Everything About AI
NextFin News - The U.S. economy could reach $44 trillion by 2030, nearly 50% above today's level, if artificial intelligence keeps improving at its current pace, Anthropic co-founder Jack Clark told finance ministers and central bankers gathering in Bangkok. The figure is the extreme end of a three-scenario forecast from an internal Anthropic economic model — one that also maps out a modest $34.1 trillion path and a substantial $36.3 trillion path. Clark himself called the number "very outlandish," then added the line that carries the real weight of his argument: "it's implicated by the technology continuing to get better."
"I know, but it's implicated by the technology continuing to get better. And if GDP were to grow this much, then economic doctrine suggests that new industries and inventions would need to have been created to power such a major economic expansion."
The forecast, delivered on October 17, 2026 at the IMF and World Bank annual meetings, is less a point prediction than a stress test. It asks what the world looks like if AI systems reach recursive self-improvement — the point where models can make themselves better — by 2027, and if the productivity gains from that capability spread through the economy fast enough to create industries that do not yet exist.
The Forecast: Three Scenarios, One Outlandish Ceiling
Clark framed the $44 trillion figure against current U.S. economic output of just over $30 trillion. Official data puts the gap in sharper relief: nominal GDP stood at $32.5 trillion at an annual rate in the second quarter of 2026, according to the U.S. Bureau of Economic Analysis, up from $30.8 trillion for all of 2025. Reaching $44 trillion by 2030 would require roughly 32% more real economic activity in four years — a pace with few peacetime precedents outside of postwar rebuilds or the late-1990s technology boom.
The projection comes from Anthropic's "Econ Scenario Explorer," an interactive model the company released in September 2026 alongside a technical report, "Economic Scenarios for Transformative AI," produced by the Anthropic Institute's economics team. The model treats the economy as a bundle of tasks that AI can augment, automate, or create anew, and it lays out three paths to 2030, calculated at 2025 price levels:
- Modest: U.S. GDP reaches $34.1 trillion, a 1.6% boost over a no-AI baseline. Anthropic frames this as an internet-scale impact — real, but hard to see in the macro aggregates.
- Substantial: GDP rises to $36.3 trillion, an 8.3% boost — roughly twice the normal growth rate over the period.
- Extreme: GDP hits $44.4 trillion, a 32.4% boost. This case assumes AI reaches high levels of autonomy and efficiency, and that recursively self-improving systems deliver the full effect.
The public, for its part, is not betting on the moonshot. In an August survey of 10,980 Americans conducted with Morning Consult, respondents were asked five questions that mirror the explorer's variables — what tasks AI can do, how much people use it, how autonomous it is, how much new work it creates, and how productivity responds. The typical respondent's answers lined up closest to the substantial scenario: about a 10% GDP boost by 2030, with the overall unemployment rate around 5%. Only about 10% of respondents held views consistent with the extreme case. The people whose jobs are most exposed to the technology, in other words, see meaningful change coming — but not a revolution.
Why $44 Trillion Requires New Industries, Not Just Faster Workers
The extreme scenario is not a simple extrapolation of today's productivity gains. Sustained growth at that scale cannot come from squeezing more output out of existing tasks. Clark stated the mechanism plainly at the meeting: "if GDP were to grow this much, then economic doctrine suggests that new industries and inventions would need to have been created to power such a major economic expansion."
That is a Solow-style insight dressed in frontier-AI language. Robert Solow's famous 1987 observation — "you can see the computer age everywhere but in the productivity statistics" — was resolved only when the internet created entirely new categories of economic activity: e-commerce, digital advertising, app economies, cloud computing. The 1995–2000 productivity boom, which lifted labor productivity growth to 2.5% from 1.4% in the prior period, was driven by information-technology investment growing at an average of 24% a year and contributing more than three-quarters of a percentage point to annual GDP growth, according to Federal Reserve Bank of San Francisco research.
The parallel is instructive, and it cuts both ways. The 1990s boom required three things that had to arrive together: cheap computing hardware, a general-purpose network, and new business models that could monetize it. Clark's extreme scenario requires the same kind of convergence: AI capabilities that keep improving, adoption that keeps pace with capability rather than stalling at early adopters, and labor that is redeployed into newly created work rather than sitting idle.
Clark moved up his own timeline for the first of those links. He now expects full recursive self-improvement capabilities by 2027 — a shift he tied directly to faster progress in the technology. That date is the hinge between the substantial and extreme outcomes. If systems can improve themselves on that timeline, the rate of scientific discovery could accelerate in fields from drug development to materials science, compounding into the kind of invention wave the model requires. If they cannot, the forecast collapses back toward the substantial case.
The model's own architecture explains why the gap between the scenarios is so wide. It does not ask "how much faster will workers become?" It asks "what share of tasks can AI perform, how widely will it be adopted, how autonomously will it operate, and how much new work will it create?" Small differences in those inputs compound into the $10 trillion spread between the modest and extreme outcomes. That is the difference between AI as a tool and AI as an inventor.
The Bear Case: Acemoglu's 0.7% and the Adoption Wall
The strongest counter-thesis comes from Daron Acemoglu, the MIT economist and 2024 Nobel laureate, who has estimated that AI will raise U.S. productivity by roughly 0.71% over a decade — a small fraction of what the extreme scenario requires. Other mainstream estimates cluster far below Anthropic's ceiling: Goldman Sachs projects AI adding about 1.5% to annual productivity growth over ten years, while the OECD models barely 0.5%.
