
Economic outcomes (not part of the score)
Each country page shows two economic outcomes next to the index. They are context, not inputs: if they were part of the score, the index could no longer be tested against them.
- Labour productivity: output per hour worked in 2025, in constant 2021 international dollars at purchasing power parity, and its average annual growth from 2019 to 2025 (ILO modelled estimates, November 2025; 2025 values are estimates).
- GDP per person: average annual growth of real GDP per person from 2019 to the latest year available (World Bank, constant dollars).
In Ireland and Luxembourg, GDP and productivity are inflated by multinational profits and cross-border workers; their pages carry a warning.
Versions
- v1.x: a correction to a grid or a source, or an indicator added within an existing pillar, with the eight pillars unchanged. Past editions are recalculated with the new version, so comparisons stay valid.
- v2.0: a change to the pillars or their weights.
Principles
- Filter, never correct: the index checks which observations fit each definition; it does not adjust a source’s figures.
- Goalposts are fixed before calculation and do not move to improve a ranking.
- Every change to a grid creates a new version, with the history recalculated.
Limits
This is an experimental index. Some indicators rely on a single source, coded indicators involve judgement, and coverage is uneven across regions. The Compute pillar only sees compute sold through listed cloud zones. Rankings that differ by a few points should not be over-interpreted.
Data
The full dataset can be downloaded: one row per economy and indicator, with the raw value, its unit, the source and its vintage, the normalised score, the pillar coverage, the status of coded components and any methodological caveat. Missing values are empty cells, never zero. The source data file is also available as it is.
- Download the dataset (CSV)
- Download the source data file (JSON)
- Normalisation bounds: see the table below.
Main sources
World Bank (World Development Indicators, Worldwide Governance Indicators 2025), Ember, OpenAlex, OECD.AI, OECD PISA 2025, Stanford AI Index 2026, Microsoft AI Diffusion Report (Q2 2026), Epoch AI, International Labour Organization (ILOSTAT), plus the public sources cited for each coded score.
Normalisation bounds (goalposts)
Each indicator is converted to a 0–100 score against fixed bounds, written down before any score was calculated. Values outside the bounds are clipped to 0 or 100.
score = 100 × (f(x) − f(min)) / (f(max) − f(min)), where f is the identity (linear scale), log₁₀ (logarithmic scale) or log(1 + x).
| Indicator | Measure | Min | Max | Scale |
|---|---|---|---|---|
| CMP-B | Cloud availability zones, weighted by tier | 0 | 100 | log(1 + x) |
| NRG-A | Electricity generation, kWh per person | 100 | 25,000 | Logarithmic |
| NRG-B | Growth in generating capacity over five years, % | 0 | 100 | Linear |
| NRG-B | Generating capacity added, GW per million people | 0.01 | 2 | Logarithmic |
| TAL-B | Top-10% cited AI publications per million people | 0.1 | 200 | Logarithmic |
| EDU-A | Top performers in PISA 2025, % | 0 | 60 | Linear |
| CAP-A | Private AI investment, % of GDP | 0.001 | 1 | Logarithmic |
| CAP-B | Capital markets and private credit, % of GDP | 10 | 300 | Logarithmic |
| ADO-A | Working-age population using AI, % | 0 | 100 | Linear |
| INS-A, INS-B | Worldwide Governance Indicators score | 0 | 100 | Linear |
| SOV-A | Notable AI models, last three years | 0 | 400 | log(1 + x) |
EDU-B and SOV-B are coded directly on a 0–100 scale and have no goalposts.