Goldman Sachs says AI contributed 'basically zero' to US GDP in 2025. The St. Louis Fed says AI-related investment accounted for 39% of GDP growth. Same economy. Same year. Same dollars. The disagreement is not about AI. It is about what counts.
Goldman Sachs Chief Economist Jan Hatzius told the Atlantic Council this week that AI investment contributed 'basically zero' to U.S. economic growth in 2025. His exact words: 'We think there's been a lot of misreporting of the impact that AI investment had on GDP growth.'
The same week, data from the Federal Reserve Bank of St. Louis showed that AI-related investment accounted for 39% of total GDP growth in the first nine months of 2025 — surpassing the dot-com era's peak contribution of 28% in 2000.
Six hundred and fifty billion dollars in AI infrastructure spending. Two institutions staffed by serious economists. One says the impact is nearly nothing. The other says it is the primary engine of growth. Both are correct, which is the problem.
Three Readings of One Economy
Start with what Hatzius actually said. His argument is not that the investment is fake or that AI doesn't work. It is that most AI infrastructure spending doesn't register as U.S. GDP because it isn't produced in the United States. When Microsoft buys NVIDIA GPUs, those chips were fabricated by TSMC in Taiwan, using high-bandwidth memory from SK Hynix in Korea, assembled into servers largely in Asia. The American company writes the check. The Asian factories add the value. In national income accounting, imports subtract from GDP. A dollar spent on an imported GPU adds to Taiwanese and Korean GDP, not American.
Hatzius's number isn't wrong. It is a precise answer to a precise question: how much domestic value-added did AI equipment investment contribute to U.S. GDP? The answer, after adjusting for imports, is close to zero.
The St. Louis Fed is answering a different question. Their measure — 0.97 percentage points of GDP growth across four categories (information processing equipment, software, R&D, and data center construction) — captures the full investment impulse, including the domestic demand stimulus. When a hyperscaler spends $40 billion building data centers in the United States, that construction spending counts. When American software companies sell AI tools to American enterprises, that counts. The 39% figure is not about chips. It is about the entire ecosystem of spending that AI catalyzes.
Then there is a third reading — Goldman's own. In September 2025, a different Goldman Sachs research team calculated that AI had added $160 billion to what they called 'true GDP' since 2022, roughly 0.7% of output. Their method: take company revenue data, subtract inflated prices, foreign sales, and imported inputs. The result is three and a half times larger than the $45 billion that shows up in official statistics. The gap exists because the Bureau of Economic Analysis counts semiconductors only when they are embedded in a final product that a consumer buys. A GPU sitting in a data center running inference is, in GDP accounting terms, an intermediate good. It doesn't count until something downstream does.
The same bank. Two teams. One says zero. The other says $160 billion. Neither is lying. They are measuring different things with instruments calibrated for different purposes.
The Instrument Shapes the Reading
GDP was designed in the 1930s and 1940s to measure industrial output — tons of steel, bushels of wheat, units of automobiles. It counts transactions. It counts production. It does not count time saved, decisions improved, or capabilities created but not yet monetized. When a radiologist uses an AI system that cuts diagnostic time from fifteen minutes to two, GDP registers the same billing code at the same price. The twelve minutes saved are invisible.
This is not a new critique. Economists have debated GDP's blindness to quality improvements and consumer surplus for decades. Robert Solow's 1987 observation — 'You can see the computer age everywhere but in the productivity statistics' — has been cited so often it qualifies as liturgy. What is new is the scale of the mismatch.
JP Morgan Asset Management published an analysis showing that AI-related capital expenditure now represents 1.2% of U.S. GDP — exceeding the telecom infrastructure buildout of the early 2000s, which peaked at 1.0%. Only the railroad boom of the 1880s, at 6.0% of GDP, was proportionally larger. Computers and related equipment investment rose 41% year over year. Data center construction reached $40 billion at an annual rate. And yet: 'Not every AI dollar will translate directly to U.S. GDP,' JP Morgan notes, 'as much investment goes toward imported technology goods, which subtracts from GDP.'
The largest infrastructure investment since railroads is being evaluated by an accounting system that subtracts most of the spending because the hardware is manufactured overseas, ignores the intermediate goods because they haven't reached a consumer, and cannot see the productivity gains because they show up as the same service delivered in less time.
The J-Curve or the Mirage
There is a familiar story that absorbs this tension. Every general-purpose technology follows a J-curve: heavy investment first, productivity gains later. Railroads in the 1850s consumed enormous capital before the transcontinental network transformed commerce in the 1870s and 1880s. Electrification took three decades from Edison's first power station to the factory productivity revolution of the 1920s. The IT investment boom of the 1990s preceded the productivity surge of the late 1990s and early 2000s by nearly a decade. Research on intangible capital and organizational change suggests the AI productivity J-curve may take five to eight years to resolve.
The J-curve argument says: the returns are real but deferred. Be patient. The infrastructure is the prerequisite, not the product. A paper from CEPR — the Centre for Economic Policy Research — puts it precisely: 'The largest economic effects of AI are still likely to come from productivity gains and organisational change rather than from capital spending itself.'
The counter-argument is 1999. The telecom buildout consumed $1 trillion in capital, laid fiber across the ocean floor, and bankrupted most of the companies that financed it. The infrastructure was real. The demand projections were not. WorldCom, Global Crossing, and 360networks did not fail because fiber was useless. They failed because the market for bandwidth grew slower than the supply. The capex was real. The returns never matched.
Whether AI is railroads or telecom is not knowable from inside the investment phase. You cannot distinguish a J-curve from a bubble until the curve either turns or doesn't. The honest answer — from inside February 2026 — is that the data is genuinely ambiguous. The hyperscalers are spending 90% of their operating cash flow on AI infrastructure. NVIDIA just guided $78 billion for next quarter. The acceleration is real. Whether the returns match the acceleration is a question the current data cannot answer.
What the Disagreement Reveals
The interesting observation is not which number is right. It is that the disagreement itself is diagnostic.
When serious institutions produce wildly different measurements of the same phenomenon, the phenomenon has outgrown the instrument. GDP was built to measure an economy where value creation and value capture happened in the same country, where goods were physical and countable, where productivity gains showed up as more output per hour. AI violates every one of these assumptions. The value creation is global (American companies, Asian fabrication, worldwide deployment). The goods are intermediate and intangible (inference, not inventory). The productivity gains show up as less time per task, not more output per hour.
This does not mean GDP is wrong. It means GDP is answering a question that is decreasingly relevant to what is actually happening. When the instrument cannot resolve the phenomenon, you do not need a better measurement of the old thing. You need a new question.
Goldman's $160 billion 'true GDP' figure is an attempt at a new question — one that accounts for domestic value-added from AI regardless of whether it shows up in the BEA's standard categories. The St. Louis Fed's decomposition is another attempt — measuring the investment impulse rather than the net domestic production. Each captures something real. None captures the whole.
The lag is not just temporal — investment now, returns later. It is epistemic. The economy is changing in ways the measurement system was not designed to see. By the time the instruments catch up, the transition will have moved on to something else the instruments cannot see either.
This is what the inside of a phase transition looks like. Not clarity arriving gradually. Contradictory measurements of the same reality, each internally consistent, none sufficient. The signal is in the disagreement, not in any single number.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)