AI's Inflationary
Footprint Arrives First

The letter reached Liberty Utilities customers in March 2026 with something between a warning and a shrug: in 14 months, the utility supplying power to 49,000 residents on California's side of Lake Tahoe would have to find a new source for three-quarters of its electricity. NV Energy, the Nevada-based supplier, said it could not continue the arrangement beyond May 2027 because of its "own resource needs."

What NV Energy called a "long-standing" transition plan looked different from inside the data center industry. The previous September, Jeff Brigger, NV Energy's director of business development, had told a data-center-focused conference that tech companies were driving requests to triple the utility's peak demand — which runs around 9,000 megawatts during heat waves. "These are unprecedented times," Brigger said, according to the Las Vegas Review-Journal.

Northern Nevada has become one of the fastest-growing data center corridors in the country. Google, Apple, and Microsoft have built or are planning facilities around the Tahoe-Reno Industrial Center east of Reno, according to reporting by Fortune. A January report from the Desert Research Institute, part of the University of Nevada, counted more than 40 projects underway. NV Energy's own 2024 resource plan lists 12 of them, projecting 5,900 megawatts of new demand — nearly three times the Hoover Dam's capacity — and load growth by 2033 that would exceed half of Nevada's entire 2024 net electricity generation.

NV Energy disputed any link to Lake Tahoe. "Data centers did not influence this decision," the company said. "The decision for Liberty to move to its own power supply is based on long-standing agreements and planning assumptions that date back more than a decade — well before data center growth became a factor." But a California Public Utilities Commission spokesperson told CapRadio, Sacramento's public radio station, that according to Liberty, NV Energy's decision was "partially due to projected demand by data centers in the region."

Rates in the area were already climbing before the supply problem went public. Danielle Hughes, a North Tahoe resident and consumer advocate who also works in California's Energy Commission efficiency division, said she is one of roughly 17,000 to 20,000 year-round residents bearing costs that state energy models don't capture. The rate structure hits permanent residents with demand charges that vacation homeowners don't face. "Our rates are going to be the highest in the country most likely, and we are not being considered," she told CapRadio.

Cody Bass, South Lake Tahoe's mayor, said he learned about the supply disruption when residents did. "We really didn't get any notice prior to the public press release," Bass told CapRadio. He has since met with Liberty and state regulators, who assured him the power won't simply cut out. That's not what worries him. "I think we're pretty aware it's going to cause higher rates, and that of course becomes a major issue for our residents, businesses and for everybody else because our rates are already pretty high."

Lake Tahoe is a preview of a tension running through the broader U.S. economy. The infrastructure costs of the AI buildout — power, grid upgrades, data center construction — are landing in electricity bills now, while the technology's promised productivity gains remain years away. That timing gap sits at the center of one of the more consequential debates in monetary policy.

Kevin Warsh, the newly confirmed Federal Reserve chairman, has argued publicly that artificial intelligence will reduce costs across the economy, lift productivity, and create room for lower interest rates without reigniting inflation. The argument has gained traction in market commentary and Fed-watcher debate — particularly as Warsh navigates political pressure to cut rates ahead of November's midterm elections.

The thesis rests on a straightforward supply-side logic: AI raises output per unit of labor and capital, reducing unit costs across adopting sectors. If the productivity dividend is large enough and diffuses quickly enough, the aggregate price level falls — or at minimum inflation decelerates — even as the economy grows. Under that scenario, the Fed can cut rates without sacrificing its inflation mandate.

The problem is timing. The 1990s productivity boom took years to diffuse from early adopters to the broader economy. AI adoption today is at an earlier stage. More than three in four U.S. businesses have yet to incorporate AI, according to the Census Bureau. And unlike the 1990s, the current cycle requires building a new physical infrastructure — power grids, data centers, semiconductor fabs — that is itself inflationary before it produces any disinflationary output.

The Cato Institute has argued that while Warsh may be right about the need to overhaul the Fed's governance, his inflation argument is a trap: turning a forward-looking productivity thesis into a confident monetary-policy framework assumes both rapid diffusion and a clean pass-through from productivity gains to lower prices — two assumptions the evidence does not currently support.

The most direct challenge to the Warsh thesis comes not from theoretical models but from the price data itself. A quarterly U.S. AI-intensity index, published in April 2026 by Professor Salem Abo-Zaid at the University of Maryland, finds that AI has already made a positive and rising contribution to U.S. inflation — with the most AI-exposed sectors tending to show the highest inflation rates (SSRN working paper 6529799). This is the opposite of what the Warsh thesis predicts. The buildout costs are arriving before the productivity gains that are supposed to offset them.

U.S. electricity inflation ran at 6.9% year-on-year through December 2025 — more than double the headline PCE gauge — according to Bloomberg data cited by Nuveen in its June 2026 monthly commentary. Consumer electricity prices are likely to stay near 6% through 2027, while data center demand is on course to nearly double from its current roughly 4% share of total U.S. electricity by 2030, per Nuveen estimates. The chart below shows the divergence between electricity costs and the broader price level from 2023 to 2025.

