Methodology & sources
How the atlas works — and where it’s uncertain
The atlas treats AGI as a scenario for analysis and asks a conditional question: if a transformation of this kind occurs, how do national growth outcomes differ? All results are illustrative rather than forecasts. Explosive growth remains a minority view in economics, and the atlas assigns no probability to any scenario. This page documents the model, data, known artifacts and further reading.
The model, in four steps
1 · Factor scores
Each country receives a score from 1 to 5 on nine factors, where 3.0 is approximately the global median and 5.0 is the highest range. Scores are data-derived (v5, September 2026), combining public datasets (World Bank WDI & WGI, Epoch AI compute shares, WB Climate Portal heat, measured AI adoption, cybersecurity indices, CommonCrawl language data) with documented rubric inputs for geopolitics and language; imputed or estimated values are flagged. The nine factors and four families are explained in the Framework.
2 · The AGI Growth Index (0–100)
A weighted average of the nine scores, rescaled from 0 to 100. Because the weights change by scenario, a country can occupy a different position under each one. The scenario specifies the available growth rates; the index places the country within that range (see “What growth means in this atlas,” below).
3 · Post-AGI growth
The index maps to a growth rate within a frozen band. Gradual and Fast replace the previous trend with an absolute post-AGI rate. The lower anchor of the mapping is each country’s own no-AI trend, capped at the scenario “floor” (5% Gradual, 3% Fast), so the least-ready country is not handed the floor rate; the upper anchor is the ceiling. Under Gradual, the no-AI trend is also a hard minimum because the slower transition allows time to adapt. Fast has no such minimum, so some countries grow more slowly than they would without AGI; this represents the export ladder effect. Modest instead adds a small temporary increase to the existing trend. Across the site, growth is shown as this post-AGI rate (for Modest, the 2035 peak) unless a calendar year is chosen in the lens bar; the year-by-year path is the ramp in step 4.
anchor = max( 0, min( floor, BAU ) )
gradual: g = max( anchor + (ceiling − anchor) × norm^elasticity, BAU )
fast : g = anchor + (ceiling − anchor) × norm^elasticity — no BAU minimum
modest : g = BAU + upliftCeiling × norm^elasticity
4 · GDP to 2060
Under Gradual and Fast, growth follows a convex path from the ramp start S (2025 by default, adjustable below) to the AGI year. The contribution of automation is back-loaded because the final tasks automated produce the largest gains (Erdil & Besiroglu, fig. 2). Growth then remains at the post-AGI rate through 2060; the early part of the path matches skeptical estimates by construction. Under Modest, the uplift applies only during diffusion (~2025–2035), after which growth returns to trend: a one-time level gain of up to ~2% of GDP, consistent with the literal reading of Acemoglu’s estimate. Per-capita paths divide GDP by projected population (World Bank projections to 2050, damped extrapolation to 2060). The derivation of the 15% and 30% ceilings appears in Why hypergrowth?
doubling time = ln 2 ÷ ln(1 + g/100)
What growth means in this atlas
The introduces this logic; here it is with the formulas, the frozen bands and the per-factor mapping.
Economists account for growth with three sources: productivity (how much output the same inputs yield), capital (the machines, buildings and infrastructure workers use) and labor (people, hours and skills) — roughly, output growth ≈ productivity growth + capital accumulation + labor growth. The Gradual and Fast lenses assume frontier AI rewires this arithmetic: when AI can do the work, effective labor stops being the fixed input — it becomes something you can build, capital that behaves like labor — so productivity jumps with each model generation and accumulation is freed to run at the speed of construction rather than the speed of population. That is the global engine. The nine factors do not create it. They are the complements that decide how much of it a country converts into domestic growth: the standing to buy the frontier, generation after generation (geopolitics); the capacity to run it — chips, data centers, and above all power (compute); the speed at which the physical economy around it gets built (build); the ability to keep redeploying a changing technology across firms, the state and languages (diffusion, language); the headroom to grow by adopting what already works elsewhere (catch-up); the workforce to grow and redeploy (demographics); and protection against the recurring losses that eat capital and uptime (cyber, climate).
The mapping is deliberately reduced-form. The scenario specifies the growth rates, while the index determines a country’s position within them. The lower anchor is the country’s own trend, capped at the scenario floor (formulas in step 3), and the upper anchor is the scenario ceiling. “Best-positioned” means that a country has the strongest bundle of the nine complements under the selected weights. Scores remain fixed during the simulation, preserving each country’s relative starting bundle; the policy editor compares alternative bundles — comparative statics, a different entering position rather than a modeled reform path. The resulting growth rates, doubling times and GDP paths are conditional consequences of these assumptions rather than equilibrium forecasts.
