How we build the two indices

The Compute Capitals of AI IndexThe Citadels of Crypto IndexFull dataset (JSON)

Edition 1, Q4 2026, revised 13 October 2026. The Compute Capitals of AI Index and The Citadels of Crypto Index. Data frozen 13 October 2026. Updated every quarter.

What the two indices answer

Most rankings of AI or crypto countries ask where the industry is biggest. We ask a narrower question: if you build AI, or hold or build in crypto, where can you live and work well, and can you actually get the right to live there?

So each index has four pillars about the place and one pillar about you getting in. A country with a strong industry and no realistic way to get residence can't top the list. The list is global: we score every country, the United States included, the same way for someone of any nationality, whether they're moving from France to the United States or from the United States to France.

We don't run surveys and we don't score countries by feel. Every number comes from a statute, a regulator's register, a government page, an openly licensed dataset or our own database of mobility assets (residence and citizenship programmes). We only use sources we're allowed to republish, and every value is published with its source and the date we read it, so anyone can check our work.

No country is placed by hand. If a rank changes, it's because a rule changed, and the same rule applies to every country.

The score

• Each index has five pillars. Each pillar scores 0 to 10.
• In the AI index, each pillar weighs 20%. In the crypto index, tax weighs 30% and the other four pillars 17.5% each, because for people who hold crypto, what they keep after tax matters most.
• The index score is 10 × the weighted sum of the pillars, so it runs from 0 to 100.
• Every country, the United States included, is ranked on all five pillars.

Turning a rank into a score

Some sources publish ranks. We convert a rank into 0 to 10 like this:

rank score = 10 × (N + 1 − rank) ÷ N

N is the number of economies that source ranks. First of 139 in the Global Innovation Index scores 10.0; 139th scores 0.07.

The Compute Capitals of AI Index

Pillar 1: Compute and infrastructure

What it measures: access to computing power, data centres, connectivity and the energy behind them.

Pillar = average of the three.

Why a logarithm: the largest country has hundreds of times more tracked compute than the next tier. On a straight-line scale, every other country would score close to zero. On a log scale, going from 1,000 to 10,000 H100-equivalents counts the same as going from 100,000 to a million. A country with no tracked operational cluster scores 0 on this indicator. Epoch tracks the largest clusters; it doesn't see every server room.

Pillar 2: Rules, training-data law and model access

What it measures: how much legal friction there is in building, training and launching AI models, and whether you can use the leading models at all. Three tests read from the statute or official guidance, plus one read from the model makers' own lists of supported countries:

Pillar = 10 × (2 × R1 + R2 ÷ 2 + R3 + 2 × A) ÷ 6.

The access rule. If you can't use the leading models, a country can't be the best place to build AI. So a country where none of the three services is offered ranks below every country where at least one is, whatever its score. Its score is still published.

Pillar 3: Adoption

What it measures: how far AI is actually used, by government, businesses and people.

Pillar = average of the two.

Microsoft models its figure from anonymised usage data; it's an estimate, not a survey. Where Microsoft gives one shared figure to a group of neighbouring economies instead of an estimate for the country (Nigeria shares one with 12 West African economies), we treat it as missing.

Pillar 4: Talent and research capital

What it measures: the people, the research and the money that turns it into companies.

Pillar = average of the three parts. We publish only WIPO's own sub-pillar ranks, not the third-party figures inside them. Visas for AI talent are scored once, in pillar 5.

The Citadels of Crypto Index

Pillar 1: Rules, licensing and state commitment

What it measures: whether you can legally run a crypto business there, and how committed the state is to crypto.

Pillar = average of the two.

How we read licensing. An exchange or custodian in a country with no crypto regime at all scores 0, even if nothing bans it, because a business there has no clear legal footing. Software that never holds customer funds is different: where there's no regime and no ban, it needs no licence, so it scores 1. If we can't establish an activity from the regulator or the statute, we leave it out of the share and flag it, rather than guess.

State commitment to crypto. Each test scores what is true on the freeze date, not what a country once did or plans to do.

Central bank digital currencies don't count; they're not crypto. Seized coins awaiting sale aren't a reserve. El Salvador shows how the tests work. Its Bitcoin Law is still in force, but since the 2025 amendment, businesses choose whether to accept bitcoin, it's no longer defined as currency and taxes can't be paid in it. So it scores 1 for money status, not 2.

