AI Cost Watch

AI Cost Watch — a recurring, dated deflation note on AI-infrastructure spend. Each issue states one falsifiable forward call: whether, and when, the buildout's demand-side and physical-supply assumptions break. The standing tell is a down-revision in hyperscaler capex guidance — until it fires, the read is "still expansion"; next read is the late-July 2026 Q2 reports. 

This record collects Issue 1 (4 Jun 2026) — spend hasn't rolled over; DeepSeek's permanent 75% price cut against intact ≈$660–725B capex, the bifurcation as the read — and Issue 2 (13 Jun 2026) - the month AI spend got metered (coding tools and buyers self-capping), forced-buyer index flows (the SpaceX IPO), and a government model-off-switch (the Fable/Mythos export suspension). Findings map onto the six-factor risk register of the companion Capital-Infrastructure Barbell (DOI 10.5281/zenodo.20586864). 

Changes in v2.1. Corrected Issue 2's SpaceX figure: Morningstar's $780B fair value is ~55% below the $1.75T IPO target, not "48%" (an orphaned stale-base figure - 48% is $780B against an earlier ~$1.5T valuation, never recomputed when the mark moved to $1.77T; the fix widens the fair-value gap, i.e. moves conservative). Repackaged to portfolio standard - Markdown sources, house-style PDFs, and a README - replacing the v2 .docx files. Issue 1's then-current  Anthropic ≈$30B run-rate is retained as a dated snapshot; Issue 2 transparently restates it to ≈$47B (the trail isthe point). 

Author: N. Milton (ORCID 0009-0003-4213-7769), independent, no positions in the companies named. Produced with AI assistance (Claude Opus 4.8, Anthropic). Conflict of interest: Anthropic is among the companies tracked; lines where Anthropic is the subject are non-neutral and marked [1] in-text, with primary coverage preferred. Open-science documentation only — not investment advice. CC BY 4.0.