How to read a country’s result. Each rate operates as a ceiling given the country’s current complements, the pace of AGI in the selected scenario and sustained government support for deployment. Improving the complements raises that ceiling; the policy editor shows the modeled amount. Reaching it depends on political economy: whether the groups that lose from deployment — professions, rightsholders, incumbents — hold the power to block it, and whether the country captures enough of the surplus to compensate them. The model does not measure either. The same interpretation applies to countries below trend under Fast: with their current complements, the lower value is their modeled ceiling after a rapid takeoff replaces the conditions supporting their current trend.
How each factor is built — indicators, sources, weights
GDP, GDP per capita and the 2016–2025 mean real-growth rate come directly from the World Bank, World Development Indicators (NY.GDP.MKTP.CD, NY.GDP.PCAP.CD, NY.GDP.MKTP.KD.ZG; levels are 2025 actuals, growth the 2016–2025 mean). Each of the nine factor scores is built from the components below (v5 · September 2026).
Each data component is percentile-ranked across the included countries and combined using the weights below. The combined value is ranked again and mapped to a 1–5 scale, making 3.0 the median by construction. Geopolitics is the exception because it uses an absolute rubric with meaningful plateaus. Compute also preserves observed concentration because the United States holds about 80% of known AI compute.
Compute & Energy Access
computeThe factor combines training compute and inference compute. Training compute refers to clusters used to develop more capable models and is highly concentrated; the United States holds about 80% of tracked capacity. Inference compute refers to data center capacity used to run AI throughout an economy and is more widely distributed, including hubs such as Ireland, Malaysia and the UAE. The scoring preserves the observed concentration rather than stretching all countries across the full scale.
Caveat: Epoch tracks publicly known clusters — English-source-biased, so China is likely under-counted. The data center capacity tier and pipeline are hand-compiled from industry reports and public announcements (inspectable). Electricity lives in build, not here, to avoid double-counting.
Geopolitical Positioning
geopoliticsPosition in the AI supply chain and the blocs that control it: who can get chips, frontier models (via trusted API access or open weights), capital and technology.
Caveat: An absolute rubric — the plateaus carry meaning (4.0 = trusted ally), so it is NOT percentile-spread. Every value is inspectable.
AI Diffusion Capacity
diffusionHow fast AI is absorbed into firms and daily life — connectivity, absorptive capacity, measured AI adoption, and the absolute size of the online workforce.
Caveat: Broadband was dropped (collinear with internet use). Microsoft covers 147 economies (some small states missing); EGDI excludes Taiwan and Hong Kong — the blend renormalizes where a source is missing. Microsoft measures usage of tracked genAI tools and is platform-tilted; EGDI measures government digitalization, not firm adoption — hence the modest weights.
Language Inclusion
languageHow well the country’s languages are served by frontier AI models, plus English exposure.
Caveat: Language-to-country mapping and EF bands are hand-entered from the published sources; EF covers ~113 countries, so some small states use the official-language fallback.
Physical Build Speed
buildThe factor measures the rate and scale of physical investment across the economy, including but not limited to data centers. Investment intensity is combined with a scale gate based on absolute investment so that one large project in a small economy does not produce the same score as sustained national construction.
Caveat: A construction-permits or floor-space series would be ideal, but none has global coverage since Doing Business was discontinued. The scale gate replaced the v3 additive formula, which let small high-investment economies (Maldives, Gambia, Niger) outrank the UAE.
Catch-up Headroom
catchupRoom to grow by adopting what already exists — the income gap to the frontier, discounted by state capacity to convert it.
Caveat: Taiwan uses IMF/DGBAS income data (absent from WDI).
Demographics & Migration
demographicsThe workforce: its trajectory, its size — and its quality.
Caveat: Migration pull is only captured insofar as it is in the UN projections. HCI moved here from catch-up: workforce quality belongs with the workforce.
Cyber Resilience
cyberThe factor measures digital-infrastructure security through implemented capacity, formal commitments and secure infrastructure.
Caveat: NCSI is Europe-tilted (misses parts of Asia); ITU excludes Taiwan and Hong Kong — the blend renormalizes over the sources available for each country.
Climate & Environmental Resilience
climatePhysical suitability for AI infrastructure and an AI-era economy: heat load first, water for cooling second.
Caveat: Heat is weighted above water: cooling design can partially substitute for water, not for heat. Temperatures are area-weighted over territory, not population — this flatters countries with large cold hinterlands (Canada, Chile) relative to where their grids actually are.
