AI investment has become large enough to rewrite the Federal Reserve’s reaction function. The mechanism runs through the neutral rate and how the committee classifies inflation, not through the headline capex figure.
Net of imports, AI investment adds roughly 0.5–0.8 percentage points to U.S. growth, worth 0.1–0.25 points of core inflation through the output gap. That justifies 40–80 basis points of extra tightening relative to a rule that looks through supply shocks. However, the funds rate barely touches AI capex. Spreads and term premia do.
The Nominal-to-Domestic-Value-Added Decomposition
The headline capex figure overstates U.S. domestic value-added in two ways.
First, part of the increase is price, not volume. Microsoft attributed about $25 billion of its 2026 guide to higher component pricing, and Meta raised guidance citing higher component and data center costs. Extrapolating, 10–15% of the 2026 level is price, a quarter to a third of the nominal increase.
Second, and larger, imports are the leak. Fed staff estimates show AI-related investment contributing 0.14 points to GDP in late 2025 against about 0.75 points gross, with net imports of AI equipment subtracting 0.61 points.
| Measure | Gross | Net of imports | Source |
|---|---|---|---|
| AI investment contribution to GDP | 0.75pp | 0.14pp | Fed staff, late 2025 |
| Net imports of AI equipment | — | -0.61pp | Fed staff, late 2025 |
| Bhide gross-to-net | 0.9pp | 0.4–0.5pp | Bhide |
| ING gross-to-net | ~50% of Q2 y/y growth | 36% after netting computers/semis | ING |
| Wells Fargo | ~quarter of growth | — | Wells Fargo |
Across estimates, domestic value added is 30–70% of the gross figure, central 50%. Per $100 of nominal capex, about $12 is price and about $88 is volume. Half of the volume is U.S. value added, so about $44 becomes U.S. output.
Applied to the +$320 billion nominal increase in 2026, that is about $140 billion, or 0.45% of GDP. Adding induced consumption from construction wages and equity wealth, with a multiplier of about 1.5, gives roughly 0.7 points of growth.
That sits inside the 0.14–0.73 point range of Fed-staff quarterly readings.
The Transmission Loop, Link by Link
| Link | Size | Feedback to the policy rate |
|---|---|---|
| Imported equipment | Net imports subtract 0.45–0.61 points of growth in Fed-staff readings | Demand leaks to Asia and Mexico. U.S. output gains are smaller, and component prices rise |
| Domestic construction | Data center construction was a $75 billion annual rate in July, about 20% of all-in cost. In August it rose 7.5%, up 73% y/y to a record $84.95 billion | About 0.25% of GDP directly. Rate-sensitive construction is already contracting. The price spillover is in trades and electrical components |
| Electricity | Data centers use 240–280 TWh, about 5.5–6.5% of U.S. electricity in 2026. EIA sees residential prices rising from 17.30 to 18.29 cents/kWh | Residential prices rise about 5.7%. With electricity at about 2.5% of CPI, the data-center share is under 0.1 point. This link is a small stabilizer |
| Corporate borrowing | Hyperscaler debt issuance passed $100 billion in 2025, up from a $35 billion annual average in 2020–2024. AI-related corporate borrowing ecosystem-wide approaches $450–570 billion. Tech/AI debt in IG indices expanded from under 9% to over 13% | The market has already added about 37bp of spread, more than a 25bp hike. The binding constraint is shrinking investor appetite, not the policy rate |
| Policy rate to capex | Goolsbee says AI investment has not been very sensitive to rates but is becoming more debt-financed. Waller says buyers with perfectly inelastic demand will pay whatever price | A 25bp hike passed fully into the debt share moves blended financing cost by about 8bp. With a user-cost elasticity of 0.5, capex falls about 0.3%, roughly $2 billion against a $320 billion annual increase |
The last row is the key result. Hikes land on housing, autos, and non-AI business investment, not on the demand source. Waller notes that 10- and 30-year yields above 5% are doing some of the work.
That is the same compositional squeeze, delivered through term premia.
The Reaction Function Translation
The parameters, all assumed:
- Pre-AI growth was about 1.5%, against potential of about 1.8–2.0%.
- AI lifts growth by 0.7 points, which closes the slack and leaves an excess gap of 0.2–0.4 points per year.
- The Phillips-curve slope is 0.1–0.2.
- One hundred basis points of tightening lowers output by about 0.5%.
On these numbers:
- AI explains 0.1–0.25 points of a core PCE overshoot that has been above 3% every month of 2026. Oil and tariffs explain most of the rest.
- The excess gap is worth 40–80bp of policy rate. The Fed has delivered 25bp, to 3.75%–4%, and projects one more this year and none in 2027. The path sits inside the AI-justified range.
- The classification matters more than the size. Goolsbee argues that rates may need to be higher even if the tariff and Iran-war supply shocks are one-time. A supply shock gets looked through, but a persistent demand shock does not. Waller puts the duration at two to three years. Expectations raise the cost of waiting: the NY Fed one-year measure is 3.9%, the highest in three years.
- The neutral rate shifts too. Cook raised the possibility that the neutral rate is higher than before the pandemic because investment is supporting strong demand.
- The committee has not converged on the demand diagnosis. Cook says the boom looks like a demand shift toward AI sectors rather than economy-wide demand, and that hiking against it could be a mistake. The September vote was still unanimous.
The Correction Scenario and Timing
The IMF’s downside scenario has a steep AI investment drop cutting global GDP by 1.1% by 2028, including 1.8% in the U.S. The reported staging is a direct hit of about 0.4% of global GDP, 0.6% with wealth effects and tighter financial conditions, then bond-market and demand spillovers. Amplification of roughly 2.75x comes from financial channels, not the investment decline itself.
On our 0.5%-per-100bp parameter, a 1.8% U.S. level hit needs about 360bp of easing. The funds rate now has about 375–400bp of room, and long yields and oil limit how much of that is usable.
A correction is not required for the demand impulse to fade. The contribution tracks the growth rate of capex, not its level. If capex growth slows from about +78% to about +20%, the growth contribution falls by roughly 0.5 points with no decline in spending.
Goldman expects hyperscaler debt issuance to reach a record $420 billion in 2027, up 60%, so capex is still rising next year, but the fade will begin. Hikes made now operate with a four-to-six-quarter lag, so they peak in effect in 2027–28, when the AI tailwind is shrinking.
Leading Indicators
- Bond cover ratios and AI-issuer spreads. A cover ratio at or below 2x with spreads over 150bp is where the rate elasticity of capex jumps.
- Census data center construction against Fed-staff net-of-import contributions. Divergence means import timing is distorting the read.
- Memory and DRAM export prices, for the price component of capex.
- Services inflation and wage growth in electrical and construction trades. This is Goolsbee’s test of whether demand is “jumping out of its lane.”
The Strategic Consequence
The Fed is tightening into a demand engine it cannot directly throttle. The funds rate is the wrong instrument for AI capex; the real transmission runs through spreads, term premia, and the classification of the inflation source.
If the committee continues to treat AI investment as a persistent demand shock, the neutral rate repricing is already underway, and the policy path will stay higher for longer than the pre-AI baseline.
The falsification variable is the bond cover ratio: if it holds above 2x with spreads under 150bp, the rate elasticity of capex stays low, and the Fed’s tightening has less bite on the demand source than the reaction function assumes.