"The clearest empirical signal so far cuts against the consensus that AI is disinflationary," wrote Laura Cooper, Nuveen's head of macro credit, and Quinn Brody, a senior macro strategist, in the June commentary. "The steep costs of the AI buildout are arriving faster than the productivity gains that are supposed to offset them."

The broader academic literature agrees that the inflation picture is fundamentally ambiguous. A Bank for International Settlements working paper — one of the most rigorous central bank examinations of AI's macroeconomic effects — finds that AI raises output, consumption, and investment, but that the inflation response depends critically on expectations. If firms and households anticipate future productivity gains, they may consume and invest today against tomorrow's expected income, pulling inflation forward before the gains have materialized.

IMF research reaches a similar conclusion. Model simulations in a 2025 working paper show that AI can raise global productivity and GDP materially over a decade, but near-term inflation may edge higher because investment and demand outpace the initial supply response. In those simulations, central banks respond with modest tightening in the short run — not rate cuts.

A 2026 SUERF policy note captures the two-sided transmission precisely: AI can be disinflationary by lifting supply and reducing unit labor costs, but inflationary through stronger investment demand, higher expected income, and labor-market frictions during adoption. The note frames AI as "conditionally disinflationary" rather than inherently deflationary — a distinction that matters considerably for anyone mapping it onto a rate-cut argument.

The clearest sector-level evidence for AI as a disinflationary force comes from European manufacturing. A 2025 paper in Economics Letters by Borowski, Fidrmuc, and Jaworski finds that a 10-percentage-point increase in EU firms using AI is associated with roughly a 0.3–0.6 percentage-point decline in producer-price inflation, with effects concentrated in services and visible only beyond a threshold of adoption.

This is a real effect — and the most direct empirical support Warsh's thesis can claim. But producer-price effects in adopting sectors don't automatically translate into broad consumer-price disinflation. Aggregate inflation also depends on wages, non-AI sectors, energy and housing costs, and whether AI-related capital spending lifts demand elsewhere.

Economists surveyed by the Financial Times largely rejected the idea that AI would meaningfully reduce inflation or policy rates over the next two years, with many expecting the effect to be negligible. A June 2026 World Economic Forum survey found economists now expect AI-driven productivity gains to take at least another two years to materialize across most sectors. (WEF Chief Economists Outlook, May 2026)

The bond market is sending a mixed signal. When NBER researchers studied U.S. Treasury, TIPS, and corporate yields around major AI model releases in 2023 and 2024, they found yields fell consistently by more than 10 basis points on average, remaining lower for over two weeks. (Andrews & Farboodi, NBER WP 34243, 2025) On the surface, bond markets priced AI as a disinflationary shock. But the NBER paper attributes that reaction partly to downward revisions in expected consumption growth — a considerably less bullish read. Meanwhile, the financing of the AI buildout is generating structural upward pressure on term premiums through long-duration corporate bond issuance, a dynamic the Dallas Federal Reserve flagged earlier this year as structural rather than cyclical. (Searls, Dallas Fed, Feb 2026)

Beneath all of this sits the neutral interest rate — the rate at which monetary policy is neither stimulating nor restraining the economy. A productivity-driven AI economy implies a higher neutral rate, since stronger growth prospects raise the equilibrium cost of capital. But if AI-driven gains accrue disproportionately to capital owners while displacing portions of the workforce, higher precautionary savings and greater social uncertainty could offset some of that upward pressure.

The Cleveland Fed's current estimate of the nominal neutral rate spans a confidence band from 2.9% to 4.5% (Zaman, Cleveland Fed, Sep 2025) — a range so wide it renders confident policy prescriptions difficult to defend. The Warsh case for lower rates implicitly assumes the actual neutral rate is near the lower end of that band; the inflationary buildout dynamics outlined in this report suggest it may be closer to the upper end.

Since 2024, 64 cents of every dollar of U.S. GDP growth has been attributable to tech spending, and hardware investment has risen from nearly 2% of GDP to more than 3%, per Bureau of Economic Analysis data. Yet more than three in four U.S. businesses have yet to incorporate AI, per the Census Bureau. The productivity growth already running above 2% this cycle largely reflects post-pandemic labor-market dynamics rather than any AI contribution.

The employment picture adds another layer of uncertainty. The IMF estimates that close to 40% of global employment is exposed to AI, and roughly four in five U.S. workers have at least 10% of their tasks exposed to large language model capabilities (Eloundou, Manning, Mishkin & Rock, arXiv 2303.10130, 2023).

Yet since the introduction of generative AI, economy-wide job losses and wage declines have not materialized. More than 142,000 U.S. tech workers were laid off year-to-date as of mid-2026 — about a 33% increase over the same period in 2025 — but analysts put only roughly a quarter of those cuts down to AI and automation, with the rest reflecting cost discipline and the unwinding of pandemic-era over-hiring, according to TrueUp and Challenger, Gray & Christmas.

The labor market picture is one of uncertainty without the wage deflation that would most directly support the Warsh thesis. If AI were already generating broad productivity gains and reducing labor input, we would expect to see declining unit labor costs and softening wage growth. Neither is clearly present at the aggregate level.