Three countries are special-cased and flagged in the artifacts list below (Taiwan, Ireland, Guyana). The full per-country provenance — including every rubric, imputed and estimated value — is in the inspection workbook (data/factor-grounding-v5.xlsx) and data/grounding-v2.json.
The three scenarios
Set the takeoff timing
Growth ramps from the year AI starts adding to measured growth (2025 now) to each scenario’s post-AGI rate in its AGI year (Gradual, Fast), or to Modest’s peak. The ramp is back-loaded, so an earlier start adds little in the first years. Move the years to test a slower or faster arrival: every map, table, chart and country page recomputes.
Change the other assumptions
The allows users to change each scenario’s ceiling, the cap on its lower anchor (the “floor”) and the factor weights as well as the timing above. Every map, ranking and chart recalculates when those settings change, and “Share this view” in the lens bar copies a link that carries them.
Country archetypes, in two tiers
The index says how well a country can capture AI-driven growth. The archetypes add what the index does not: where a country sits in the AI economy. Most writing on AI and the world economy looks at the two AI superpowers, the holders of chokepoints in the chip supply chain, and the twenty or so middle powers with the capital, markets or institutions to bargain for access (Leicht & Ball 2026; Chatham House 2026; Tony Blair Institute 2026). The first tier covers them. The second tier covers the hundred-plus countries that writing mostly leaves out, sorted by the channel through which AI reaches them: offshored work, visitors and diaspora earnings, smallholder farms, or adoption at home.
| How the literature groups countries | Archetypes in the atlas |
|---|---|
| AI superpowers: the United States and China | Frontier superpowers |
| Middle powers: about 15 to 25 advanced or strategically placed economies | Supply chain chokepoints, Fast-build capital hubs, High-trust diffusers; at the edges, the largest Regional leaders |
| Regional or “pivotal” states (Chase, Hill & Kennedy 1996) | Regional leaders |
| Small states and the least developed, rarely analyzed | The second tier, and most Resource suppliers |
Frontier superpowers. A compute score of at least 3.8: the United States and China.
Supply chain chokepoints. Taiwan, South Korea, Japan and the Netherlands, named for fabrication, memory, equipment and lithography.
Fast-build capital hubs. The Gulf builders and Malaysia’s Johor corridor, plus countries with a build score of at least 3.8, income above $15,000 per person and diffusion of at least 3.0, or build of at least 4.6 and diffusion of at least 3.5 with trend growth below 4.3%.
High-trust diffusers. A diffusion score of at least 3.9 and income above $20,000 per person, or 3.3 and above $35,000.
Resource suppliers. Fuels and ores make up at least half of goods exports, averaged over 2022–2025, and income is below $40,000 per person. Guinea, Mauritania and Angola qualify by documented exception, as does Trinidad & Tobago, whose gas-based ammonia and methanol count as manufactures. Gold and diamonds do not count, because the build-out does not use them.
Regional leaders. The supply chain’s regional gateways: each is the largest economy in its UN subregion, with at least 50 million people, income below $20,000 per person and an established data center market.
Offshoring destinations. Manufactures and modern services exports of at least 15% of GDP, larger than tourism and remittances; or 10% where manufactures exceed 80% of goods exports.
Tourism and diaspora economies. Travel exports plus remittances of at least 12% of GDP, larger than offshored exports.
Smallholder economies. At least 60% of workers in agriculture, or 40% in countries under 100 million people.
Sheltered economies. The rest: no channel dominates, so the outcome turns on adoption at home.
High-risk / volatile. Government effectiveness at −1.05 or below, or −0.9 with trend growth under 3%; cut off from the AI supply chain while stagnating; or a military government after a coup (currently Mali, Burkina Faso, Niger, Madagascar and Myanmar).
Rules fire in order and the first match decides: frontier, chokepoints, high-risk, the named capital hubs, regional leaders, resource suppliers, diffusers, the other capital hubs, then the four channels. Export and employment shares are World Bank WDI averages for 2022–2025; government effectiveness is the Worldwide Governance Indicators estimate averaged over the last three years, because single years move within the estimate’s margin of error; military governments follow the African Union’s suspensions. The workbook lists the rule and the values behind every country’s assignment. Unlike the index, the archetypes use economic structure; the next section explains why the index does not.
Economic structure: shown, not scored
Countries meet AI with very different production and employment structures. Formal, digitally mediated services are often easier to augment or automate than agriculture, construction and informal work. Lower exposure can provide temporary insulation while also limiting near-term productivity gains. The atlas therefore shows sectoral composition, informality and current automation exposure on country pages, but does not currently use them in the AGI Growth Index or the growth scenarios. The direction and size of the effect are scenario-dependent, and today’s structure will itself change.
