AI-Capex Capital Competition — Pre-registration

The thesis

We register a falsifiable claim, not a domain label: private AI-infrastructure capex — hyperscaler data-center buildout plus the semiconductor supply chain feeding it — has become a first-order, independently observable macro driver, operating through two separable legs tracked independently and never collapsed into one verdict.

Financial leg, capital competition: sustained double-to-triple-digit capex growth absorbs private savings and creates correlated corporate financing needs, competing with Treasury supply for the same capital pool and bidding up the long end independent of inflation expectations. Our sibling US Treasury yield-repricing series lists this competition among its term-premium drivers but does not instrument it (pre-registration, 2026-08-12); capex decelerating while long yields stay elevated would cost that thesis its non-fiscal demand pillar.

Physical leg, power demand: data-center load is the marginal driver of US electricity demand growth, and because buildout leads energization by one to three years, capex is its leading indicator. A downward capex inflection would crater the data-center demand layer of the power market; policy moratoria — Texas, August 2026 — are early indicators of such an inflection.

The legs share one impulse and should decouple only when the decoupler is identifiable — power constraints capping operation while hardware is bought ahead of sockets, for example. We claim no monopoly: fiscal supply is the rival driver, and discriminating between them is part of the job.

What would prove it wrong

Four conditions, all not-met at registration.

F1 and F2 bracket the financial leg from opposite sides. Capex negative year over year for two consecutive quarters while the 10-year holds at or above 5.0% means elevated long yields are not capex competition. A boom sustained at or above +20% for four or more consecutive quarters while the 10-year falls below 4.30% and stays there means the channel is insufficient — fiscal supply owns the long end. F3 guards the physical leg: an absolute-downward EIA or IEA revision to data-center demand while capex still booms would mean the demand layer breaks before the investment does. F4 is the independent-observability channel: a central-bank document naming private AI-capex financing demand as a source of long-end pressure would corroborate from outside; persistent absence marks it a dead channel.

Nothing is close today. The live discriminating tests are F2 and F3, though F2’s streak clock — sitting at its first observed quarter — is not yet testable. No falsifier met is not confirmation: survival is not score, and confirmation requires positive evidence.

Where things stand

Support: the five US hyperscalers’ most recent earnings releases, as of 2026-08-28, put capex growth year over year at Microsoft +109.6%, Alphabet +100.1%, Amazon +76.7%, Meta +82.1%, Oracle +81.6% — median +82.1%. The aggregation is our own computation from public filings, not a national statistic; fiscal calendars are mixed (Oracle’s latest quarter ends in May, the others’ in June). The demand pull is reported sales, not speculative inventory: NVIDIA’s most recent reported quarter as of late August 2026 printed $96.2bn of revenue, +106% year over year (earnings release, 2026-08-26; quarter ended 2026-07-26), after $81.6bn and +85.2% in the prior release (2026-05-20). The level leg (a) must explain is elevated — 10-year 4.67% then 4.73% (FRED DGS10, 2026-08-27/28), real 10-year 2.32–2.42% (FRED DFII10) — but the sibling-series cross-check is consistency under a shared backdrop, not isolation of the channel; we grade it weak.

Against, at full strength: the physical leg shows its first policy crack. On 2026-08-03 the Governor of Texas halted state-level approvals for new data centers pending a grid-operator audit of interconnection-seekers (Governor’s statement and Texas Tribune, 2026-08-03; Houston Public Media, 2026-08-27). The EIA then cut its Texas 2027 electricity load-growth forecast from 14% in the July outlook to 6%, citing the August 3 order directly (EIA press release, 2026-08-11; EIA Short-Term Energy Outlook, August 2026 edition). The boom has not moved the broad commodity complex: the same outlook unwinds the supply-driven oil spike to a $69-per-barrel Brent average in 2027, at its 2025 level. And the 10-year breakeven at about 2.31% (FRED T10YIE, 2026-08-28) keeps the transmission channel clean but unproven.

The opposing view, strongest form: capital competition is financial folklore — Treasury net supply dwarfs even booming hyperscaler issuance by an order of magnitude, the long end reprices on fiscal durability and term-premium normalization, and capex is narrative wallpaper along for the ride. The +82% median partly reflects a base effect (the year-ago quarter is itself the boom’s start, not a normal base), and the leg that would confirm the thesis most cleanly is the leg already cracking.

Registered response: the base-effect critique is partially accepted — against an elevated base, +82% reads momentum conservatively, and F1/F2 test sign and persistence, not levels. The scale point we concede, restating the claim narrowly: marginal explanatory power at the capex inflection, where F1/F2 discriminate — not dominance over fiscal supply. Texas is the first F3 test case, not explained away.

Honesty: the five-name set misses ex-US, sovereign, and neocloud capex, so a global plateau could look US-masked; restatements and mixed calendars add ±1-quarter noise; a median of five noisy series is not a national-accounts quantity; a composition shift — who capexes — would read as deceleration before it is one. The +20%/5.0%/4.30% thresholds are pre-registration round numbers, not calibrated output. Open as of this date: the funding mix of marginal capex (debt vs cash flow), which sets how directly it competes with Treasury supply; whether guidance leads reported capex reliably; whether Texas is idiosyncratic or the first of a moratoria wave; where the electricity-constraint ceiling sits; whether any central bank ever names the channel. Confidence at registration: medium — the capex boom is verified; the two transmission claims remain to be tested.

How to read this series

Every episode carries the full ledger forward, scoring each condition against its threshold as of that date; conditions are never dropped silently, and amendments appear only as dated revised entries stating the reason. Cadence is irregular and follows catalysts — earnings prints and guidance (Q3 2026 in late October), the EIA’s annual demand revision (November 2026), permitting decisions at any time, central-bank documents — not the calendar. The series closes with a sunset ledger and one terminal verdict — confirmed, falsified, or unresolved — scored against these exact numbers.