Deep Dive · No. 027 · May 23, 2026

AI CapexAI LaborMacro

AI Has Three Inflations

We now also have two economies, but still only one CPI number.

AI is going to be inflationary in the short run, inside a narrow set of datacenter-adjacent inputs. It will be disinflationary in the medium run as job-displacement fears curb spending. However, in the long run, it will be massively deflationary as productivity gains flow through to prices.

That is my thesis.

Most of the inflation debate I have read collapses these three horizons into one, takes a directional view on whether AI is, on net, inflationary or deflationary, and trades that view. That is the wrong question. The three are different forces, with different mechanisms, hitting at different times. They require different policy responses.

The harder question is what the Federal Reserve does in the meantime, when CPI prints (and unemployment numbers) will not show the Fed the split between the three.

The Short Run: Inflation Inside the Buildout

Anything associated with the AI buildout is going up in price. Anything outside it, mostly, is not.

The buildout itself is enormous. Combined AI capex from Microsoft, Alphabet, Meta, and Amazon is tracking toward roughly $700 billion in 2026, nearly double last year. Bridgewater estimates this cycle alone adds about 140 basis points to US growth in 2026. The macro footprint is large. The inflation footprint is concentrated.

The clearest signal is power. PJM, the largest US grid operator, saw wholesale prices rise 76% year-over-year in Q1 2026, with capacity prices up 398%. The same forces are pushing through every other input that touches the buildout: copper, specialty land, electrical labor, AI engineer comp. Specialist electricians on hyperscaler sites in Texas are reportedly earning $240,000 to $280,000 a year.

This is real inflation. It is also a sector story. The April CPI print landed at 3.8% headline, but the overshoot is shelter, transportation services, and energy. Core goods were flat. AI capex feeds the energy channel. It is not most of the print.

If you are looking for AI-driven inflation in macro aggregates, you will find it in categories that touch the buildout. Beyond those, you will not find it.

The Middle: Spending Cools as the Rest of the Economy Bends

In the medium run, the same forces driving short-run inflation do not extend outside the buildout. The dominant macro variable shifts. It becomes consumer sentiment, and consumer sentiment is starting to break.

The May 2026 preliminary University of Michigan reading printed at 48.2, comparable to the June 2022 inflation panic. Sixty-eight percent of consumers expect unemployment to rise in the next year. Goldman Sachs estimates AI has already reduced monthly US payroll growth by roughly 16,000 jobs over the past year. Amazon eliminated approximately 30,000 corporate roles in the year after Andy Jassy's June 2025 memo warning AI would shrink the corporate workforce.

When people are worried about losing their jobs, they spend less. This is the medium-term force this thesis turns on. The cohorts most exposed to AI substitution (entry-level white-collar, routine knowledge work, junior analysts, customer support) are also the marginal spenders in the discretionary economy. When that cohort tightens, the pullback concentrates in spending categories the Fed's tightening tools cannot reach.

The catch is that the pullback has not yet shown up in the aggregate saving rate, which fell from 4.5% in January to 3.6% in March. That looks like confident spending. It is not. The top 10% of US households own roughly 93% of all equities and have been spending freely against Magnificent 7 gains. The bottom 90% has nothing left to set aside. The aggregate saving rate is one number averaging two different consumer behaviors.

My read: the disinflation is on the way. The aggregate data is masking it.

The Long Run: The AI Promised Land

Eventually the buildout finishes, the models diffuse, and AI behaves like every other transformative technology has behaved. Production gets more efficient. Goods and services get cheaper. The cost savings flow through.

AI lifts productivity and GDP by 1.5% by 2035, 3% by 2055, and 3.7% by 2075. Those are large numbers compounded over time. The mechanism is the standard one. If a process becomes cheaper to run, competitive pressure passes the savings to the buyer, and prices fall.

The historical analog is hard to argue with. Computers used to be multimillion-dollar investments that filled rooms. Cell phones used to cost thousands and only made calls. Today you carry both in your pocket for $500, and most of what they do, they do for free. That is what technology proliferation does. It compresses prices for the things it touches and expands what each dollar can buy. The deflationary endgame is not controversial. It is just years away.

What the Fed Sees, and What It Doesn't

Aggregate CPI is what the Fed sees. It is one number that averages the short-run input inflation, the medium-run disinflation, and the long-run deflation into a single print, with the short-run weighted heavily because it is the loudest.

At a funding desk, the difference between what the inflation indicator says and what the underlying flows are doing is the trade. You watch the same CPI print the Fed watches, and you watch the funding-market and consumer-credit flows the Fed reacts to with a six-month lag. When those diverge, the Fed makes mistakes. The 2022 hiking cycle into a softening labor market was one. The 2007 "inflation is the bigger risk" moment is the one nobody wants to repeat.

If the Fed reads aggregate CPI as inflationary and tightens, the short-run inflation does not heal. You cannot print copper. You cannot print electricians. You cannot print 1-gigawatt substations. What rates do touch is the rest of the economy: mortgage affordability, small-business credit, consumer credit, the labor market for everyone outside the AI buildout. Tightening into the short-run inflation breaks the part of the economy the medium-term disinflation is going to hit hardest anyway, without disciplining the short-run inflation itself.

The risk of a policy mistake here runs in only one direction, and that direction is tightening.

What to Watch

Three indicators, one per horizon.

Short run. PJM and ERCOT wholesale power prices, Comex copper, hyperscaler capex guidance in Q2 earnings. The inflationary side. Sector data, not macro inflation.

Medium run. Monthly job-cut announcements, UMich expected unemployment, JOLTS quits rate, and the recent-graduate unemployment gap. The disinflationary side, emerging in real time.

The bridge. The personal saving rate. The signal is not its level. The signal is whether the rate stops falling.

From a macro perspective, the two most important things to watch out for are always inflation and unemployment. AI will disrupt both, and in different ways across different time horizons.