Sector shares come from the World Bank World Development Indicators, informality from ILOSTAT, and present-day task exposure from the Global Automation Atlas (Garg, Crosta & Baier, 2026 — task-level exposure classified under country-specific workplace conditions). The latest available year is shown for each country and measure; the panel’s labels are coarse display bands over the raw values, documented in the source data. The panel’s framing follows the World Development Report 2026 (The Promise of Artificial Intelligence) and its background paper Disruption without Dividend? (Gmyrek, Viollaz & Winkler), which document how task composition, informality and connectivity split AI’s near-term impact across countries. The WDR’s own decomposition supports keeping structure out of the index: adoption differences account for about two-thirds of the projected rich–poor gap in near-term AI productivity gains, task structure for about one-third — and the structural share shrinks in faster scenarios, as AI’s capabilities reach more of the task mix.
Securing access: alignment, relevance and the procurement trap
Geopolitical positioning and compute & energy access measure how a country obtains frontier capability. Their policy guidance follows the current effect of export controls and alliance frameworks on access to chips and models: alignment with supplier blocs, and the ability to offer energy, sites, minerals, talent or market access, can affect the score. The score describes a country’s position under current regimes rather than recommending which bloc it should join.
The WDR 2026 gives greater emphasis to a middle path based on open-weight models and shared regional compute, and argues that full AI sovereignty may become a development trap. The atlas gives more weight to the possibility that open-weight availability may narrow: the WDR reports that the open share of notable new models fell from 56% to about 40% between 2020 and 2025, and continued access also depends on release decisions by a small number of labs, including prominent Chinese labs.
Both approaches support careful procurement. Technology contracts can create long-term dependence regardless of the supplier’s country. Agreements should preserve data control, exit rights, audit rights and the ability to change providers. A domestic data center does not by itself create sovereignty, and the buyer’s long-term bargaining position may matter more than the initial price once systems are embedded in government workflows.
Known model artifacts and limitations
- The between-country Gini is continued to 2060. The framework page extends Milanović’s between-country series (1820–2018) with the same measure computed from World Bank data: the population-weighted Gini of GDP per person in purchasing-power terms (constant 2021 international $) for a balanced panel of 184 economies, 1990–2025 (scripts/build-between-gini.mjs; Taiwan from IMF WEO). Computed this way it falls 15 points from 1990 to 2018, against 16 in Milanović’s survey-based series, and sits about a point below it, so the chart adds 1.1 points to align the two in 2018. From 2025 each country’s PPP level is grown along the atlas’s paths to 2060. The trends are 2016–2025 averages held constant, so after the mid-2040s they carry China’s income per person past that of the United States even without AI, which turns the no-AI line up; the gap between the scenarios and that line is the finding. Post-AGI rates are also held constant, so the later levels overstate any real path.
- Post-AGI growth is held constant. Each country compounds its post-AGI rate unchanged to 2060, so the 2060 GDP levels — especially under Fast takeoff — are illustrative, not forecasts. The charts cap the axis and let runaway paths climb off the top. The timing and the divergence are the point, not the dollar figures.
- No fading catch-up, no resource ceiling. The rates are held constant, so the model does not yet endogenize three things: the erosion of catch-up as countries approach a moving frontier, the continued innovation a frontier economy needs to keep converting new technological opportunities into growth, and the tightening of global energy, land, material and construction limits after repeated output doublings. Nor does it require the independently projected country paths to clear one global resource market. The Fast ceiling compounded 2030–2060 implies roughly fourteen income doublings for the best-positioned country — read that as a transitional stress test of relative exposure and timing, not a physical forecast.
- Investment finance is proxied, not scored. Savings, financial depth, sovereign borrowing capacity and FDI access have no factor of their own — they enter only indirectly through compute, build and geopolitics. A country could hold every other complement and still be capital-starved; the model cannot see that directly. During a takeoff the problem can compound: capital historically flows toward the highest returns, which a frontier acceleration would place in the frontier itself (the Lucas paradox; Alonso, Berg, Kothari, Papageorgiou & Rehman 2020 model the AI version) — so financing constraints may tighten for followers just as the machines become worth buying.
- Wage-based adoption incentives are not modeled. Automation pays when machines undercut wages, so identical robots that are profitable in Seattle can be uneconomical in Lilongwe — low wages can rationally delay adoption even where access exists. The atlas has no wage channel: it captures the export ladder risk (fast-growers falling below trend under Fast) but not this adoption-incentive margin, which would slow diffusion most in the very countries with the lowest incomes. How cheap AI-embodied machines must get to compete with the world’s lowest wages is an open empirical question. The WDR 2026 documents the pattern today: automated checkout and garment machinery sit unadopted in Bangladesh, Cambodia and Viet Nam because the workers cost less than the machines.
