AI's net impact on services inflation

Empirical synthesis of capital spending, productivity, and labor channels through 2032.

Headline numbers

Basis points per year, services CPI. Central case combines probability-weighted capex trajectory with moderate productivity offset.

2026 U.S. AI capex
~$950B–1.05T
Big-5 hyperscalers: $602B1
Peak net pressure
+92 bps
2028, central case
Crossover year
Beyond 2032
Under central case
Cumulative impact
+587 bps
2025–2032, probability-weighted

Net impact trajectory, 2025–2032

Capex pressure dominates near-term. Productivity offset builds gradually. Central case stays inflationary through the horizon.

Capex pressure (inflationary) Productivity offset (disinflationary) Net effect (probability-weighted)

Three capex trajectories

The 2032 outcome depends heavily on which path capex actually takes. Current hyperscaler guidance has shifted probabilities toward continued elevation.

Bust · ~15% probability
Dot-com style adjustment
Capex peaks 2026-2027, falls 75% through 2032. Requires demand disappointment or investor revolt. Current evidence does not support this: Microsoft has an $80B Azure backlog it cannot fulfill due to power constraints; Oracle has $523B in contracted performance obligations2.
Plateau · ~55% probability (central)
Peak extends through 2028-2029
Capex grows modestly through 2028, plateaus, then gradually declines as replacement capex takes over from new-build. Best fit for power and supply chain constraints. 2032 net impact: +48 bps. No disinflation crossover.
Growth · ~30% probability
Continued acceleration
Capex grows 10-15% annually through 2029. Alphabet's 2027 guidance of $250B (up 39% from 2026) is consistent with this path. Peak +110 bps in 2029; 2032 still at +88 bps. No crossover within horizon.

Empirical checks

Four mechanisms the disinflation thesis relies on. Each tested against current data.

Check 1 · Wage pattern
Gap widens, not compresses
High-exposure wages grew faster, not slower, post-LLM. The wage gap widened by +1.20pp. Two independent lines of evidence converge: a worker survey found 70% of respondents said AI gains flow to themselves rather than employers, and a calibrated GE model from Huang (2025, IMF) finds aggregate wages can rise up to +1.0% per 0.1pp increase in AI-adopting firms under realistic cost-savings assumptions. The pattern is consistent with Acemoglu & Restrepo (2018): capital accumulation in response to automation causes productivity gains to flow to labor in the long run, even as the labor share declines.
Check 2 · Capex scale
2–3× larger than forecast
2026 U.S. AI capex is ~$950B–1.05T, far above the $300–500B earlier models assumed. Analyst consensus has underestimated hyperscaler capex by 30+ percentage points in both 2024 and 2025.
Check 3 · Productivity signal
Early, fragile
MFP acceleration of +0.74pp above baseline, comparable to the late-1990s IT boom3. But only 3 post-LLM observations and decelerating within the window (1.63% in 2023 → 0.83% in 2025). Confounded by pandemic recovery and stimulus.
Check 4 · Surplus capture
Workers, not firms
Survey of 81,000 Claude users found 70% of those identifying a beneficiary said AI productivity gains flow to workers; only 10% cited employers. Standard disinflation channels require firm-side capture and pass-through to consumer prices. Worker capture weakens the pass-through.
Methodological note — local vs. aggregate effects: Studies finding negative AI impacts on employment and wages typically measure local effects (e.g., comparing high- vs. low-exposure commuting zones). These designs difference out general equilibrium effects like aggregate income gains and inter-regional reallocation. Huang (2025, IMF) formalizes this distinction: her calibrated GE model finds AI's local effect on wages is negative (-2.34%) while the aggregate effect can be positive (+1.0%) under realistic cost savings. This dashboard analyzes aggregate effects on services CPI; the companion Occupation Risk Explorer measures local exposure pressure. Both findings can be true simultaneously without contradiction.

Productivity test — the bull case's one empirical leg

Annualized growth rates, post-LLM (2023–2025) vs post-IT baseline (2005–2019). Data: BLS3.

