The AI Economy Is Leaving the Rest of America Behind

The University of Michigan's consumer sentiment index hit a record low on 22 May 2026. Fifty-seven percent of respondents told pollsters their personal finances were being eroded by high prices. The same day, Polymarket traders were pricing a 92 percent probability that an AI data-center moratorium would pass before year-end — not because the technology is failing, but because the physical infrastructure is consuming so much power that communities and regulators are pushing back. These two facts sit inside the same country. They are not in tension with each other. They are the same story, told from different floors of the same building.

What the market is pricing and what the grocery store is charging have diverged. Since the start of 2024, AI-linked equities have outperformed the non-AI S&P 500 by 121 percentage points. That spread is not a temporary dislocation. It reflects a structural reassignment of capital away from the broad economy and into a concentrated cluster of infrastructure, semiconductor, and hyperscaler companies whose customer base is overwhelmingly other businesses rather than households. The Federal Reserve's mandate covers both sides of that ledger. When it hikes to cool inflation, it slows consumer spending. When it cuts to support growth, it lowers the cost of capital for data-center debt. The people holding AI equities are, on balance, insulated from the first effect and beneficiaries of the second. The people filling up their petrol tanks and replacing worn appliances are not.

The IRS disclosure question adds a different dimension. The agency's reported consideration of a citizenship-status question on next year's tax forms is, on its face, a compliance and enforcement matter. But in a moment when AI is projected to automate significant portions of routine cognitive work —tax preparation included — a question about legal status on a government form is also a question about who the automated future is designed to serve. Highly automated filing systems work well for citizens with stable employment histories and W-2 income. They tend to break against gig workers, undocumented residents, and people navigating complex multi-state situations. If the IRS is building its next-generation infrastructure on a definition of normal that fits Silicon Valley's idea of a taxpayer, it will systematically misread a large fraction of the actual population.

Google's AI search stumble — specifically the breakdown triggered by the word "disregard" returning garbled results — is being treated as a product anecdote. It is not. It is a demonstration of what happens when language-model integration is deployed at scale before the underlying systems are understood. The Polymarket markets price only a 23 percent chance that Google holds the best AI model by the end of June. That is a low bar for a company whose name became a verb for internet information-seeking. The "disregard" bug exposes something the price does not: the gap between the rhetorical confidence of AI deployment and the actual reliability of the systems in front of hundreds of millions of users. When a single word can break a search result, the product is not mature. The market knows this — it is assigning a probability rather than a certainty — but the deployment timeline is not waiting for the probability to cross 50 percent.

The data-center moratorium is the most structurally revealing signal. Ninety-two percent odds mean the market is not betting on frictionless expansion. It is betting that the environmental cost of running large language models at hyperscale will become politically untenable before the model's productivity gains fully materialize for ordinary workers. That is a reasonable bet. The power consumption of a single large training run can exceed the annual output of a mid-sized industrial facility. Communities near data-center clusters are already reporting groundwater depletion, noise, and grid strain. This is the physical economy — not the stock market version — raising its hand and asking for a time-out.

The policy lever that might begin bridging these two economies does not yet exist in coherent form. Capital gains treatment favours AI equity holders. R&D tax credits have accelerated hyperscaler construction. Trade restrictions on advanced chips have, if anything, concentrated the investment case for domestic data centres. Meanwhile, retraining programmes for workers displaced by AI are chronically underfunded and administratively slow. The political economy of AI is set up to deliver the investment case efficiently. It is not set up to deliver the transition costs efficiently. Until those costs are treated as first-order policy inputs rather than residual externalities, the divergence between the AI economy and the consumer economy will deepen.

What remains genuinely uncertain is the timeline. AI productivity gains may, at some point, show up in wage data, consumer confidence, and mainstream employment in ways that close the gap. The technology is capable enough that a genuine demand-side surge is not implausible. But the mechanisms that would translate a 121-point equity outperformance into broadly shared economic improvement — wages, pricing power, public infrastructure, accessible retraining — require deliberate political choices that the current policy environment is not obviously oriented to make. The record-low sentiment numbers and the AI outperformance are not contradictory facts. They are the same structural observation, seen from the ground and from the terminal respectively.

The wire services covered the consumer sentiment crash on 22 May as a data release. They covered AI stock performance as a market story. These are being processed as separate items. They are not. What is being priced in the equity market and what is being felt at the kitchen table are diverging along a fault line that is political as much as it is economic — and that divergence will shape the 2026 midterms and the next cycle of AI governance debates in ways the current coverage is not yet reflecting.

This piece ran in the Opinion desk. Monexus led with the University of Michigan sentiment data and the AI equity spread as a single structural argument; the wire services treated them as separate releases.