The numbers were always going to be large. What nobody expected was for them to be simultaneously enormous and nearly worthless — at least from the perspective of American economic growth. Goldman Sachs Chief Economist Jan Hatzius has delivered what may be the most sobering data point of the AI boom: despite the biggest peacetime capital investment surge in US history, artificial intelligence contributed "basically zero" to American GDP growth in 2025.
That's not a critique of AI's long-run potential. It's a structural accounting problem — one that reveals something important about where the money is actually going, and who is actually benefiting.
The Import Problem Nobody Was Talking About
Speaking at the Atlantic Council, Hatzius was unusually direct about what he called widespread misreporting of AI's economic impact. Of the 2.2% US GDP growth recorded in 2025, AI investment contributed only around 0.2 percentage points — a figure that already looks modest against the hype. But the real story is starker still.
GDP measures domestic production, not domestic spending. It counts what is made inside the country, not what is bought. When American tech companies spend hundreds of billions on AI infrastructure, a significant share of that spending flows directly to imported components — the GPUs, memory chips, and specialized hardware that power data centers. Roughly 75% of the cost of building a data center comes from imported parts, according to TechRadar's analysis.
The result is an economic paradox. As Hatzius put it, a significant portion of US AI investment "adds to Taiwanese GDP, and it adds to Korean GDP" — countries that manufacture the semiconductors and memory chips that underpin the AI build-out. The domestic investment and the import outflows cancel each other out in the GDP calculation, leaving the US running on something close to an economic treadmill.
Tech companies were projected to spend $700 billion on AI infrastructure in 2025 alone. The US economy registered almost none of it as growth.
The Inflation Trap No One Modeled
The GDP story is only part of Goldman's increasingly complicated picture of AI's near-term economics. A separate Goldman research note from economist Manuel Abecasis found that AI-related price pressures have already added roughly 0.3 percentage points to annual core PCE inflation and 0.1 percentage points to core CPI over the past year — with similar pressure forecast for the next twelve months.
Three channels are driving this inflation. First, surging demand for AI infrastructure has pushed up prices for digital memory, storage batteries, and related electronics across the consumer supply chain. Second, a tight labour market for AI engineers has created wage pressure in high-skill technology roles. Third — and perhaps most consequential — data centres are consuming enormous amounts of electricity. After more than a decade of flat US electricity production, output rose 2.5% in 2024, 2.4% in 2025, and was running 3% higher year-over-year in early 2026, largely driven by data centre demand. Consumer electricity prices are up 4.6% year-over-year as a result.
Stifel analyst Thomas Carroll flagged the magnitude of this shift, calling it a macro "regime shift" and noting that 2026 marks the first time in 65 years that tech goods prices are rising faster than wages. Goldman's own economists estimate electricity costs will add 0.1 to 0.2 percentage points to headline PCE inflation over the next couple of years — making AI, paradoxically, a headwind to the purchasing power of the workers it's supposed to benefit.
And as of March 2026, Goldman found no meaningful relationship between AI adoption and economy-wide productivity — though the bank noted isolated gains of around 30% on specific tasks where companies were actively measuring outcomes.
The Physical Economy: Where Goldman Sees the Next Wave
If the first chapter of the AI boom was about software, cloud, and frontier models, Goldman's latest research argues the second will be written in steel, electricity, and industrial automation. In a report shared exclusively with Axios, Goldman's banking division says AI's next frontier is the physical economy — factories, mines, utilities, and oil rigs.
"We're only just beginning to see AI's impact on industrial businesses," Mark Sorrell, Goldman's global head of industrials, told Axios. Goldman estimates roughly $7.6 trillion will be invested globally in AI infrastructure between 2026 and 2031 — spanning compute, data centres, and power infrastructure. The bank's framing is deliberately expansive: software accounts for less than 0.5% of global GDP, leaving "the other 99.5%" of the economy as AI's next frontier.
The urgency in corporate boardrooms is already palpable. "There's definitely a little bit of, 'If I don't move, do I get left behind?'" Sorrell noted of conversations with industrial clients. Tech M&A has already reached $566 billion in 2026, up sharply from $334 billion in all of 2025 — a signal that capital is moving, fast, from pure AI plays toward industrial deployment.
