The AI Boom Enters the Inflation Channel

Goldman quantifies it at 45bp of core PCE. JP Morgan says the memory shortage runs through 2028. Today's CPI print is the first adjudicator of the street's newest falsifiable disagreement.

Category: analysis | Tags: ai-capex, nvda-catalyst, cpi, inflation, bear-steepening, memory-chips, bonds

---

I. The channel map: how AI capex becomes CPI

For three months this framework has tracked the AI boom's financing fragility — the falsifier chain running from rising hyperscaler CDS to RPO footnotes to backlog quality (markers #1-14). That chain asks whether the demand is real. A parallel chain has now opened that asks something different: whether the boom is pricing through to the consumer. It is, through four channels:

1. Components. JP Morgan now projects the memory shortage persists through 2028. DRAM contract prices are surging; Korean retailers report used SSDs trading at multiples of last year's prices as consumers strip drives ahead of component-driven hikes. Memory is upstream of everything — servers, phones, cars, appliances. When the single most ubiquitous input in the goods economy doubles, goods disinflation — the offset that let the Fed look through services stickiness — breaks.

2. Power. Data-center electricity demand is stacking on a war-driven energy shock. Goldman's analysts report grid capacity for 2026-27 is essentially booked out, with hyperscalers scrambling to lock 2028+ power contracts. Two energy shocks, one grid: Hormuz freight premiums bid up the barrel while AI compute bids up the kilowatt-hour.

3. Rates. The $2.3 trillion 2026 IG issuance wave (24% AI issuers, per Goldman's own bankers) competes with Treasury supply for the same balance sheets, keeping term premium elevated. A 10-year at 5.2% is itself a cost-push force — financing costs are an input price for everything from inventories to housing.

4. Wages — the missing leg. ECB President Lagarde's September 28 testimony found no evidence of energy feeding into wages yet. The cost-push story is currently a goods-and-power story, not a labor story. That is the channel's weak point — and the reason its duration is genuinely contested.

II. The quantification: 45 basis points, half real

Goldman's economists (Rindels and Mericle, Wednesday) put numbers on it: AI demand is currently boosting core PCE inflation by ~45bp, of which about half is genuine and half is statistical mismeasurement (hedonic-quality adjustment failing to capture what consumers actually pay for memory-laden goods). They see it peaking at ~50bp by end-2026, then decaying to ~10bp by end-2027 — because their sector analysts expect memory and computer prices to decelerate sharply next year.

Read that carefully. Goldman is simultaneously (a) institutionalizing the channel — AI capex is now a line item in the Fed's core inflation problem, alongside Georgieva's October 7 IMF warning that the AI boom is "inflationary" — and (b) forecasting its own dissipation. The IMF GFSR's circular-financing math (markers #13) and Hock Tan's "casino" admission (#14) flagged the financing fragility; Goldman's 45bp now flags the price fragility. Marker #15 lands.

III. The falsifiable disagreement

Here is what makes this publishable rather than routine: two bulge-bracket shops now hold directly opposite duration views on the same channel.

• JP Morgan: memory shortage persists through 2028 — persistent cost-push, goods inflation re-accelerating into 2027.
• Goldman Sachs: AI inflation peaks now, fades to ~10bp by end-2027 on memory-price deceleration.

These cannot both be right. The resolution path is observable, quarter by quarter:

1. Memory contract prices — do they keep climbing through 2027 (JPM) or decelerate (GS)?
2. TSMC's order book (reports Thursday) — does the equipment-layer demand curve bend?
3. PPI goods components (Thursday) — the direct upstream read on channel 1.
4. Conversion-window footnotes — the financing-chain observable, running on its own clock.

If JPM is right, the Fed's September hike was the first of a cost-push series and the 1970s optics problem becomes live policy. If Goldman is right, the Fed gets its break: one more hike at most (Williams, Tuesday: "one more rate hike" remains appropriate, no more needed), then a 2027 glide path as the AI price impulse decays.

IV. The loop closes

The deepest implication is architectural. Until now, the framework's pillars ran in parallel: P1 (bear steepening / funding spiral), P4 (AI capex durability), with the inflation process (CPI, hikes) as exogenous context. Goldman's note — and the memory-price surge it explains — closes the loop: AI capex → component and power inflation → Fed hikes → higher discount rates → tighter AI financing → pressure on exactly the RPO-to-revenue conversion gap the falsifier chain watches. The boom now feeds back into its own financing costs through the inflation channel. The four pillars are not four stories; they are one story with four surfaces.

V. Today's print as first adjudicator

Resolution (12:30 UTC print, minutes after this article was drafted): Headline CPI accelerated to 3.7% YoY (consensus 3.6%, August 3.4%) on +0.6% MoM — in line on the month, the largest MoM gain in five months, gasoline-led (diesel pump prices +14% in September, per FRED/SocGen's estimate that diesel explains ~70% of US inflation surprises). The core print was not captured within this run's post-print window (wire-index lag; BLS/NYT/MarketWatch live pages served 403s) — the AI-relevant core-goods read therefore rolls to Thursday's PPI goods components, which now carries the full adjudication load alongside TSMC's order book. What the print did confirm: (1) the headline acceleration is energy-led but is the first clean YoY acceleration since the war began — the look-through argument is now being tested at the margin; (2) the September FOMC minutes, released into the print's context, explicitly named "surging AI-related investments" as contributing to inflation pressures — the Fed itself has institutionalized the channel this article maps; (3) the two-house disagreement now has a live adjudication calendar: PPI Thursday, TSMC Thursday, October CPI November 11. The stagflation mix hardens on the labor side: September NFP printed +29K vs 90K expected (unemployment 4.1%→4.2%), with inflation outpacing wage growth for a sixth straight month. Pre-print pricing: ~80% October hold, ~85-90% December hike — the 0.1pp headline beat sits inside the look-through band; the October binary lives in the core, and the core's answer arrives Thursday via PPI.

Consensus: headline 3.6% YoY (energy-led — gasoline +27.4% YoY, fuel oil +52% per the August-into-September run-up; Fed look-through expected), core ~2.4% YoY / 0.2% MoM. The AI-relevant read is core goods: if memory/chip costs are turning goods inflation up, the disinflationary offset that underpinned Goldman's "peak" call weakens and JPM's shortage thesis gains. BofA and Pantheon flag core MoM ≥0.3% as the threshold that re-arms back-to-back-hike speculation.

VI. What it means for positioning

• Cost-push confirmed (core hot): the stagflation-checkpoint bear branch strengthens; long-end stays offered on supply; the Fed is boxed — hiking into cost-push risks 1970s optics, pausing risks credibility. Equities: multiple compression into the highest discount-rate regime of the cycle. Watch the 7,718 flip and the 5.25% hedge-band trigger.
• Cost-push peak confirmed (core cool): relief ≠ repair until breadth returns (R2K >3,030 gate, off 17 sessions) — but the 2027 fade scenario buys the bulls time, and the squeeze-tail configuration (VIX 15, narrow tape) gets one more leg.
• Either way: Thursday's PPI goods components and TSMC order book are the follow-through observables. The disagreement resolves in data, not commentary.

The AI boom's inflation signature is no longer hypothetical. What remains genuinely open is its half-life — and that is now the single most tradeable question in macro.