Fed's Williams on AI Productivity: Why the Boom Means Higher Rates, Not Cuts
New York Fed president John Williams says he is not yet seeing AI in the productivity data. That takes apart the argument Kevin Warsh brought to the Fed chair. Even if the boom arrives, it pushes the neutral rate up, not down. The desk expects a second hike on October 28.
John Williams said on Friday that he is not seeing an impact on productivity from artificial intelligence yet. For a New York Fed president, that was a small remark. It matters because it undercuts the idea that got the current Fed chair his job.
Kevin Warsh arrived arguing that AI would be "a significant disinflationary force", meaning the economy could grow faster without prices rising and the Fed could cut. Nine days ago his committee voted 12-0 to raise rates by a quarter point, to 3.75-4%, the first increase since July 2023. Sixteen of the 18 officials who submitted projections expect another move up this year. Williams, who votes every year and sits near the committee's centre, said on Thursday that the market's expectation of one more hike by year-end was "a reasonable way of thinking about it". Bloomberg reported the remarks.
The two comments go together. The Fed now treats AI as a demand shock. Williams has said since July that AI spending is his main inflation worry. Data-centre building has bid up the prices of semiconductors, power transformers and grid equipment faster than new supply can come on stream. Add an energy shock and the result is 3.4% headline consumer-price inflation in August, with energy prices up 16.3% on the year, according to the BLS. Inflation has been above the Fed's target for five straight years. Nobody on the committee wants to look past a sixth on the promise of a productivity dividend.
The case for patience rests on real numbers. Output per hour in the non-farm business sector grew 2.2% in the year to the second quarter, the BLS says, well above the roughly 1.5% trend of the decade before the pandemic. Williams himself told a Reykjavik audience in May that AI could plausibly produce something like the 1996-2005 boom, when productivity grew about 3% a year. On this view the Fed is making Alan Greenspan's famous mistake in reverse. In 1996 Greenspan refused to tighten because he suspected productivity was running ahead of the official statistics, and history proved him right.
The Greenspan comparison works against the doves for two reasons.
The first is what the investment does in the meantime. In the late 1990s the capital spending that raised productivity was mostly cheap: PCs, networking gear and software whose prices were falling fast. It pushed inflation down as it spread. The current build-out is heavy. It runs on power, copper, turbines and chips that are in short supply. Before a single model makes an accountant faster, the capex competes for the same scarce electricity and equipment as the rest of the economy. Greenspan's bet was that supply was already expanding. Today's bet would be that supply will expand later, while demand is bidding up prices now.
The second reason matters more. A genuine productivity boom raises the neutral rate of interest. Williams made the point himself in his May speech: the productivity slowdown of the 2000s was "a significant influence on driving down the natural rate of interest." A sustained AI boom would lift it. Faster productivity growth raises the return on capital, pulls in more investment and lets households borrow against higher expected incomes. The rate that neither stimulates nor restrains the economy goes up. Williams's own estimate of the real neutral rate is about 1.1%. A 1990s-style boom would push that estimate higher, and policy would have to follow.
The Warsh thesis loses both ways. If AI productivity is not arriving, the Fed is facing an investment-driven demand surge with inflation above target, and it hikes. If AI productivity is arriving, the neutral rate is rising, and today's policy rate is less restrictive than it looks. That argues for holding higher for longer, not cutting. AI justifies easier money only if productivity surges and prices fall while investment and consumption stay calm. That describes 1998 more than 2026.
The bond market has started to price this. Williams has argued that the rise in long-term Treasury yields reflects the economy's strength and the scale of AI and data-centre investment more than fear about inflation. That sounds soothing. For equity investors, higher real yields driven by investment demand are exactly what a rising neutral rate looks like from the trading desk, and they fall hardest on long-duration assets. The same AI franchises whose spending is causing the problem are the most exposed.
There is also a financing loop. Hyperscalers are increasingly paying for the build-out with bonds and leases rather than free cash flow. Each Fed hike raises the cost of the capex that Warsh hoped would bring rates down. The Fed is not trying to stop the AI boom. It is making the boom pay market rates for the resources it uses, an uncomfortable result for anyone modelling data-centre returns on 2024 funding costs.
The measurement problem makes all of this slower. Williams's May speech noted that quarterly productivity growth swings between minus 2% and 7%, and that the 1970s slowdown was not properly recognised until 1979. The Fed will not credit AI with a structural gain on one or two good quarters. It will want evidence that gains are spreading beyond a handful of industries, and research from the Kansas City Fed has found the recent pickup is not yet broad-based.
The desk expects the committee to raise rates again on October 28, to 4-4.25%. We expect Warsh to keep AI in his press conference as a long-term hope rather than a reason for patience. The next real test is the preliminary third-quarter productivity report in early November. A strong print will be read at the Fed as evidence for a higher neutral rate. Investors hoping it opens the door to cuts will be disappointed.