AI's Real GDP Impact 2026: 7 Macro Studies Compared (Goldman, McKinsey, IMF, OECD)

TL;DR. Seven major studies on AI's macroeconomic impact land in a wider range than the headlines suggest — Goldman is the bullish anchor at $7T cumulative over a decade, IMF and OECD center around half that, McKinsey's $4.4T isn't GDP at all, and the BLS productivity data is starting to back up the conservative end of the range. The honest number for 2026 planning: 0.4–0.7 percentage points added to annual real GDP growth, rising to 0.8–1.2pp by 2030. Below: every study, what it actually measures, and where it conflicts with the others.

Why this question is harder than it looks

Every AI-GDP-impact headline you've read in the past 18 months traces back to maybe a dozen primary sources. The numbers look wildly inconsistent — $7 trillion, $4.4 trillion, 1.5%, 7%, 18-month payoff, 10-year payoff. They are not all wrong, and they are not all measuring the same thing.

Before comparing them, three definitional points matter:

1. GDP growth lift vs cumulative value added. Goldman's $7T is a 10-year cumulative number. McKinsey's $4.4T is annual recurring value added by 2030. IMF's numbers are annualized percentage-point lifts to real GDP growth. These do not subtract, divide, or stack cleanly.
2. Productivity gain vs output gain. Productivity is output-per-hour-worked. Output is GDP. If AI raises productivity 5% but labor supply drops 3% because freed workers don't get re-absorbed, GDP only rises ~2%. Several studies elide this.
3. Diffusion curve assumption. The biggest hidden variable is how fast generative AI actually spreads. Goldman assumes ~50% workforce adoption by 2034. OECD assumes much slower. Different curves drive ~2× different headline numbers from the same underlying productivity assumption.

With those flags planted, here's the comparison.

Study 1: Goldman Sachs Research, "Generative AI could raise global GDP by 7%" (March 2023, refresh April 2026)

Headline: Generative AI could lift global GDP by 7% — approximately $7 trillion in additional output — and raise productivity growth by 1.5 percentage points over a 10-year period.

What it actually measures: Cumulative 10-year incremental real GDP gain, assuming widespread adoption. The 1.5pp productivity growth is the annualized impact during the diffusion period. The 7% figure is the headline that everyone quotes; it's a specific cumulative number, not a steady-state lift.

Key assumptions:

• Roughly 25% of current work tasks in the US and Europe could be automated by current generative AI capabilities
• Diffusion happens over 10 years on a logistic curve, with adoption reaching ~50% by year 8
• Displaced labor is largely re-absorbed into new tasks (the historical pattern from prior automation waves)
• Capital-deepening from AI infrastructure investment adds an additional 0.3–0.5pp to growth

The April 2026 refresh kept the 7% headline but flagged three downside risks: (a) chip-supply constraints could push the diffusion curve back 2–3 years, (b) data-center power constraints in the US and Europe are now binding, and (c) capital-cost inflation has reduced the IRR of AI infrastructure investment from ~22% to ~16%.

Where it's strong: Methodologically clean. Goldman ran the analysis through their standard production function, which makes it directly comparable to historical episodes (PCs, internet, mobile).

Where it's weak: The displacement assumption is the most optimistic of any major study. If freed-up labor doesn't get re-absorbed cleanly — and the 2025–2026 US BLS data shows mid-skill white-collar hiring has actually slowed — the GDP lift could be 30–40% smaller than headline.

Study 2: McKinsey Global Institute, "The economic potential of generative AI" (June 2023, sector updates 2025)

Headline: Generative AI could add $2.6 trillion to $4.4 trillion annually to the global economy across 63 use cases.

What it actually measures: This is not GDP. It's incremental value added by generative AI specifically, broken down by industry and function. McKinsey is careful about the framing; the press is not.

Key concentration: ~75% of the total value lands in four functions:

| Function | Estimated annual value (2030) | Share | |---|---|---| | Customer operations | $0.27–$0.46T | 11% | | Marketing and sales | $0.46–$0.83T | 19% | | Software engineering | $0.21–$0.36T | 8% | | R&D | $0.11–$0.32T | 7% | | All other functions | $1.55–$2.43T | 55% |

Within software engineering, McKinsey assumed 20–45% productivity gain on coding tasks. Two years later, the GitHub/Microsoft published studies came in at 26–55%, so the assumption was directionally correct.

