Which world are we in?
The signals that separate the scenarios — and where they stand
Under Gradual and Modest the late 2020s add only fractions of a point to growth, and the two do not separate until the early 2030s; under Fast, US growth would already be above 4% in 2027 and near 8% in 2028. So the question “which world are we in?” is answered by leading indicators, not by GDP. Five tells matter most. Each card gives the latest reading, its trend, the world it is most consistent with so far, and what each scenario would predict for it. Readings are refreshed with the monthly data check (last: 1 October 2026); every source is live, so check it rather than trust the snapshot.
Fast, so far · 3 of 5Gradual, so far · 0 of 5Modest, so far · 0 of 5Consistent with all three · 2 of 5
A reading “consistent with Fast so far” means the indicator is doing what the Fast scenario needs it to do at this point; the leading indicators run ahead of the economic ones by construction. The atlas attaches no probability to any scenario.
Actual growth against the three modeled paths
Black: real GDP growth 2021–2025 (World Bank). Colored: this atlas’s modeled path for the same country under each scenario, ramping from 2025 to the post-AGI rate in 2040 (Gradual) and 2030 (Fast), or to Modest’s peak. Fast separates from the others from about 2027, Gradual and Modest only in the early 2030s, which is why the tells below arrive earlier than the growth data.
Frontier capability
Does the length of tasks frontier models complete on their own keep doubling every four to five months?
Fast, so farThe doubling has held at about 4.3 months for three years with no observed slowdown, which is what the Fast world needs through 2027–28. A Gradual world would show the rate easing toward seven months by now; a Modest one, the curve bending.
Latest reading
≈ 16 hours
METR’s 50%-success time horizon for Claude Mythos Preview (early), evaluated March 2026 — at the top of what its task suite can measure. The doubling time has been about 4.3 months since 2023, against seven months over 2019–2025.
METR page last updated 8 May 2026
Trend · 50%-success time horizon (log scale)
Stylized from METR’s measurements (GPT-2 → Claude Mythos Preview); above 16 h, METR’s current task suite cannot measure reliably.
| Scenario | What it would predict |
|---|---|
| Modest | The curve bends: horizons stall short of economically meaningful work, or success rates fail to keep up with length. |
| Gradual | The rate slows toward the long-run seven months, or horizons plateau at days for several years; month-long tasks arrive in the 2030s. |
| Fast | Doubling holds at five months or less: week-long tasks in 2027, month-long by 2028–29, with reliability rising alongside length. |
Live source: METR — task completion time horizons
AI capital spending
Does spending on AI chips keep rising year on year, so each round of build-out is larger than the last?
Fast, so farQuarterly spending is up about 70% on a year earlier and has risen every quarter since 2023, so the build-out is still accelerating. A Gradual world would show a plateau while revenue catches up; a Modest one, revenue falling short of spending for long enough to slow it.
Latest reading
≈ $76 billion a quarter
Estimated spending on leading AI chips reached about $76 billion in the second quarter of 2026 (5.8 million H100-equivalents), up from about $42 billion a year earlier and $24 billion two years earlier. 2025 as a whole was about $190 billion.
Epoch AI, chip sales estimates as of 1 October 2026
Trend · $bn spent on leading AI chips per quarter
Quarters from Q1 2023 to Q2 2026 (Q3 2026 incomplete). Epoch’s median estimates from chip volumes, prices and earnings commentary.
| Scenario | What it would predict |
|---|---|
| Modest | Revenue falls short of spending for long enough that the build-out slows and frontier compute stops getting cheaper. |
| Gradual | Spending plateaus for a while as revenue catches up, then resumes: no bust, but no runaway either. |
| Fast | Spending keeps rising year on year through the late 2020s, financed by revenue and by capital that expects the frontier to pay. |
Live source: Epoch AI — AI chip sales
Frontier lab revenue
Does frontier lab revenue keep growing several-fold a year, so the spending above is paid for?