Acemoglu's argument is not that AI is useless. It is that most tasks cannot be fully automated, that firms adopt new technology slowly, and that the tasks AI can do profitably today account for a limited share of total economic value. The adoption channel is the bear's strongest ground: capability does not equal deployment. Regulation, liability rules, legacy IT systems, and organizational inertia all sit between a working model and a rewired economy.
History offers a cautionary analog. The personal computer arrived in the 1980s, but the productivity payoff took fifteen years to show up in the statistics — the so-called Solow paradox. The lesson is not that the technology was overhyped. It is that the economic return depends on complementary investments: reorganized workplaces, new skills, updated regulation. Those complements take time, and time is the one input the extreme scenario assumes away.
"I know, but it's implicated by the technology continuing to get better."
There is also a physical constraint the model must contend with. Data centers are already the fastest-growing source of electricity demand in the U.S. power system. The International Energy Agency projects global data-center electricity consumption will more than double by 2030, reaching about 945 terawatt-hours — just under 3% of total global consumption — with the United States accounting for nearly 45% of the market and a 130% increase in data-center demand by the end of the decade. The Department of Energy's Lawrence Berkeley National Laboratory found data centers consumed about 4.4% of U.S. electricity in 2024, and the Electric Power Research Institute estimates that share could reach 9% by 2030. An economy expanding at the extreme-scenario pace would need commensurate growth in compute, and commensurate growth in the power plants to run it.
The Second-Order Question: Who Owns the Gains
There is a second-order implication that the headline growth number hides. Reviewers of Anthropic's work, including Acemoglu and David Autor, pushed the company to model a distributional channel: as AI makes capital more productive, a larger share of the expansion accrues to whoever owns the compute, the data, and the models — not to labor.
This reframes the headline. A 32% rise in GDP is not automatically a 32% rise in median household income. If the expansion concentrates in the owners of frontier infrastructure — chipmakers, cloud providers, model developers, and the institutional investors behind them — the macro aggregate can soar while the median worker sees wage pressure or displacement. The extreme scenario itself anticipates unemployment spikes in some configurations. Growth and shared prosperity are not the same variable.
This is why Clark's regulatory message is inseparable from his forecast. He told the Bangkok audience that international cooperation is needed to erect "unusual and serious rules" for the transformative technology.
Clark said international cooperation was needed to erect "unusual and serious rules" for the transformative technology.
A world where AI adds $10 trillion of output but concentrates it requires different policy — tax systems, social safety nets, labor-market interventions — than a world where the gains diffuse broadly. The model is, in Clark's framing, a tool for stress-testing assumptions rather than a prophecy: set your variables for capability, adoption, autonomy, and redistribution, and see which 2030 you get.
What Would Prove It Wrong
The strongest falsifying signal is the 2027 recursive-self-improvement timeline. If frontier systems have not demonstrated the ability to improve themselves by then, the engine of the extreme scenario is missing, and the forecast reverts to the substantial case. A second signal is business investment: sustained growth in AI-related capital expenditure and software spending is the adoption tell. If capex stalls while capability advances, the adoption wall Acemoglu describes is real. A third signal is labor-market data for knowledge workers — displacement showing up in employment and wage figures would move the debate from conference panels to legislation quickly.
And there is a market signal already flashing. Technology shares have been volatile as investors weigh AI's promise against its cost. The Nasdaq recently posted its worst week since late June with a 4.5% slide as the AI trade suffered, and an earlier Anthropic tool launch in February 2026 triggered a one-day software sell-off that wiped more than $280 billion from software and data-centric stocks. The market is pricing both the upside and the disruption — sometimes in the same week.
The Verdict: Structural Claim, Not Cyclical Forecast
Clark's $44 trillion is a structural claim about a regime shift in how discovery and production work, not a cyclical forecast. Cyclical booms revert; they are driven by inventory swings, credit cycles, and demand pulses that fade on their own. A structural shift changes the production function itself — the set of tasks the economy can perform and the cost of performing them. If Clark is right, the 2030 economy will not look like a linear extension of today's. If Acemoglu is right, the AI boom will show up in corporate margins and stock buybacks long before it shows up in GDP.
The base case is the substantial scenario: meaningful but not revolutionary growth, with adoption lagging capability and labor markets adjusting unevenly. The upside case requires recursive self-improvement arriving by 2027 and spreading into scientific and industrial discovery fast enough to create new industries at pace. The downside case is the Acemoglu world: slow adoption, limited task automation, and growth barely above trend.
For investors, the asymmetry is clear. The substantial scenario supports continued spending on AI infrastructure — chips, data centers, power — even if the macro payoff disappoints, because that spending is the price of admission to either future. The extreme scenario, if it materializes, rewards the owners of the invention machine most of all. The downside scenario punishes only those who paid for revolution and got efficiency instead.
The $44 trillion number is not really a GDP forecast. It is a bet that AI will invent industries we cannot yet name. The economy will either validate the invention machine, or it will validate the skeptics — and the 2027 self-improvement milestone will tell us which.
Explore more exclusive insights at nextfin.ai.