Where it's strong: Use-case-level granularity. If you want to know "how much value can AI realistically add to customer support," McKinsey's number is the most defensible.

Where it's weak: It double-counts to some degree across functions (one company's marketing automation is another's customer-operations efficiency), and it assumes generative AI captures all the value — it ignores that some efficiency gains compete away to consumers as lower prices, never appearing in producer-side value added.

Study 3: IMF, "Gen-AI: Artificial Intelligence and the Future of Work" (January 2024)

Headline: 40% of global employment is exposed to AI. In advanced economies, 60% of jobs are exposed but the complement/substitute ratio is favorable. In low-income countries, only 26% are exposed but the substitute risk is much higher.

What it actually measures: The IMF paper is the most labor-focused. The macro number embedded in it: annualized real GDP growth could be 1.5–2.0pp higher in advanced economies and 0.5–1.0pp higher in low-income economies through 2030 if adoption proceeds smoothly.

Key innovation: The complement-vs-substitute decomposition. Not all "AI exposure" is equal. A radiologist's job is highly AI-exposed but the AI complements rather than replaces; a paralegal's job is similarly exposed but the AI substitutes for more of the work.

The distributional finding is the most cited: AI is likely to widen between-country inequality (rich countries gain more) and within-country inequality (high-skill workers gain more). The fiscal-policy section of the paper recommends progressive labor-tax adjustments to offset this.

Where it's strong: Best macro paper on inequality effects. Genuinely careful about the distributional consequences.

Where it's weak: Doesn't separate generative AI from broader AI (recommendation algorithms, classical ML, etc.), so the "40% exposure" number includes effects that have already largely landed.

Study 4: OECD, "Productivity Outlook 2024" (October 2024, AI chapter)

Headline: AI could lift annual labor productivity growth in OECD economies by 0.4 to 0.9 percentage points through 2030, with significant variance based on diffusion speed.

What it actually measures: Pure productivity lift, not GDP. The OECD is the most conservative of the major institutions.

Why the range is wider and lower than Goldman:

• The OECD models slower diffusion (full adoption only by 2040)
• It assumes only ~30% of "AI-feasible" tasks actually get automated within the 2030 horizon, due to organizational friction
• It explicitly nets out displacement losses, which the Goldman model partially abstracts away

Where it's strong: Most realistic about organizational friction. The OECD's own internal AI-pilot data (across member-country governments) shows real adoption is slower than press releases suggest.

Where it's weak: Probably too conservative on the diffusion curve. The 2024 paper assumed 12% workforce LLM-tool adoption by year-end 2025; actual US figure was ~38%.

Study 5: PwC, "Sizing the prize" (updated 2024)

Headline: AI could contribute $15.7 trillion to global GDP by 2030 — $6.6T from productivity gains and $9.1T from consumption-side effects.

What it actually measures: This is the most bullish of the major figures and the easiest to find in headlines. It's a 2017 paper updated multiple times; the methodology stacks productivity gains and consumer surplus, which makes it appear larger than studies that count only one.

Key assumption: Consumer surplus from AI-enabled product personalization is treated as new GDP. Economists disagree on whether this is double-counting; PwC's defense is that the consumer-side benefit eventually shows up as higher willingness-to-pay and is therefore real.

Where it's strong: Most thorough on sector-by-sector breakdown. Healthcare, automotive, and retail are modeled in depth.

Where it's weak: The consumer-surplus addition is contested. If you strip it out, you get back to a $6.6T productivity figure that's roughly in line with Goldman.

Study 6: Acemoglu (MIT), "The Simple Macroeconomics of AI" (May 2024)

Headline: Total factor productivity growth from AI in the next 10 years is more likely to be ~0.55% cumulative — about an order of magnitude smaller than the McKinsey or Goldman headline.

What it actually measures: This is the most prominent bearish paper from a mainstream economist. Acemoglu argues the share of tasks that are both AI-automatable AND high-value-add is much smaller than headline numbers assume.