Pillar 2: Tax and transparency

What it measures: what you'd pay on crypto gains as a resident, and what a licensed crypto business actually pays. Many countries tax a short-term gain differently from a long-term one, so we look at a personal gain held for 6 months and the same gain held for 24 months. We also report whether the country takes part in the OECD's Crypto-Asset Reporting Framework (CARF), but that doesn't affect the score.

Pillar = half personal tax, half business tax. The personal half is the average of the two holding periods. People who hold crypto and people who build a crypto business both use this index, so what you keep and what your company keeps count the same.

We use the rate on the top slice of the gain, including surtaxes that always apply. For Canada we include Ontario's provincial tax. A wealth tax or a tax on a deemed return (the Netherlands' Box 3, Switzerland's cantonal wealth tax) isn't a tax on the gain, so it isn't in this score; we note it in the country's data.

Which business rate we use. We use the lowest rate a licensed virtual-asset business can generally get in that jurisdiction, where a member could realistically set one up:

• a statutory regime for digital-asset businesses (El Salvador exempts income from digital-asset transfers and operations);
• a special zone that licenses crypto firms (Kazakhstan's AIFC, Malaysia's Labuan, the UAE's free zones);
• a refund or exemption scheme (Malta's shareholder refund brings 35% down to an effective 5%).

We don't apply territorial exemptions for profit earned abroad, because they depend on how a business is structured, not on the country's rules for crypto. Where a business rate is distributed-profit only (Estonia, Georgia), we use the rate on distributions.

A draft law or a minister's announcement doesn't count until it's enacted. Germany's proposed 25% tax on coins bought from 2027 isn't in the score until it passes.

Pillar 3: Institutions and banking

What it measures: whether the institutions and the financial system are deep enough to bank you and your business, and whether people there actually use crypto.

Pillar = average of the first two, plus the Chainalysis bonus, capped at 10.

Pillar 4: Talent and innovation

What it measures: the people and the technology base a crypto business would hire from and build on.

Pillar = average of the two. Visas for founders and talent are scored once, in pillar 5.

Pillar 5 in both indices: Can you move there?

What it measures: whether someone of any nationality can get the legal right to live there, how hard it is, and where it leads. It comes from our database of mobility assets (residence and citizenship programmes), as it stood on the freeze date, which links every programme to its official government page.

Two kinds of programme don't count, because most people can't use them: programmes you can only use through ancestry (citizenship by descent), and programmes open to the citizens of one country only. In this edition that's Israel's B/5 investor visa and the Dutch-American Friendship Treaty, both for US citizens. Programmes open to the citizens of a list of countries, such as the US E-2 treaty investor visa or a working holiday visa, still count, and so do programmes that turn away only a few nationalities.

Pillar = sum of the six (maximum 10).

Missing data, coverage and rank stability

Missing indicators. No source covers every country. When a country lacks one or more indicators in a pillar, we don't treat the gap as zero and we don't simply average what's left, because sources score on different scales. Instead we fit a straight line, across every country that has all of that pillar's indicators, from the average of the indicators the country does have to the full pillar score, and use that line. Where a pillar weights its indicators, both averages use the same weights. We publish the line's two numbers for every pillar and every combination of missing indicators.

If a country has none of a pillar's indicators, it gets the median of all scored countries for that pillar, and we flag it.

Coverage. Next to every score we publish how much of it rests on real data. A pillar with every indicator observed counts 1. A bridged pillar counts the share of its indicators we observed. A pillar filled with the median counts 0. "Can you move there?" and the AI rules tests are always observed. The score's coverage weights the pillars like the score does, so 0.90 means 90% of the score rests on observed values.

Rank stability. A rank depends partly on the weights we chose. So for every country we also publish a rank range: the best and worst rank it gets across 17 versions of the index. Those are the published weights; each pillar's weight cut by 20% and raised by 20% (with the others rescaled); all pillars weighted equally; and each pillar left out in turn. A narrow range means the rank holds whatever reasonable weights you pick. A wide one means it leans on one pillar.