Labor productivity vs multifactor productivity, by era

Metric 2005–2019 baseline 2023–2025 post-LLM Delta 1995–2004 IT boom
Labor productivity 1.52%/yr 2.42%/yr +0.90pp 2.89%/yr
MFP (multifactor) 0.57%/yr 1.31%/yr +0.74pp 1.38%/yr
Unit labor costs 1.59%/yr 2.20%/yr +0.61pp —
Caveats on the productivity signal
  • Only 3 post-LLM observations; confidence is weak
  • MFP decelerating within the window (1.63% in 2023 → 0.83% in 2025), the opposite of what AI ramp-up would predict
  • Pandemic recovery, immigration, and fiscal stimulus are confounds that can't be cleanly separated from AI attribution
  • 2005–2019 baseline was historically depressed; comparing to 1990–1994 levels would shrink the apparent gain
  • Unit labor costs are still rising faster than pre-pandemic, meaning wage pressure still exceeds productivity gains in the production function

Four productivity scenarios

The inflation offset depends on which productivity curve AI follows. The smooth-acceleration curve is a shape with no clear historical analog.

Historical reference: IT productivity boom
Peak 2003–04 (5-yr MA 3.2%)
AI productivity offset (basis points of inflation reduction)
2025–2032 model range
Probability weights: Faster than IT 35%, Same as IT 35%, Slower than IT 20%, Original model 10%. Fast adoption observed in survey data — 81,000 personal Claude users reporting 48% scope gains and 40% speed gains — supports the faster-diffusion trajectories over slow institutional diffusion. Under every combination except original-model-with-bust-capex, AI stays inflationary through 2032.

Scenario envelope — where does 2030 land?

36 combinations of productivity × pass-through × capex taper assumptions.

Disinflationary Neutral Inflationary
Synthesis

AI's net impact on services CPI is inflationary through the full 2025–2032 horizon under central-case assumptions. The probability-weighted expected path shows +64 bps in 2025 rising to +92 bps in 2028, then declining but remaining positive at +53 bps by 2032. Peak pressure arrives in 2028. Cumulative 8-year impact: +587 bps.

The disinflation crossover by 2032 requires a combination of: capex bust (~15% probability), smooth productivity acceleration (~10% probability of that specific curve shape), and firm-side surplus capture (the standard pass-through mechanism). Survey evidence shows workers are capturing 70% of productivity gains, not firms — which weakens the pass-through even when productivity gains are real.

The honest read from February 2026: AI is inflationary, period. Through 2030 across all plausible scenarios; through 2032 under the central case and the bull case. The 2023–2024 consensus that "AI is disinflationary" continues to look backwards — not just on timing, but on direction during the current build-out. Whether this reverses in 2033–2035 depends on whether productivity gains compound, whether capex genuinely saturates, and whether firm capture of AI surplus increases as enterprise deployment matures. None of these are resolved by current data.

References & sources

  1. CreditSights, "Hyperscaler Capex 2026 Estimates" — Big-5 capex projections, capital intensity ratios.
  2. Futurum Research, "AI Capex 2026: The $690B Infrastructure Sprint" — Hyperscaler guidance, Oracle RPO, Stargate project.
  3. IEEE ComSoc Technology Blog (Dec 2025) — Microsoft Azure power-constrained backlog, debt-financing analysis.
  4. CNBC (Feb 2026) — 2026 hyperscaler capex guidance; Morgan Stanley 2027 Alphabet projections.
  5. Goldman Sachs Research (Dec 2025) — Analyst underestimation pattern; $500B investment forecast.
  6. U.S. Bureau of Labor Statistics, Productivity and Costs program — Labor productivity, MFP, unit labor cost data; historical productivity series.
  7. BLS Monthly Labor Review, "The U.S. productivity slowdown" — 1995–2004 IT boom period productivity analysis.
  8. Massenkoff & Huang (2026), "What 81,000 people told us about the economics of AI" — Worker survey on AI productivity gains, job threat perception, surplus capture by career stage and wage quartile.
  9. Huang (2025, IMF Working Paper), "The Labor Market Impact of Artificial Intelligence: Local vs. Aggregate Effect" — Calibrated general equilibrium model finding aggregate wage effects of AI between -0.8% and +1.0% under realistic cost-savings assumptions; key methodological contribution distinguishing local (negative) from aggregate (potentially positive) effects.
  10. Acemoglu & Restrepo (2018), "The Race Between Man and Machine", American Economic Review 108(6) — Foundational theoretical model of automation versus new task creation; predicts capital accumulation causes productivity gains to flow to labor in the long run; identifies parameter regions where economy moves to full-automation BGP with collapsed labor share.