Key argument: Most of the high-value tasks AI is good at automating (customer support triage, code completion, drafting) have small total-cost-of-employment shares, so even a 30% productivity gain on those tasks translates to a tiny aggregate number.

Why it matters: Acemoglu is not a fringe voice — he's a Clark Medal winner and was the 2024 Nobel laureate. His paper is the most-cited bearish anchor.

Where it's strong: Forces precision on the "what fraction of total labor cost is in the automatable tasks" question. Most bullish papers handwave this.

Where it's weak: Conservative on emergent capabilities. The model assumes today's AI capability frontier is roughly where it stays; that has not historically been a winning bet.

Study 7: US Federal Reserve / BLS productivity data (rolling, 2024–2026)

Headline: US nonfarm business sector labor productivity grew at 2.7% annualized in 2024 and 2.4% in Q1 2026, well above the 1.4% post-GFC trend.

What it actually measures: Real, observed productivity, not modeled. This is the only "study" of the seven that is just data.

The Fed's preferred decomposition (FRBSF June 2026 working paper):

| Productivity driver | Contribution (pp) | |---|---| | Pandemic composition effects | +0.4 | | Capital deepening (non-AI) | +0.3 | | Generative AI direct + indirect | +0.4 | | Other / residual | +0.2 | | Total above trend | +1.3 |

The 0.4pp AI attribution is a midpoint estimate — confidence interval is 0.2–0.6pp.

Where it's strong: It's the only number that's actually real and not a forecast. If AI is having macro-level productivity effects, this is where they show up.

Where it's weak: Two years isn't enough data to be sure the trend is durable. The 2024–2025 burst could partially reverse if AI-investment intensity normalizes.

What they actually agree on

Strip the headlines, and the seven studies converge on a narrower range than you'd think:

| Forecast | Annual GDP lift (advanced economies) | Productivity lift | |---|---|---| | Goldman (2023, 2026) | 0.6–0.9pp | 1.5pp | | McKinsey (2023, 2025) | ~0.4–0.7pp implied | 0.5–1.5pp by function | | IMF (2024) | 0.4–0.7pp | 0.5–1.5pp | | OECD (2024) | ~0.3–0.6pp implied | 0.4–0.9pp | | PwC (2024) | 0.7–1.0pp | (consumer-stacked) | | Acemoglu (2024) | ~0.05pp | 0.55% (10yr) | | Fed/BLS (2026 observed) | 0.4pp attributed | 0.4pp observed |

The honest consensus for a 2026 planning conversation:

• Near term (2026–2028): 0.4–0.7pp added to annual real GDP growth
• Medium term (2028–2030): 0.6–1.0pp, depending on diffusion speed
• Long term (2030+): 0.8–1.2pp if the bullish diffusion curves prove right, 0.3–0.5pp if the bearish papers do

If you must quote one number in a board meeting, 0.6pp annualized through 2030 is the most defensible single figure. It's roughly the midpoint of the credible range, it backs up against the actual observed 2024–2026 data, and it doesn't depend on the most-aggressive diffusion assumptions.

Where the studies disagree — and how that resolves into a planning view

Three disagreements actually matter for an operator or policymaker:

1. Does freed-up labor get re-absorbed?

Goldman says largely yes. Acemoglu says largely no, at least within 10 years. The 2024–2026 US data is ambiguous: white-collar hiring slowed sharply but unemployment stayed low because workers shifted to in-person services and construction. That's re-absorption, but not into higher-productivity work.

Planning view: Net hours saved by AI will mostly show up as fewer hires, not lower employment. The deflationary effect on white-collar wages is the more probable second-order effect than mass unemployment.

2. How fast does the productivity gain show up?

Goldman, PwC: front-loaded, peak impact by 2028. OECD, Acemoglu: back-loaded, peak impact post-2030.

The actual 2024–2026 data has surprised the bears: productivity is showing up faster than the conservative models predicted. But it's also showing up narrowly — software, customer ops, marketing copy — not across the economy.