Unverified values. A tax rate or licensing status we couldn't confirm from a primary source in the current quarter is either published with a flag or left out and bridged as above. We never fill a gap with a guess.

Ties. Scores are shown to one decimal. If two countries tie, the one that's easier to move to (pillar 5) ranks higher; if they still tie, pillar 1 decides.

Rounding. We calculate with three decimals and round only the published score.

The quarterly refresh

Data freezes on the last day of each quarter. The new edition publishes in the second week of the next month (January, April, July, October). The first edition froze on 13 October 2026.

Every edition publishes the full dataset: each indicator's value, its source address, the date we read it, the pillar scores, the bridging lines, coverage, rank ranges and the final ranks. Anyone with the sources and the dataset can recompute every score.

When a source corrects itself or we find an error, we fix it in the next edition and list the change in that edition's notes. We don't silently revise a published edition.

Licence and attribution

The published dataset is licensed under Creative Commons Attribution-ShareAlike 4.0 (creativecommons.org/licenses/by-sa/4.0). It adapts values from the Oxford Insights Government AI Readiness Index 2025, which is published under the same licence, so you can reuse, adapt and share the dataset, commercially too, as long as you credit it and share what you build from it under the same licence.

Credit it as: CitizensOS, The Compute Capitals of AI Index and The Citadels of Crypto Index, Edition 1 (Q4 2026). The dataset includes data from:

• Oxford Insights, Government AI Readiness Index 2025 (CC BY-SA 4.0);
• WIPO, Global Innovation Index 2026 (CC BY 4.0). We converted WIPO's ranks into 0 to 10 scores. The Secretariat of WIPO assumes no liability or responsibility with regard to the transformation or translation of the original content;
• Epoch AI, AI data centers and Data on GPU clusters (CC BY 4.0);
• Microsoft AI Economy Institute, Global AI Diffusion data (MIT licence, copyright 2026 Microsoft), described in Misra and others (2025), *Measuring AI Diffusion: A Population-Normalized Metric for Tracking Global AI Usage*, arXiv:2511.02781;
• ESMA's interim MiCA register. This document has been drafted using material downloaded from ESMA's website; ESMA doesn't endorse it;
• Chainalysis, 2026 Global Crypto Adoption Index. We use only its published top-20 ranks, cited as facts.

Each source's licence is also listed with the source in the dataset.

What changed in the 13 October 2026 revision

• We dropped two sources we can't republish or reproduce: the Henley Crypto Adoption Index and Tortoise Media's Global AI Index (last published September 2024). WIPO Global Innovation Index ranks and Microsoft's openly licensed AI diffusion data take their place.
• AI rules now include access to the leading models, and a country where none of them is offered can't rank above one where they are.
• Crypto licensing is scored by what's actually permitted, across four business models, instead of a single regime level.
• Personal crypto gains are scored at two holding periods, 6 and 24 months, and personal and business tax now weigh the same.
• "Can you move there?" uses our mobility asset database as it stood on the freeze date, including Kazakhstan's programmes added that day. Kazakhstan and Mauritius were scored as having a self-sponsored residence permit, but on your own income they offer only a visa (Mauritius's retiree permit is limited by age), so they now score 2, not 3, under the rule above.
• The list is global. The United States was scored as a home country and left unranked; it's now ranked like every other country. "Can you move there?" is scored for someone of any nationality: a programme open to the citizens of one country only no longer counts, and the citizenship test asks whether you can keep the citizenship you already have, not whether you can keep a US passport. The United States ranks 1st in AI and 10th in crypto. The Netherlands and Israel score lower on moving there, because their programmes for US citizens only no longer count.
• Every score now shows its coverage and its rank range, and the dataset carries an open licence.

What these indices are not

• They're not legal, tax or immigration advice. Governments decide every application and change the rules often.
• A high score doesn't mean you'll pay less tax. If you're a US citizen, the United States taxes your worldwide income wherever you live, so moving abroad doesn't end your US tax return.
• They don't score political risk, personal safety or cost of living. Where those matter, we say so in the country write-up.
• The AI index counts compute that is running, not compute that is announced. A data centre counts once Epoch AI's tracking shows it operating. That's why a site under construction, such as the first phase of Stargate UAE, doesn't count yet.