Chapter 8 · Forecasts and scenarios

Whose forecast, in which units?

Big AI numbers differ because they measure different things: global value potential, U.S. GDP levels, productivity growth, jobs supported. They are compared here on their own terms and never added together.

Every row is a forecast or model, not an observation. Global estimates are never relabelled as U.S. estimates.
SponsorPublishedGeographyMeasuresFlow or levelHorizonValueKey assumptionsVerification
Penn Wharton Budget Model2025-09-08United StatesLevel of GDP and productivity vs no-AI baselineLevel effect (cumulative); Real2035 / 2055 / 2075+1.5% (2035), ~+3% (2055), +3.7% (2075)Near-full within ~15 years. Automates 23% of exposed tasks; labor savings 25% rising to 40%. Costs: Not netted. Baseline: No generative AI.verified
Daron Acemoglu (NBER w32487)2024-05United StatesTotal factor productivityCumulative over 10 years; Real10 years≤ +0.66% TFP (< +0.53% adjusted)Task share impacted from exposure studies. Cost savings at task level. Costs: n/a. Baseline: No AI.verified
McKinsey Global Institute2023-06GlobalEconomic value potential of use casesAnnual flow; Nominal (2023 USD)Annual potential, no date$2.6T to $4.4T per yearFull capture of modeled use cases. Assumes redeployment. Costs: Not netted. Baseline: Unclear.not re-verified this pass
SIEPR (Stanford)2026-10United StatesManufacturing jobs supported by data-center capexStock at a run rate; n/a2026 run rate~304,000 jobs (150k to 450k+)n/a. No substitution. Costs: n/a. Baseline: Zero data-center capex.verified

Adoption scenarios

Low, base and high, with a backtest.

Left to choose its own ceiling, the fit runs to the 100% bound: three years of a still-rising series cannot tell us where adoption saturates. So every ceiling here is an explicit assumption: low 65%, base 80%, high 95%. No scenario can exceed 100%. Census history under a consistent question is only one year long, too short to fit honestly, so it is not projected.

Ramp paid adoption: observed, and three scenarios to 2030

Observed only. Switch on scenarios to see assumption-driven paths.

Observed

Observed adoption reaches 56.8%. Backtest predicted 50.9% for 2026-10 versus 56.8% observed.

0%20%40%60%80%100%202420252026Last observationObserved 57%

Backtest. Fitted only on data through 2025-10, the base (80% ceiling) curve predicted 50.9% for 2026-10; Ramp observed 56.8% (error -5.9 points; in-sample RMSE 1.08 points).

Caveat. Logistic curves fitted to a vendor sample. A ceiling is an assumption, not a finding. Paid adoption says nothing about productive use, output or wages.

UnitPercent of businesses on RampGeographyUnited States, Ramp customersObservedObserved Jul 2023 to Oct 2026; scenarios Nov 2026 to Dec 2030SourceRamp AI IndexRetrieved2026-11-18
Inspect the evidence
What this chart does not show
  • National adoption (use Census for that, which has too little consistent history to fit).
  • Intensity of use per adopter.
  • Any economic outcome.
Lineage: source to transformation to chart
  1. Ramp overall series (observed)
  2. Grid-search least-squares logistic fit with fixed ceilings (65%, 80%, 95%); an unconstrained fit hits the 100% bound
  3. Backtest: fit through 2025-10, predict 2026-10

Text description. Observed adoption reaches 56.8%. Backtest predicted 50.9% for 2026-10 versus 56.8% observed.

CSV not offered: Ramp reuse terms not yet verifiedLink

Assumptions lab

Move the assumptions, watch the number.

A first-order model in the style of Acemoglu (2024). It shows which assumptions drive the big forecasts. Every output here is a scenario.

Forecast assumptions lab

First-order level effect on U.S. GDP: labor share × exposed share × tasks changed × cost saving × adoption × implementation success.

Scenario, not a prediction

40% of labor income exposed; 23% of exposed tasks automated; ~30% cost saving (25% rising to 40%); near-full adoption.

Scenario output

+1.44%

higher U.S. GDP level at the horizon than without AI, under these assumptions.

With Penn Wharton-like inputs this returns about 1.4%, close to their published 1.5% by 2035, a consistency check rather than independent support. With Acemoglu-like inputs it returns about 0.65%, in line with his bound of 0.66% TFP over ten years.

Left out on purpose: new tasks and products, demand expansion, reallocation of displaced workers, investment and energy costs, price pass-through, import leakage. Time saved does not automatically become paid output, and higher productivity does not automatically raise employment.