AI Spending Has Become An Inflation Problem Before A Productivity Gain

The consensus view of artificial intelligence in financial markets is that it is a disinflationary force. Cheaper cognitive labor, more output per worker, falling unit costs. Federal Reserve Chair Kevin Warsh has said so in public, describing AI as "structurally disinflationary" and comparing it to the early internet. As a statement about the 2030s that is a defensible view. As a working assumption for a portfolio in September 2026, it is the most expensive idea in the market.

Here is the claim. The AI build-out has grown large enough to work as a macroeconomic demand shock, and it is now the marginal reason a Federal Reserve rate rise is live at all. Not a complication to the tightening story. A cause of it. Investors holding AI-linked equities as shelter from a hawkish central bank are holding the thing that is making the central bank hawkish.

Friday's jobs report contained both halves of the argument

US nonfarm payrolls rose 162,000 in August 2026, against a consensus near 53,000, with the unemployment rate steady at 4.1 percent. The Bureau of Labor Statistics also revised June up to +31,000 and July up to +21,000, adding 55,000 jobs to the prior two months. Average hourly earnings rose 0.3 percent on the month to 37.75 dollars, up 3.1 percent over the year.

Buried in the same release was a line that gets quoted as evidence for the other side of the argument. Information employment fell 23,000 in August, against an average monthly loss of 8,000 over the prior twelve months, with declines in computing infrastructure and data processing (down 8,000), publishing (down 7,000) and broadcasting and content (down 5,000).

The market did not treat that as disinflationary. The 10-year Treasury yield rose to about 4.79 percent on Friday 4 September, having touched 4.812 percent after the data, and futures moved to roughly a coin flip on a 25 basis point increase at this month's meeting. The S&P 500 fell 0.38 percent to close at 7,718.60, having ended Thursday at 7,747.71. The bond market read a labor report containing the sharpest information-sector job losses of this cycle and concluded that policy was too loose. It was right to.

A build-out this size is a demand shock long before it is a supply improvement

The five largest US cloud and AI infrastructure providers have guided to somewhere between 630 billion and 700 billion dollars of capital expenditure in 2026, against a record 388 billion dollars in 2025. J.P. Morgan puts the 2026 figure near 697 billion dollars. Whatever the precise number, it is close to a doubling in a single year.

That money does not buy intelligence. It buys concrete, transformers, switchgear, turbines, copper, high-voltage cable, industrial electricians and semiconductors, most of it inside the United States and most of it on delivery schedules that cannot flex. Investment in AI-related data center construction, compute hardware and networking equipment came to roughly 0.8 percent of US GDP in the first quarter of 2026. Computer, peripheral and software investment alone contributed about 1.09 percentage points of the 2.0 percent annualized growth reported for that quarter.

An economy running near full employment that adds a demand impulse of that size to a fixed short-run supply of skilled trades and grid equipment does not get lower prices. It gets higher prices, and it gets them first. The productivity gain, if it arrives, arrives on a different and much longer clock.

The inflation is already visible, and it is in the electricity bill

Electricity prices rose 4.2 percent over the year to July 2026, up from 4.0 percent in June, with the average price at 19.7 cents per kilowatt hour. Wages in the utilities sector rose 8.1 percent over the same year, the highest rate on record for that series. Utilities are not a sector that generates 8 percent wage growth on its own. It happens when someone else is bidding for the same linemen.

The physical demand is not disputed. Peak load on the PJM grid reached 162.6 GW on 2 July 2026, roughly 2 GW above the previous record, and PJM power demand is projected to rise by 5,400 MW across 2026 with data centers responsible for the majority of the increase.

Meanwhile core PCE inflation held at 3.3 percent over the year to July 2026, unchanged from June, and headline CPI ran at 3.4 percent. Energy prices were up 14.7 percent over the year, dominated by a 24.6 percent rise in gasoline tied to Middle East supply disruption. The oil shock is not AI's doing. But it means the AI demand impulse is landing on an economy that has no spare disinflationary capacity to absorb it, which is precisely when a central bank has to act on a demand shock rather than look through it.

The disinflation people cite is the wrong kind of disinflation

Census Bureau data show 39.7 percent of information-sector firms now using AI against a 19.8 percent average across all industries, and it is in that sector that jobs are disappearing fastest. This is offered as proof that AI is bringing prices down. It proves something narrower.

Job losses concentrated in one sector are not disinflation. Disinflation is a fall in the general price level, and it requires that the output per worker gain shows up in aggregate measured productivity and then in unit labor costs. None of that has happened yet. Average hourly earnings are still growing at 3.1 percent, which with trend productivity anywhere near 1.5 percent is not consistent with 2 percent inflation.

What we can currently observe is displacement without a measured productivity offset, funded by a capital expenditure boom that is bidding up energy and construction inputs. That combination is not a positive supply shock. In miniature, it is the opposite: fewer jobs and firmer prices at the same time. Chair Warsh's disinflation is a forecast. Kashkari's demand element, as he put it, is in the data now.

The utilities sector is the market quietly admitting this

The cleanest evidence sits in a corner of the equity market that most people file under something else entirely. Utilities should be the purest structural winner from AI. Load growth is real, rate base expansion is real, and most companies in the sector are guiding to 6 to 8 percent earnings growth or better.

Utilities in the S&P 500 rose more than 11 percent through the end of February 2026. They are now close to flat for the year and rank second worst of the eleven sectors in the index. The volume growth arrived exactly as advertised, and rising Treasury yields ate it.

Investors usually discuss two separate things about utilities: that they are an AI beneficiary, and that they are rate-sensitive. Those are not two stories. The AI demand that lifts utility volumes is a meaningful part of what is lifting the discount rate applied to them. The sector has spent 2026 running a controlled experiment on whether the AI boom is inflationary, and the answer it returned is yes.

What this means for how the AI trade is actually held

Most portfolios in 2026 hold AI exposure as a growth position assumed to be indifferent to policy, on the reasoning that earnings growth of this magnitude overwhelms the discount rate. That reasoning has a hole in it. A capital expenditure program approaching 700 billion dollars a year is increasingly financed rather than funded from cash flow, which makes the build-out itself a rate-sensitive asset. And it is the build-out, not the software, that is currently doing the tightening.

The practical point is about correlation, not about any individual security. An investor who owns index-level US equity exposure, owns utilities for defensiveness, and owns duration as a hedge is, on this reading, holding three expressions of the same underlying bet on the AI capital cycle. That is a concentrated position wearing the costume of a diversified one.

What would change my mind

This is a falsifiable claim, so here is what would break it. If measured nonfarm business productivity growth runs sustainably above roughly 2.5 percent while unit labor costs decelerate, the productivity channel is real and arriving faster than I think. If core services inflation excluding shelter falls toward 2.5 percent while capital expenditure guidance keeps rising, then the build-out is being absorbed without price pressure and Warsh is right. If electricity CPI inflation falls back below 2 percent while data center load keeps climbing, the grid constraint I am leaning on is not binding.

There is one other way to be wrong, and it is worth naming because it is not comforting. If hyperscaler capital expenditure guidance is cut sharply, the demand impulse reverses and inflation falls quickly. The thesis that AI capex is what is holding up prices would be vindicated by the disinflation that followed. It would also be a very poor environment in which to own equities. Being right about the mechanism and wrong about the direction of the shock is the specific risk in this trade.

For now, the position is this: the Federal Reserve should raise rates this month, the reason it needs to is sitting in a grid interconnection queue rather than in a wage-price spiral, and the disinflation everyone is waiting for is a story about the next decade being used to price this one.