The argument
Why hypergrowth?
The atlas examines the possibility that artificial general intelligence (AGI) — AI able to perform any job done at a computer — could raise annual growth in the best-positioned countries from about 2% to 20% or more. This page presents the historical comparisons, economic mechanism and capability trends behind that scenario, followed by the main reasons it may not occur.
1 · Growth regimes have changed before
Two earlier transitions changed the rate of economic growth: the shift from foraging to farming, and the Industrial Revolutions. Between 1 CE and 1700, average living standards rose by roughly 0.03–0.05% a year. Between 1820 and 2020, as economies adopted steam, electricity, combustion engines and computing, the rate rose to 1–1.5% a year — a 20–40-fold increase in the annual rate of improvement. In the long-run record, almost all of the rise in GDP per capita occurs at the very end.
2 · National economies have sustained much faster growth
After the Second World War, catch-up growth helped Korea, Taiwan, Singapore and China sustain annual increases of 6–9% in income per person for 30–40 years. Over those periods, income per person rose by 10 to 26 times. These episodes show that national economies can sustain growth far above the current rich-country trend for several decades. The explosive growth literature asks whether AI could produce another proportional increase, from about 2% a year in advanced economies to 20% or more.
3 · The mechanism: capital that behaves like labor
The relevant mechanism is not the current generation of chatbots. It is the possibility that sufficiently capable AI turns capital into an input that behaves like labor, first digitally and increasingly through embodied systems. If machines can perform the work people do, the effective workforce can grow with industrial production rather than human reproduction. This removes the fixed labor supply constraint used in earlier growth models. Forethought describes the possible result as “a century of progress in a decade.” The principal counterargument is the bottleneck view: growth remains limited by tasks and inputs that machines cannot provide. The atlas applies that logic to differences between countries.
4 · How fast is AI actually improving?
METR measures the length of tasks that frontier models can complete autonomously with a 50% success rate. On that measure, capability has doubled every 4.3 months since 2023, compared with seven months over 2019–2025, without an observed slowdown. Exponential trends eventually become sigmoids, but the turning point is difficult to predict. The atlas uses a prior that a stable trend continues for about as long as it has already run, which would bring tasks that take skilled people weeks or months within reach around 2029, or 2033 at the slower long-run rate. Physical AI is scaling at the same time: Waymo passed 220 million rider-only miles by early 2026, Amazon a million deployed robots in 2025, and industrial and humanoid robotics keep growing (IFR; IEEE Spectrum).
5 · Estimates range from modest gains to explosive growth
Most growth economists reject the explosive growth hypothesis; most people building frontier AI expect it. Tom Cunningham’s compendium of forecasts for 2025–2035 puts economists at 0.1–1.5% a year of extra growth and AI insiders at 3–30%. The Modest lens represents the economists’ range, Gradual and Fast the insiders’. The policy case does not require certainty: a 10–20% probability of a change in the growth regime should still affect decisions by governments, investors and funds today.
Modest · The skeptics. Acemoglu and Nordhaus put AI’s contribution at a fraction of a percentage point a year; bank forecasts at one or two points.
Gradual · Moll and Imas (2026). AI capabilities may explode, but frontier growth is unlikely to rise above 4–5% a year within 10–15 years.
Fast · The explosive case. Machines of Loving Grace, AI 2027 and Situational Awareness, from people building the systems and observers of the capability trend.
A similar exercise with three named scenarios, for the United States only: Anthropic Economics’ scenario explorer (modest, substantial and extreme, to 2030).
6 · From the argument to the atlas’s numbers
The atlas represents the argument with two growth ceilings and a transition path. Each ceiling matches published forecasts in Cunningham’s compendium; the path follows Erdil & Besiroglu’s model of automation. Both ceilings, the floors and the AGI years are yours to change in the ; every map, table and country page recomputes.