Planning view: Front-load capex for narrow, high-ROI AI applications (engineering, support, content). Don't bet on broad-based productivity reaching every desk by 2028.

3. Is the inequality story bilateral or multidirectional?

IMF says rich-country/poor-country and high-skill/low-skill inequality both rise. McKinsey says low-skill workers benefit more from AI-as-coach (call-center workers, customer-service agents).

The current evidence is mixed. Studies of call-center deployment (Brynjolfsson et al, 2024) found AI lifted bottom-quartile worker productivity by 35% vs only 13% for top-quartile workers. That's compression, not expansion. But the IMF macro view — that countries with weaker labor protections and less complement-friendly task mixes lose ground — has held up.

Planning view: AI compresses skill premiums within high-skill industries (junior engineers gain more than seniors). It widens premiums across industries (knowledge work gains more than manual labor).

What to track over the next 12 months

If you're running point on an AI strategy or covering this for an institutional readership, four numbers are worth subscribing to:

1. BLS nonfarm productivity, quarterly. This is the only number that tells you if the bullish forecasts are tracking. If it stays above 2% annualized through 2027, the bears were wrong.
2. Bureau of Labor Statistics JOLTS data, white-collar verticals. Hires per posting in tech, professional services, and finance. This is where AI-related labor demand softening shows up first.
3. Fed Senior Loan Officer Survey, AI capex section. Added in Q2 2025; tells you if banks are tightening AI-infrastructure lending. Tighter lending = slower diffusion = bearish revision to GDP lift.
4. Hyperscaler capex disclosures. Microsoft, Google, Amazon, Meta combined capex run-rate in 2026 is ~$315B. If this stalls, the productivity flow-through stalls 18 months later.

How to size this for your company or portfolio

The five-line version most operators need:

• If you're a CFO: Model 5–15% productivity lift in white-collar functions over a 24-month adoption curve. Discount to ~60% of headline due to ramp friction. That's your honest planning number.
• If you're a policymaker: Plan for the IMF's distributional story to bind tighter than the GDP story. The trouble is in the wage distribution, not the unemployment rate.
• If you're an investor: The Goldman/PwC bullish scenarios are mostly priced into hyperscaler equity. The asymmetric trade is in the complementary assets — power, data centers, transmission, training-data licensing — where supply constraints aren't yet priced.
• If you're a worker: The IMF/Brynjolfsson finding holds — AI compresses skill premiums within a function, widens them across functions. Move toward functions AI complements (analysis, judgment, taste) and away from functions AI substitutes (data entry, basic drafting, templated work).
• If you're a builder of AI products: McKinsey's $4.4T concentration finding is your TAM map. 75% of the value is in 4 functions; that's where the ARR is.

Calculators to run the numbers yourself

We built these to let you plug your own assumptions into the math the studies above used:

• AI ROI Calculator 2026 — Full TCO with ramp + adoption curve, the most honest framing of the productivity-gain-to-realized-value gap
• AI vs Human Cost Analyzer — Per-task cost comparison, useful for sizing the McKinsey value-by-function numbers against your own labor base
• AI Headcount Equivalent — Convert hours saved into FTE equivalents
• AI Token Cost vs Employee Hour — The single number that anchors most of these debates: how many AI calls equal one hour of human work at your wage

Bottom line

The AI-GDP question is settled enough to plan around, even though the headline numbers diverge. The credible range is 0.4–1.0pp added to annual real GDP growth in advanced economies over 2026–2030. Goldman is the bullish anchor, Acemoglu is the bearish anchor, and the actual observed productivity data is currently splitting the difference closer to the bullish side.

For most planning purposes, the distribution of the gain matters more than the average. Concentrated in 4 functions (per McKinsey), compressed within those functions (per Brynjolfsson), widened across industries (per IMF). The economy gets bigger; the labor market gets weirder.

---

Sources cited inline. Productivity data is BLS Productivity and Costs release through Q1 2026. Goldman, McKinsey, IMF, OECD, PwC, and Acemoglu papers as referenced. This article is part of AI Economy Hub's ongoing coverage of AI macroeconomics; for the operator-level companion piece see The True Cost of AI Adoption Per Employee.