The Projection — a symmetric watercolor butterfly

The Projection

The surface is never the system.

AI

THE CAPEX PICTURE: four actors, one table

· part of Hyperscaler Capex

What happened

*Four parallel destination crawls (findings + 167 sources in artifacts/bundles/-capex-2026-07-27/). The comparable read: | | capex (2026) | sites | silicon | power secured | payoff claim | |---|---|---|---|---|---| | Google | $195–205B guided | 15+ states + India — a land-grab | TPU v7; Nvidia split undisclosed | solar-heavy + Kairos Gen-IV '30 | $514B backlog, Cloud +82% (press-only) | | Microsoft | $80.1B 9-mo FY26 (FY total Wed) | WI+GA superfactory ~2GW + Pecos ~2GW | Maia 200 dual-purpose; split undisclosed | TMI/Crane '27 + Kilby 2.7GW | $37B AI run-rate +123%; $250B OpenAI unmapped | | Amazon | ~$200B plan | Rainier live + MS/PA/NC $55B | 1.4M Trainium; split undisclosed | Talen 1.92GW firm + X-Energy 5GW '39 | $364B backlog excl. Anthropic $100B; landlord model | | Meta | $135–145B guidance (our $76B TTM flagged) | Hyperion 5GW — 80% sold to Blue Owl; El Paso pushback | MTIA "millions" H100-eq; Broadcom '29 | nuclear ≤6.6GW RFP BUT RE100 exit → 7.5GW gas | superintelligence framing softening; Moody's warning | The four cross-cutting findings: 1. The silicon split is the industry's common non-disclosure — 4 of 4 refuse the own-chip vs Nvidia breakdown. The labs leak what the hyperscalers won't (Anthropic: 1M TPUs and 1M Trainium). 2. The payoff layer is the weakest — backlogs and run-rates are press-corroborated; primary filings were unreachable or the mappings (e.g. $250B OpenAI → capex) simply don't exist publicly. 3. Power is the binding constraint being solved three ways — nuclear restarts ('27) → Gen-IV/SMR options ('30–'39) → and now Meta's open gas pivot, the first break in the clean-power consensus. 4. Capex is migrating off balance sheets — Meta→Blue Owl 80%, BlackRock's $12.5B bonds, Microsoft's lease book: the buildout is becoming other people's balance sheets. (Feeds Circular Financing / Nvidia as Lender.)

Sources 0

No source link was recorded for this entry — the prose above is the record.

Credibility ratings combine two public datasets: Lin et al. 2023, PNAS Nexus 2(9):pgad286 (CC BY 4.0) for the domain-quality score, and Wikipedia's perennial-sources list (CC BY-SA 4.0) for the reliability tier. A rating describes an outlet's general record, never this particular article. Unrated means absent from those datasets — not a judgement.

The wider story

This event sits inside Hyperscaler Capex, which carries how it developed before and since.

Story 4 of 14 in this thread's timeline, oldest to newest

Adjacent threads (shared entities):