Gradual ceiling: 15% a year from 2040
Gradual sits at the center of the named forecasts from people building frontier AI. Dario Amodei puts developed-world growth at 10–15% a year; Anthropic’s economists put the United States near 15% in their most transformative scenario; Korinek & Suh’s baseline AGI model reaches 18%. The atlas sets ~15% for the best-positioned country, about seven times today’s frontier trend. That is transformative rather than explosive: short of the order-of-magnitude jump Erdil & Besiroglu use to define explosive growth. A takeoff that matures around 2040 still leaves regulation and slowly accumulated inputs such as land, energy and capital more than a decade to shape it. At this rate an economy doubles about every five years.
Fast ceiling: 30% a year from 2030
Fast is the explosive case. Thirty percent a year is the threshold the literature uses to define explosive growth, and the path Epoch AI’s GATE model reaches for world output by 2035 under its default settings; Leopold Aschenbrenner expects 30% or more. One of GATE’s authors, Ege Erdil, does not himself predict it, citing model uncertainty and longer timelines. The atlas sets ~30% a year for the best-positioned country from 2030, twice the Gradual rate, on the assumption that regulation, permitting and reallocation constrain growth less over five years than over fifteen. Weighted by GDP, the world then grows about 22% a year, still an order of magnitude above today’s trend. At the ceiling an economy doubles in under three years, well short of Erdil & Besiroglu’s illustrative upper limit of annual doubling.
The approach, 2025–2040: a back-loaded ramp
The model does not keep growth at trend and then introduce a discrete jump. Because tasks are complements, the largest gains arrive when the final tasks are automated: Erdil & Besiroglu show that “a constant rate of task automation … leads to initially slow growth that becomes extremely fast towards the end.” The atlas implements that pattern with a cubic transition from 2025 to each country’s post-AGI rate. In the late 2020s, the path adds a tenth or two of a percentage point a year, in line with skeptical estimates. Around 2032 it adds about one point, in line with the cited Goldman Sachs estimate, and by 2035 about three. The main disagreement concerns the eventual rate rather than the first years of the transition.
The scenario “floors” (5% Gradual, 3% Fast) are not guaranteed rates. Each country’s lower anchor is its own pre-AI trend, capped at the floor, so the least-ready country keeps roughly its old growth and gains only through spillovers that scale with its index: cheaper imported goods and services, remittances and commodity demand. Under Gradual, a country’s rate can never fall below its pre-AI trend because the slower arrival allows time to adapt. Under Fast, it can: countries whose previous growth depended on labor-intensive exports may grow more slowly than they would without AGI. The model represents this export ladder effect by allowing the Fast rate to replace the previous trend. Quotes: Erdil & Besiroglu (2023), abstract, §1 and §3.4 (fig. 2).
Why growth may eventually slow again
These ceilings describe a transition rather than a permanent rate. Fast initial growth would draw on a large stock of newly feasible technologies. It could persist only while research creates valuable opportunities as quickly as firms deploy existing ones, and while energy, materials and construction continue to scale. As catch-up fades and difficult-to-reproduce inputs become binding, growth would probably follow a hump or a sequence of regimes rather than remain at 15–30%. The atlas models acceleration and cross-country divergence but holds each rate constant to 2060; it does not yet model the later slowdown. See Aghion, Jones & Jones on whether essential, hard-to-improve tasks remain, and Binder’s AI industrial-explosion series on the physical limits.
7 · Why geography affects the outcome
Training and inference require different forms of compute, with different geographic concentrations. Training compute consists of the frontier clusters used to develop more capable systems; the United States holds roughly 80% of tracked capacity. Inference compute consists of the data centers that run AI throughout an economy. It is more widely distributed and gives hubs such as Ireland, Singapore and the UAE a role even without frontier training clusters. A country can therefore supply infrastructure to the AGI economy without developing frontier systems itself.
Explosive growth models require power, connectivity, compute, capital, institutions and permission to build, and these inputs are distributed unevenly. High-income countries host 77% of the world’s colocation data center capacity; lower-middle-income countries hold 5% and low-income countries less than 0.1%. Sub-Saharan Africa has about 600 million people without electricity, and the IMF’s AI Preparedness Index averages ~0.34 there against ~0.68 in advanced economies. The rest of the atlas models whether AI narrows or widens those differences across countries.
