A landmark Federal Reserve survey finds 68% of firms reporting negligible AI productivity gains, reigniting Robert Solow's famous 1987 observation that transformative technology rarely shows up where you expect it to.
Three years and hundreds of billions of dollars into the generative AI boom, corporate America is quietly admitting something the market did not want to hear: the returns are not showing up. The Federal Reserve's Senior Loan Officer Opinion Survey, released April 15, triggered the admission at scale. Across more than 2,000 surveyed firms, the aggregate expected impact of AI on employment turned net-negative for the first time since the AI investment cycle began in 2023. Not negative in the sense of mass layoffs, but negative in the more damaging sense: the technology simply has not moved the needle in any measurable direction.
The specific number doing the most damage is 68%. That is the share of respondents reporting productivity gains of zero to five percent from AI tool deployment, a range economists note is statistically indistinguishable from the lift you get from a routine software upgrade. CEOs across manufacturing, financial services, and professional sectors echoed the same story in Q1 2026 earnings calls: the tools are present, staff are using them, and the efficiency curves are flat.
For anyone who lived through the 1980s technology debates, the current moment has an uncomfortable familiarity. In 1987, economist Robert Solow made the observation that would eventually carry his name: you can see the computer age everywhere but in the productivity statistics. The Solow Paradox became a punchline when the 1990s eventually delivered a genuine productivity surge, vindicating the patient believers in a so-called J-curve effect. Now the paradox is back in circulation, and this time economists at Goldman Sachs and JPMorgan are less certain the vindication follows on the same timeline. Both institutions have revised their 2026 US productivity growth forecasts down from 2.5% to 1.8% annually, a cut that reflects a structural reassessment rather than a cyclical dip.
The critical difference from the 1990s PC boom is the nature of the technology itself. Personal computers reorganized how work was physically done, enabling new workflows that had no prior analog. Current large language models, by contrast, appear to be operating primarily as sophisticated assistance layers on top of existing workflows. They speed up drafting, summarizing, and searching, but they have not yet restructured the underlying economic activity around them. That distinction matters enormously when justifying capital expenditure at the scale the tech sector has demanded.
The Repricing Problem
The AI trade that drove Nvidia and Microsoft to record valuations in late 2025 was built on a specific thesis: that generative AI would function as an industrial engine, compressing labor costs and accelerating output across every sector of the economy. The SLOOS data and the AEA papers released in early 2026 undercut that thesis at its foundation. If productivity gains are negligible and workforce displacement has not materialized, the return on invested capital for the hyperscalers becomes very difficult to defend at current multiples.
This is not an abstract concern. The valuation models underpinning the largest AI-exposed equities are sensitive to assumptions about efficiency gains flowing through to corporate margins over a five-to-ten year horizon. Shave those assumptions and you shave the price targets. The market volatility that followed the SLOOS release reflects investors running exactly that calculation in real time.
Worth noting is what the employment data is not saying. The hollowing-out narrative, the fear that LLMs would structurally eliminate white-collar roles at speed, has not arrived either. That is arguably reassuring for workers, but it reinforces the case that current AI operates closer to consumer productivity software than to the kind of industrial automation that genuinely reshapes labor markets. You do not price Nvidia like a picks-and-shovels play for a revolution that turns out to be a productivity plateau.
What to watch next is whether the Q-curve the optimists are betting on arrives before corporate patience expires. Enterprise AI contracts are typically multi-year commitments, and CFOs who signed in 2023 and 2024 are now approaching renewal decisions with flat productivity data in hand. If the next two quarters of earnings calls produce the same pattern of muted gains, the conversation will shift from when AI delivers to whether current LLM architectures are the right vehicle for the industrial transformation the market priced in.
Also read: Vercel Breach Exposes AI Tool Supply Chain Risk Ahead of IPO • AI Startups Face 12-Month Reality Check as Foundation Models Expand • NSA Runs Anthropic's Unreleased 'Mythos' AI Despite Federal Ban
