AI's $3.5 trillion revenue target rests on three premises with measurable empirical tests

Summary

Columbia University finance professor Stijn Van Nieuwerburgh’s projection that AI companies would need $3.5 trillion in annual revenue by 2032 — 8.8% of US GDP under his assumptions — to justify current capital commitments is contingent on three load-bearing premises.
The first premise, that AI companies continue charging today’s scarcity prices for computing power through 2032, faces a measurable test in Epoch AI’s reported 47% per quarter decline in the effective price for a given level of AI capability since 2023.
The second premise, that demand for AI services grows fast enough to offset falling per-unit prices, sits between Ezra Karger’s Federal Reserve expert panel projection of $300 billion in combined OpenAI and Anthropic revenue by 2030 and the post-1984 plateau in business computer spending documented in computing history.
The third premise, that AI productivity gains translate into completed work that customers will pay for, is challenged by a Harvard study of software projects on the Jellyfish platform that found AI’s effect on completed projects was small and “statistically insignificant.”

Columbia University finance professor Stijn Van Nieuwerburgh’s projection that AI companies would need $3.5 trillion in annual revenue by 2032 — 8.8% of US GDP under his assumptions — to justify current capital commitments rests on a small set of premises, several of which are already being tested by data sources that exist today. Van Nieuwerburgh’s analysis, first presented at the Brookings Institution, concludes that the AI buildout is now larger than any investment boom in American history. The Wall Street Journal’s chief economics commentator Greg Ip framed the implied spending in everyday terms: roughly what Americans spend on food, about twice what the nation pays for all forms of energy or for all computers and software, and seven times what consumers spend on phone, streaming, and internet services combined. Ip’s verdict on whether Americans will actually spend that much: “You should be skeptical.”

The calculation and its parameters

Van Nieuwerburgh built the estimate from the published plans of data-center builders and the average cost of a data center. He assumed 4% annual GDP growth, unadjusted for inflation, that some data centers would be canceled, that each dollar of revenue produces 50 cents of cash flow, and that builders require an unleveraged return on invested capital of 10%. The peripheral assumptions — 50 cents of cash flow per revenue dollar, 10% unleveraged ROIC, partial project cancellations — can absorb modest changes without breaking the analysis; a 200-basis-point change in the ROIC threshold or a 10-point swing in the cash-flow conversion rate would adjust the revenue target but leave the framework’s shape intact.

The principal assumption is more structural: that revenue must justify capex at conventional rates of return, with the AI companies themselves earning the revenue. Economic historians who study 19th-century railroad buildouts and the multi-decade electrification of American factories in the late 19th and early 20th centuries have characterized those episodes as cases in which the firms that built the infrastructure often failed financially while the infrastructure itself transformed the rest of the economy. The article itself cites the 1997–2001 fiber buildout as a comparable case: the buildout preceded a long-run expansion of the internet, cloud computing, and mobile services whose revenue accrued to firms that did not own the underlying fiber. Under that reading, the relevant question is not whether AI service companies earn $3.5 trillion but whether the data-center capacity being built enables economic activity whose revenue accrues elsewhere.

Premise one: scarcity pricing holds through 2032

A central assumption is that AI companies can continue charging today’s prices for computing power through 2032. Van Nieuwerburgh said the current scarcity of computing capacity is holding up prices and profit margins, and that the companies are, in remarks quoted by The Wall Street Journal, “basically saying we’re going to keep scarcity pricing in 2032, when presumably there will be tons of competition and we’ve quadrupled our capacity.”

Capacity additions announced by data-center builders make the second half of that statement — quadrupled capacity by 2032 — the more directly observable. If compute prices fall at anything close to the pace that Epoch AI reports for capability costs — 47% per quarter since 2023, roughly six times faster than the price of computing power — the scarcity-rents assumption fails on its own terms. The historical analogue in the article is the 1997–2001 collapse in transatlantic fiber bandwidth prices of 96% that followed the buildout of new capacity and accelerated the bankruptcy of long-haul carriers.