Goldman's Jung Min, global co-head of technology, media, and telecommunications, pointed to a structural shift in how tech and non-tech companies relate to each other: AI has made technology companies more dependent on non-tech firms, reversing a decade-long trend in which digital companies were largely self-contained. The convergence is no longer just a trend slide in a pitch deck — it is showing up in deal volumes, in boardroom conversations, and in the increasingly blurred lines between a chip company, a power utility, and an industrial manufacturer.
The 2027 Horizon and the Jobs Equation
Goldman's economists have maintained for some time that the real AI productivity payoff begins around 2027 — when widespread adoption across enterprises is expected to start showing up in labour productivity data. A 2023 Goldman Research report estimated AI could increase US productivity growth by 1.5 percentage points annually assuming widespread adoption over a ten-year period, a transformation that would rank among the most significant in economic history.
But the bank's own more recent research complicates that optimism. A 2025 Goldman report warned that AI could displace 6-7% of the US workforce if adopted broadly — though the bank characterises this impact as "transitory," with new job opportunities eventually absorbing displaced workers. That framing is contested. The Fortune analysis of Goldman's inflation research found that Gen Z's excitement about AI has dropped 14 percentage points over the past year to just 22%, while anger has risen 9 points to 31% — and the backlash is sharpest among daily users, suggesting familiarity breeds frustration as much as enthusiasm.
Goldman frames the broader macro picture as "up then down" — near-term inflation and economic dislocation, followed by disinflation and productivity gains once AI adoption becomes sufficiently widespread. But the St. Louis Fed has already warned that if policymakers ease interest rates prematurely based on AI productivity optimism that doesn't materialise on schedule, the result could be persistently elevated inflation. The timing risk is real, and it cuts both ways.
What It Means for AI Investors
The Goldman analysis carries specific implications for anyone tracking AI-adjacent markets. The data centre build-out is real and accelerating — as evidenced by deals like the Switch infrastructure IPO targeting an $80 billion valuation, detailed in our coverage of the data center gold rush financing the AI era. But the economic value of that build-out is, for now, accruing disproportionately to Asian hardware manufacturers and to the electricity sector, not to US GDP or corporate productivity metrics.
JP Morgan estimated in late 2025 that AI infrastructure would need to generate over $600 billion in annual revenue just to achieve a 10% return on investment. OpenAI's entire revenue for 2025 was less than $20 billion — while its capital expenditure projections have been revised toward $600 billion by 2030. As we covered in our analysis of OpenAI's $852 billion IPO S-1, the company has lost $14 billion while being valued at nearly a trillion dollars. The gap between capital deployed and revenue generated is not a rounding error; it is the central financial question of the AI era.
For enterprise buyers, the picture is similarly double-edged. AI tools are generating real productivity gains in specific, measurable workflows — but as our coverage of the "tokenmaxxing" crisis shows, unmanaged AI consumption can burn enterprise budgets at scale without commensurate output. The productivity payoff is real but conditional — on deliberate deployment, active measurement, and cultural adoption that most enterprises haven't yet achieved.
Meanwhile, the competitive dynamics of who controls AI supply chains remain in flux. Chinese AI providers — DeepSeek, GLM, Xiaomi MiMo — are capturing an outsized share of global inference demand at radically lower cost, as our analysis of Chinese models owning 45% of OpenRouter traffic showed. The import problem Hatzius identified at the hardware layer may be about to repeat itself at the model layer.
The Verdict: Real Investment, Deferred Returns
Goldman's body of research on AI economics is neither bullish nor bearish in the simple sense. It is structural. The investment is real. The infrastructure being built is genuinely transformative. The 2027 productivity horizon is a reasonable forecast, not a fantasy.
But the near-term math is unambiguous: the US has spent hundreds of billions on AI infrastructure, transferred much of the manufacturing benefit to Taiwan and South Korea, introduced new inflationary pressures into an economy still wrestling with elevated prices, and generated almost no measurable productivity gain at the aggregate level — yet.
The question is not whether AI will matter economically. It is whether the gap between investment and return will close before the capital markets that have funded this boom decide it is taking too long.
Related Reading
- Switch's $80 Billion IPO and the Data Center Gold Rush Financing the AI Era
- OpenAI's $852 Billion IPO: Inside the S-1 That Could Reshape Silicon Valley Forever
- The "Tokenmaxxing" Crisis: How Enterprises Are Burning AI Budgets
- Chinese Models Now Own 45% of OpenRouter Traffic — and the West Hasn't Noticed
