The AI bubble thesis rests on a timing error. Its proponents observe that hundreds of billions in capital expenditure have not yet produced commensurate GDP growth and conclude the investment is irrational. What they are actually observing is a well-documented feature of transformative technology adoption: the productivity gains arrive unevenly, accumulate at the firm and task level first, and surface in aggregate statistics years after the underlying diffusion has begun. The critics are not wrong that aggregate proof is thin. They are wrong to treat that thinness as dispositive.
The micro-level evidence is unambiguous and growing. Controlled studies across customer support, software development, writing, and professional consulting consistently show productivity gains ranging from 5% to over 25% on exposed tasks. These are not vendor case studies. They are peer-reviewed experimental designs — Noy and Zhang, Brynjolfsson, Li and Raymond, Peng et al. — run across different firms, occupations, and national contexts, all arriving at similar magnitudes. The gains are strongest among initially lower-performing workers, which means AI is compressing the skill distribution within organizations rather than simply amplifying elite output. That pattern has systemic implications that payroll data and GDP deflators cannot yet capture.
The firm-level data is catching up. A March 2026 NBER working paper drawing on nearly 750 corporate executive surveys found positive labor productivity gains across all measured sectors in 2025, with the largest effects in high-skill services and finance at roughly 0.8%. Those same executives project the gains will roughly double by the end of 2026. Critically, the gains are driven primarily by innovation and demand-oriented channels — new product development, expanded customer reach — rather than simple headcount reduction. That is the harder-to-measure, more durable form of productivity growth.
The aggregate picture is also beginning to shift. U.S. labor productivity growth ran at approximately 2.7% in 2025, nearly double the decade-long average of 1.4%. Morgan Stanley’s industry-level analysis attributes roughly 1.7 percentage points of that growth to top-quartile AI-exposed sectors: data processing, computer system design, software publishing, and computer manufacturing. That is not a rounding error. And the productivity J-curve framework — the empirical observation that transformative technologies depress measured output during the investment phase before the efficiency gains materialize — predicts exactly this sequence: widespread adoption, muted aggregate signal, then acceleration. Electricity and the internet both followed this arc. The lag between investment and measurable GDP contribution for general-purpose technologies has historically been measured in years, not quarters.
The Goldman Sachs critique — that $700 billion in AI investment in 2025 contributed essentially zero to GDP growth — is a data point, not a verdict. It measures one year of one indicator during the capital accumulation phase of a multi-decade technology cycle. The same institution simultaneously acknowledges isolated task-level gains of approximately 30% in firms actively measuring their AI deployments, and declines to classify current market valuations as bubble-level. The skepticism and the evidence are in tension, and the skeptics have not resolved it.
The bubble framing also smuggles in a false equivalence with the dot-com collapse. The 1990s infrastructure overbuild destroyed value because it was built ahead of any demonstrated use case and financed by companies with no viable revenue model. The current AI build-out is being financed primarily by hyperscalers — Microsoft, Google, Meta, Amazon — that are simultaneously the largest-scale deployers of the technology they are building infrastructure for. They are not speculating on future demand. They are consuming their own capacity. That is not the anatomy of a bubble. It is vertical integration under conditions of genuine demand uncertainty, which is a different and considerably less catastrophic failure mode.
None of this is to argue that current valuations are conservative or that capex deployment is optimally calibrated. It is to argue that the productivity gains from AI are real, measurable, and accelerating — and that a narrative constructed from GDP lag, timing mismatch, and the psychological residue of 2000 is not a substitute for evidence. The evidence runs the other direction.