Premise two: demand elasticity beyond the Jevons Paradox

The second assumption is that demand for AI services will grow fast enough to offset falling per-unit prices — a relationship dubbed the “Jevons Paradox” after British economist William Stanley Jevons, who identified the phenomenon in 19th-century English coal consumption. Van Nieuwerburgh’s model effectively requires elasticity greater than one.

The substrate evidence is split. Ezra Karger at the Federal Reserve Bank of Chicago led an expert panel projecting that OpenAI and Anthropic would post combined annual revenue of $300 billion by 2030; Ip called that projection “too conservative,” noting the two companies’ combined revenue run rate was already around $180 billion and was growing at double-digit rates each quarter. Against that, the article documents the post-1984 pattern in which business computer spending tripled to 0.8% of GDP between 1974 and 1984 and then plateaued, apart from secondary booms tied to the internet in the late 1990s and to AI today. Whether that trajectory continues, or saturates the way business computer spending did post-1984, is an open empirical question with current data sources positioned to answer it.

Premise three: productivity translates into revenue

Harvard economics doctoral students Fiona Chen and James Stratton, studying software projects on the Jellyfish platform, found that AI assistants boosted lines of code by 12% and pull requests — defined in the article as “instances when new code is merged into an existing code base” — by 5%, while AI agents boosted lines of code by 30% and pull requests by 23%. Even so, the authors found that AI’s effect on completed projects was small and “statistically insignificant,” because users were spending much more time reviewing, commenting on, and changing pull requests.

The article’s own caution: “AI will almost certainly raise productivity and economic growth, but not necessarily in ways that translate directly into revenue, or into the returns investors are now counting on.” If that finding generalizes beyond software — productivity gains showing up in measured inputs but not in completed outputs — the gap between activity metrics and revenue closes only if the activity metrics themselves become the billable product. The framework’s current formulation assumes they will not.

Five readings of the same figure

The 9% of GDP figure can be read at least five ways, each licensed by the substrate and each yielding a different conclusion about what the number measures.

The commercial-revenue reading treats data-center capex as a private investment requiring a private return, with 10% unleveraged ROIC and 50% cash-flow conversion. Under it, the Harvard findings, the 1997–2001 fiber collapse, and Epoch AI’s 47%-per-quarter capability-price decline cohere as warning signs: prices are falling faster than capability is rising, productivity gains have not yet translated into output that customers will pay for, and history suggests the resulting glut will bankrupt the builders. This is the framework inside which Van Nieuwerburgh’s calculation is built and inside which Ip’s verdict — “You should be skeptical” — operates.

The infrastructure reading treats the buildout as a substrate whose returns are captured downstream by users rather than by the builders. The article’s observation that “the AI buildout is now larger than any investment boom in American history” is, under this reading, consistent with infrastructure-scale investment rather than with service-scale revenue; the investment boom’s size is the data point that supports the buildout, not the data point that threatens it. The 1997–2001 fiber collapse, on this reading, is evidence about the magnitude of the bandwidth infrastructure the buildout created rather than primarily evidence about whether fiber companies would earn their construction costs.

The general-purpose-technology reading treats AI as following computing dynamics more than long-haul fiber dynamics. The article documents this comparison: from the 1960s through the early 2000s, the density of microchips doubled roughly every two years, a phenomenon dubbed “Moore’s Law,” and the real price of computers fell about 15% per year. Under this reading, falling prices do not bankrupt the builders; they expand the addressable market, and the relevant ceiling on AI revenue is the share of GDP currently paid to labor rather than the share currently paid to AI services. The article reports that “some AI advocates argue the technology could become a third factor of production alongside capital and labor, which would make 9% of GDP look modest — labor currently receives about 51% of GDP.” Ip, however, argued the current AI cycle is following a different trajectory from the computer plateau: “AI is improving and adoption is growing much more quickly than with computers, and productivity improvements have been impressive.” The article adds that “the first versions of VisiCalc and Lotus 1-2-3 revolutionized business in ways that ever more powerful versions of Microsoft Excel could not.”