Fast, so farSeveral-fold annual growth at the two largest labs is the Fast world’s signature. The question the Gradual and Modest worlds ask is whether it continues once the easy enterprise adoption is done and whether it covers the capital spending above.
Latest reading
OpenAI ≈ $70bn · Anthropic ≈ $65bn
Annualized revenue as reported: OpenAI nearly $70 billion (late September 2026, up from about $13 billion in August 2025), Anthropic about $65 billion (July 2026, up from $5 billion a year earlier). Both series have grown roughly five-fold or more in twelve months.
Epoch AI, company revenue reports to 29 September 2026
Trend · $bn annualized revenue (log scale)
Each point is a public revenue report as collected by Epoch AI (mixed run rate definitions).
| Scenario | What it would predict |
|---|---|
| Modest | Revenue plateaus well below the capital spent; some of the build-out is written off. |
| Gradual | Growth settles to a high but ordinary software rate — tens of percent a year — as adoption deepens slowly. |
| Fast | Revenue grows several-fold a year at the leading labs and broadens beyond them. |
Live source: Epoch AI — AI companies (revenue)
Growth where readiness is highest
Does growth in the ten to twenty most-ready economies, the United States and China above all, break above trend, and when?
Consistent with all threeThis is the tell that settles the question, and it cannot yet. US growth above 4% in 2027 and near 8% in 2028 would say Fast; about three points above trend by the mid-2030s, Gradual; a small bump that fades back to trend, Modest.
Latest reading
At or below trend
Real GDP growth in 2025: United States 2.2% (ten-year trend 2.4%), China 5.0% (5.6%), Germany 0.2%, United Kingdom 1.4%, Japan 1.2%, India 7.6%. Consistent with all three scenarios so far: Fast would separate from about 2027, Gradual and Modest only in the early 2030s.
World Bank WDI, updated 13 July 2026
Trend · % real GDP growth, with ten-year trend
World Bank, 2021–2025. The chart at the top of the page compares these with the modeled paths.
| Scenario | What it would predict |
|---|---|
| Modest | A small bump in the early 2030s that fades back to trend. |
| Gradual | About three points above trend by the mid-2030s, reaching double digits around 2040. |
| Fast | US growth above 4% in 2027 and near 8% in 2028, double digits by 2029. |
Live source: World Bank — real GDP growth
Diffusion
How much of the working-age population uses generative AI, and does use deepen from chat into redesigned workflows and agents?
Consistent with all threeUse is spreading steadily, about a point a quarter, which every scenario allows. What would separate them is depth — agents and redesigned workflows showing up in firm productivity — and no comparable cross-country measure of that exists yet.
Latest reading
18.8% of the world’s working-age population
Generative-AI use reached 18.8% worldwide in the second quarter of 2026, up about a point on the first; 31 economies are above 30%, the UAE highest at 73%. The gap between the Global North (28.8%) and the Global South (16.2%) widened again. Microsoft’s next report will count newer AI tools, which it expects to raise measured use almost everywhere, most of all in China.
Microsoft AI Economy Institute, Q2 2026 report (21 September 2026)
Trend · % of working-age population using generative AI
Microsoft’s global diffusion measure: H1 2025, H2 2025, Q1 2026, Q2 2026.
| Scenario | What it would predict |
|---|---|
| Modest | Use spreads the way the internet did — widely, without a step change in measured productivity. |
| Gradual | Broad use keeps rising while deep redesign lags by a decade; the North–South gap narrows slowly. |
| Fast | Use deepens into agents and automated workflows within a few years and shows up in firm-level productivity. |
Live source: Microsoft — Global AI Diffusion, Q2 2026
How to read the tells. No single reading settles the question, and the atlas attaches no probability to any scenario. The capability, spending and revenue tells lead; the growth tell decides, and it cannot yet. A reading that fits several scenarios at once is a reason to watch the earlier signals, not evidence for any one world. Each country’s brief lists the tells that arrive earlier for that country — contracts, projects, order books and adoption figures — under “How you would know which world you are in”; open one from the .