- Catch-up headroom is conditional, not automatic. History’s default is divergence: each technology wave so far has widened the gap in how intensively rich and lower-income countries use it, and the WDR 2026 projects the AI adoption gap widening through 2035 on current trends. In this atlas, headroom raises a country’s ceiling only through the index, alongside the other eight complements — a country with weak complements cannot ride headroom alone, which is why thin-complement fast-growers can end up below trend under Fast rather than above it.
- Compute reflects known clusters only. The compute share is built from publicly tracked AI clusters (Epoch AI), which are easiest to observe in English-language sources — China is likely under-counted, and secretive programs are missed.
- Climate is area-weighted. Temperatures average over territory, not population — countries with large cold hinterlands (Canada, Chile) score better than their populated, grid-connected regions would alone.
- Adoption and cyber indices have coverage gaps. Microsoft’s adoption data covers 147 economies (some small states missing); the UN and ITU indices exclude Taiwan and Hong Kong; NCSI is Europe-tilted. Blends renormalize over the sources available for each country.
- Fast takeoff can leave countries below trend. Under the Fast lens, sixteen countries grow slower than their no-AI baseline: Sahel and fragile states, but also fast growers such as Ethiopia, Ghana and Côte d’Ivoire, whose pre-AI growth was fast but whose complement bundle is thin. The mechanism (the index replacing trend) is a proxy for export ladder displacement, not a direct measure of export exposure; treat the specific list as indicative. A direct measure now exists — Kruse & Yu’s WDR 2026 background paper, “The AI Exposure of Export-Linked Jobs” — and a future version would tie the mechanism to it.
- Population projections end at 2050. Per-capita paths use World Bank population projections, which run to 2050; 2050–2060 is extrapolated by damping each country’s projected growth rate. Taiwan uses its national (NDC) projection.
- Ireland & Taiwan are special-cased. Ireland’s headline GDP/growth are distorted by multinational accounting (WDI shows ~8% growth), so growth uses the CSO modified-GNI* proxy (~3.5%); Taiwan is absent from WDI, so its figures come from the IMF; Guyana’s 10-year average is distorted by its oil startup (60%+ years), so its trend uses the IMF medium-term view (~4.5%).
Which countries are included — and which are not
The atlas covers every economy for which reliable, internationally comparable statistics exist — currently 157. A small set of states is excluded because the data does not: countries under broad sanctions, in active conflict, or not reporting through the UN and World Bank statistical system.
Planned extensions to the model
Growth paths hold each post-AGI rate constant to 2060. A fuller version would let each rate evolve through deployment, catch-up and physical saturation — and estimate after how many global output doublings diminishing catch-up and hard-to-reproduce inputs (energy, key metals, construction lead times) materially slow frontier growth. This is a genuine research extension, drawing on the endogenous growth literature (Aghion, Jones & Jones on whether essential hard-to-improve tasks remain) and recent physical-limits analysis (Binder’s AI industrial-explosion series).
The cleanest bridge from this atlas to a dynamic model: map each country’s complement bundle into an adoption lag in a frontier–follower formulation — the frontier’s automation path arrives D years later, with D set by the complements. The task model literature (Zeira; Acemoglu & Restrepo; Aghion, Jones & Jones) supplies the machinery; the atlas supplies the cross-section. That would replace the constant post-AGI rate with a transition whose speed differs by country, and make “reduce the lag” a quantitative policy target rather than a slogan.
The bottleneck moves over time — compute & capital early, absorptive capacity in the diffusion era, energy & build-out late — so the leaderboard reshuffles by phase. A future version will model this explicitly rather than as a single static index.
A high national growth rate can hide brutal internal inequality (capital vs labor). A second axis — distribution within a country — is the largest conceptual gap relative to “changing fortunes.” Anthropic’s 2026 scenario explorer models this axis for the United States (wages, unemployment, the labor share); a country-level version would need comparable task and labor data.
The jurisdiction argument implies regulation may deserve to be a multiplier on the whole index rather than one of nine weighted factors.
The pipeline can extend past 157 wherever World Bank coverage allows; other upgrades: population-weighted climatology, the IMF AI Preparedness Index once machine-readable, and measured-adoption sources beyond Microsoft.
Further reading
The sources below cover skeptical, gradual and fast views of AI-driven growth.