The geopolitical-asset reading treats AI capex as national-security expenditure rather than commercial investment. US national-security commentators and officials from successive administrations have publicly framed AI capability as a strategic asset whose development is justified by considerations that include, but are not limited to, commercial returns. The CHIPS and Science Act of 2022 and the export controls on advanced semiconductors that the Commerce Department’s Bureau of Industry and Security issued starting in 2022 have been documented in public reporting as instruments of an industrial policy that treats compute and chip capacity as strategic assets. 20th-century analogues that national-security commentators have invoked include the Manhattan Project, the Apollo program, and the construction of ARPANET — programs whose strategic objectives, rather than commercial returns, set the floor for federal expenditure. Under this reading, Van Nieuwerburgh’s analysis is a financial critique of an investment whose financial returns are not the objective.

The measurement reading holds that current measures of productivity may be missing the gains. The article reports that “users were spending much more time reviewing, commenting on, and changing pull requests” — output that absorbs labor in ways that GDP does not credit as productive. Economic historians who study the early decades of electrification and computerization have documented episodes in which new technology absorbed labor in quality, review, and integration work that the era’s productivity statistics did not credit, with the gains appearing in measured output only after a lag. Under this reading, the Harvard finding reflects an early phase in which capability gains are absorbed in quality and headroom rather than in output that current statistics capture, and the 9% of GDP figure is the wrong denominator because it measures today’s economy rather than the AI-augmented economy the buildout is designed to produce.

The article’s evidence supports the commercial-revenue, infrastructure, general-purpose-technology, and measurement readings more directly than it supports the geopolitical-asset reading; the geopolitical-asset reading draws on framings of US industrial policy and 20th-century national-security analogues that the article does not cite but that public reporting and economic-history scholarship have documented separately. The Harvard finding is consistent with all five readings — under the commercial-revenue reading, evidence of an unsolved problem; under the general-purpose-technology reading, evidence of an early phase in which productivity is absorbed in quality and headroom; under the infrastructure reading, evidence that the value is being captured downstream of the metric the Harvard researchers measured; under the measurement reading, evidence that current statistics are absorbing the gains in unmeasured work; under the geopolitical-asset reading, evidence orthogonal to the question of strategic capability.

What the tests look like

Each premise has a measurable test on a schedule shorter than 2032. Scarcity pricing would fail if compute prices tracked capability costs downward at anything like the 47%-per-quarter pace Epoch AI reports. Demand elasticity would fall short of the model’s implicit requirement if revenue run rates for OpenAI and Anthropic decelerated before reaching the trajectory the Federal Reserve expert panel projected. Productivity-to-revenue translation would fail if the Harvard finding replicated in domains beyond software.

The framework’s $3.5 trillion target is conditional on all three holding simultaneously. The article notes the framework does not, on its own, settle whether the builders will be the firms that capture the returns that history suggests the buildout will produce — a question the 1997–2001 fiber collapse raises and that the post-1984 computer-spending plateau sharpens.

The model’s report on the buildout’s scale — that it is now larger than any investment boom in American history — is one fact on which the readings agree; what the readings disagree on is whether the scale is a signal of unsustainability under commercial-revenue assumptions, of strategic priority under geopolitical-asset assumptions, or of a substrate whose returns accrue to users rather than to builders. Ip’s verdict in the article: “You should be skeptical” of the 9% figure. The substrate supports that verdict on at least one axis — the Harvard finding on completed projects — and leaves the other two as live empirical questions with data sources already in place to measure them.

Analytical techniques used in this piece

This analysis applies the methods below. Each links to a short, plain-English explainer you can read and reuse.

Paradigm Suspension: Brackets the conventional framing of a problem to see it fresh.
Principal–Agent Problem: An agent acting for a principal has its own interests, which can quietly diverge.