8 · Other consequences of a change in the growth regime
The atlas simulates growth, but a different growth regime would also affect finance, taxation, distribution, trade, physical infrastructure, government capacity and technological dependence. The questions below arise if the Gradual or Fast scenario is correct. They are conditional and qualitative; the country results do not answer them.
Each topic would require its own research program. Country-specific actions appear in the and ; this section records the wider questions around them.
Who can block deployment
Opposition to a technology is universal; what varies is whether those who lose from it hold the power to stop it. Concentrated losers organize more easily than diffuse winners, so a small group can block a change that would make the country richer overall. Countries whose elites, pensions and public finances are levered to AI paying off tend to push it through; countries that capture little of the surplus have less to offer the losers and more veto points in their way. Watch who owns the upside, and whether governments create new winners — energy discounts near data centers, free model access, AI dividends — instead of arguing the case.
Investment, interest rates and asset prices
Higher returns on AI capital could raise real interest rates and produce large changes in asset values across sectors. A government could face higher borrowing costs even without a domestic boom, because foreign-currency debt is priced against a global benchmark plus a country spread; a surge in frontier investment could therefore raise refinancing costs elsewhere. Interest payments as a share of government revenue would show the effect earlier than debt-to-GDP. During the high-growth periods in Korea and China, posted rates were held down, with some pressure appearing instead through credit rationing and misallocated capital.
Tax bases and public debt
Growth at these rates would increase revenue while changing the bases on which current tax systems depend. Governments tax wages, consumption and profits. If labor’s share falls while returns accrue to model and compute owners that book income elsewhere, GDP could grow faster than taxable domestic income. Governments might need to shift toward rents, windfalls and assets while existing bases become less productive. The composition of revenue would therefore matter as much as its level: Guyana’s GDP roughly tripled in five years while the state’s share of the new economy remained subject to negotiation with one industry.
Labor income and distribution
Cheap machine labor could allow wages and output to move separately: growth could appear through profits, rents and lower prices rather than pay. Ownership of AI systems, together with taxes and transfers, would determine how broadly the gains were shared. The adjustment could also move faster than social registries and safety nets designed for a 3% economy. The labor share of income and the coverage of transfer systems may change earlier than the unemployment rate.
Exchange rates, trade and capital flows
Capital could move rapidly toward countries that offer high returns from AI, including hosts with compute and energy, and away from economies with fewer opportunities. Garments, call centers and back-office services are among the labor-intensive exports most exposed to early automation, so the first stage of the traditional development path could weaken as capital becomes scarcer. Currency and current-account movements should be read together: depreciation raises the domestic cost of foreign-currency debt and of imported hardware. Commodity exporters have experienced a milder version of this pattern for decades.
Energy, land, housing and physical limits
An economy that doubles every few years also increases its demand for power, water, land and housing faster than existing permitting and construction systems have operated. Data centers are the most visible demand; transmission, cooling water, construction labor and housing near expanding industries would also come under pressure. Electricity prices, land prices and rents would register those constraints before an official shortage. Interconnection queues and permitting times indicate how much physical expansion a country can accommodate.
State capacity and decision speed
Budget cycles, procurement systems and courts designed for a 3% economy would face much faster decisions. Delays in approvals, tenders and dispute resolution would directly affect the rate of investment. Korea, Singapore and China created institutions for their periods of rapid growth, including planning agencies, special zones and standing delivery units. The difference between announcement and implementation on major government projects indicates how much of a boom the state could manage.
Technological dependence and sovereignty
Renting frontier capability provides faster access but can be revoked; owning it takes more time and capital and will remain beyond the reach of many countries. Data, workflows and skills accumulate around the selected technology stack, increasing switching costs over time. A supplier’s government can restrict access, and a supplier can raise prices after systems are embedded. Contract terms therefore matter alongside capability, particularly data control, exit rights and access to a second supplier.
Sources for this argument
The full methodology reading list is in Methodology; these are the works this page draws on.