Framework — AGI Growth Atlas

The framework

What converts AGI capability into growth?

The atlas starts from one proposition: AI could make intelligence abundant and inexpensive, but intelligence alone does not produce economic growth. Explosive growth would also require electricity, chips, capital, skilled builders and governments able to approve construction. Economists describe these inputs as complements because they must expand together. In each country, the complement that scales most slowly sets the growth ceiling. The model therefore places less weight on whether a country develops frontier AI and more on whether its power, capital and construction capacity can keep pace. The scenario supplies the potential growth rate; the complements determine how much of it becomes domestic growth.

2025

Early automation effects

growth begins to rise above trend; the initial increase matches skeptical estimates

2030

AGI under Fast

takeoff before countries have much time to adjust; modeled growth reaches about 30% in the best-positioned country

2040

AGI under Gradual

takeoff after more than a decade of adjustment; modeled growth reaches about 15% in the best-positioned country

2060

Model horizon

the end point used to show the difference between pre- and post-AGI paths

The scorecard

The nine complements

Every country is scored 1 to 5 on nine factors, grouped into four kinds: can it get frontier AI and the power to run it, can it deploy what it gets, how much room does it have to grow, and can it keep the lights on. Click a factor for what it measures, how it is scored and how much it counts under each scenario.

Access · slow to scale

Ability to obtain frontier chips and models and supply the power needed to run them.

Deployment · policy-sensitive

Capacity to adopt AI across firms and expand the associated physical infrastructure.

Potential · structural

Scope for catch-up growth and for workers to move into activities that remain less automated.

Resilience · slow to scale

Capacity to protect output and infrastructure from recurring shocks.

Compute & Energy Access

Policy

The capacity to run frontier AI at scale, measured through domestic or reliably contracted data center capacity and the firm, affordable electricity needed to operate it. Fabrication plants, data centers and power generation take years to build, making this one of the slowest complements to expand and therefore more heavily weighted under Fast.

Capital accumulation — growing the stock of machine labor a country can actually run · Continuing: a flow constraint for as long as the frontier moves

Share of the index under each lens

Modest 8%·Gradual 15%·Fast 24%▲

What moves it

Fast-track grid interconnection and new generation (incl. nuclear, gas, renewables + storage)
Secure chip access via alliances or domestic/partner fab capacity
Offer power and land packages to anchor hyperscale data center investment

Indicators, weights and sources for this factor:

From nine scores to a growth rate

Economists describe growth through three sources: productivity, the output produced by a given set of inputs; capital, including machinery, buildings and infrastructure; and labor, including people, hours and skills.

output growth ≈ productivity + capital + labor

The Gradual and Fast scenarios assume frontier AI changes this relationship: when AI can perform work, effective labor becomes reproducible through capital, capital that behaves like labor (the mechanism in Why hypergrowth?). Productivity then rises with model capability, and accumulation proceeds at the pace of construction rather than population growth.

Weighted together, the nine scores form the AGI Growth Index, 0 to 100. Under each scenario the index places a country between a lower anchor near its own pre-AI trend and the scenario’s ceiling: 15% a year under Gradual, 30% under Fast, a small one-time gain under Modest. These are conditional simulations, not forecasts, and every number here can be changed in the .

How the weights shift with the speed of the takeoff

Each factor’s share of the index. A fast takeoff rewards what a country already has: compute, the capacity to build and access to the frontier. A slow one rewards adoption, headroom and a growing workforce, because there is time for them to pay off.

Modest Gradual Fastthe arrow runs from Modest to Fast · hover for all three · click a factor to open it above

Result under the current model

Faster takeoff produces wider differences in national growth

In the model, faster takeoff increases the gap between countries. The gap between the 90th- and 10th-percentile countries widens from about 5 points under Modest (its peak rates, in 2035) to 8 under Gradual (rates from 2040) and 15 under Fast (rates from 2030). Fast gives more weight to compute, capital and the capacity to build, which are concentrated in richer and larger states, and less weight to catch-up potential and demographics, which account for more of the score in many poorer states. Under Fast, some countries grow more slowly than they would without AGI.

Rates here are the post-AGI rates each scenario reaches after its takeoff: from 2040 under Gradual, from 2030 under Fast, and Modest’s 2035 peak. Each dot represents a country — hover to identify it. The bar runs from minimum to maximum growth and the tick marks the mean. All three scenarios use the same axis so their dispersion can be compared directly.

Countries with the lowest modeled gains. Under Gradual, no country falls below its pre-AI trend because the slower transition allows time to adapt. Under Fast, 16 countries do, among them Ethiopia and Ghana: their current growth depends heavily on labor-intensive exports, while their complement scores are low, so the model gives them lower growth than in the no-AI baseline. The distribution widens at both the top and, under Fast, the bottom.

Result under the current model

A new Great Divergence?

To economists, convergence is the idea that poorer countries should grow faster than economies at the productivity frontier. After the first industrial revolution the world saw the opposite: a Great Divergence, as industrializing economies grew far faster than the rest, and global income inequality rose for more than a century. Only after 1989 did it begin the forty-year fall that continues today, driven above all by China and then India. Convergence has never been automatic. It depends on institutions, skills, infrastructure, capital and access to technology, which let an economy absorb frontier know-how and turn it into output. Those are the complements this atlas scores, and a fast takeoff rewards the countries that already have them.

Will AGI end the Great Convergence?

Inequality between countries: the Gini of GDP per person in purchasing-power terms, with each country weighted by its population, which is Milanović’s between-country measure. To 2018, his estimates; from 2018 to 2025, the same measure computed from World Bank data for 184 economies; from 2025, this atlas’s paths for 157 countries under each scenario and without AI (dashed). Hover for values; the shaded years after 2025 use a wider time scale. Convergence has already slowed: the measure fell by about half a point a year from 1990 to 2018 and has barely moved since. Without AI it stays roughly flat to the mid-2040s: China is now above the world average in purchasing-power terms, so its growth widens the gap, while India and several African economies narrow it. After that the no-AI line edges up, because today’s trends, held constant, carry China’s income per person past that of the United States in the 2050s; the gap between that line and the scenarios is the finding.

Richer almost everywhere, far richer in a few places. Income per person rises in nearly every country under every scenario; the paths are growth on top of today’s levels. What changes with the speed of the takeoff is the gap. The countries that score highest on the index, the ones best able to convert compute, access, skills, capital and build capacity into output, pull far ahead, and under Fast a handful of economies whose growth depended on cheap labor end up below their old path. Without AI and under Modest, inequality between countries stays roughly where it is until the mid-2040s; under Gradual it turns up in the early 2030s and passes its 1988 peak in the mid-2050s; under Fast it passes that peak before 2040.

What the model does not say. It models growth between countries only. It says nothing about household incomes or inequality within countries, and work that models the inside of an economy points the same way: faster AI shifts returns from labor to capital. It also holds each post-AGI rate constant, so the 2060 values overstate any real path; the turn in the curve is the finding, not its endpoint. Every assumption behind it is on this page and adjustable in the .

Every country across the three scenarios

Bold lines: the average post-AGI growth rate of each income group (or archetype) under Modest, Gradual and Fast, with every country as a thin line behind. Under Modest poorer countries still grow faster than rich ones, as they do today (4.3% a year for low-income countries against 2.6% for high-income ones). Under Gradual and Fast the order flips and the gap widens (14.1% against 4.5% under Fast). Hover a line for the country.

Low (15)Lower middle (39)Upper middle (45)High (58)

Source for 1820–2018: Milanović (2024), Table A2. The eleven archetypes in the fan are described on the .