Data Center Power / Permitting Constraint

Why this is in the vault

A large, sustained cluster of vault documents (dozens tagged, with a dedicated tracker brief and multiple thesis-anchor notes) keeps returning to the same structural claim: the binding constraint on AI infrastructure buildout is not chips, it's power and the local political permission to site the load that consumes it. This is distinct from the chip/compute-supply story RDCO already tracks in its memory-cycle investing thesis — it's a second, physical-world bottleneck layer (grid capacity, transmission, local permitting/NIMBY opposition) that sits underneath the compute story rather than replacing it.

The concept has two coupled halves that the vault keeps testing against each other:

• The generation/distribution half: [[2026-08-13-moonshots-200gw-grid-sodium-batteries-wave-datacenters]] cites Ramez Naam's framing that the actual bottleneck is the distribution grid, not generation capacity — there may be ~200GW of usable slack hiding in grid inefficiency rather than a hard generation ceiling. This is a more optimistic read than the headline "power crunch" framing.
• The political/permitting half: [[2026-05-18-stratechery-data-center-discontent]] documented data-center opposition going mainstream and bipartisan (Gallup: 71% oppose a local data center, exceeding historic nuclear-plant opposition), with concrete moratoria (Hill County TX, DeForest WI rejecting a 1.6GW project). [[2026-08-29-data-center-opposition-permitting-tracker]] followed up seven months later to test whether that sentiment is actually converting into stalled capacity — and found a genuine contradiction: opposition is at an all-time high (533 local moratoria, 294 in the first seven months of 2026 alone, per the Moratorium Nation dataset) simultaneously with record construction starts ($81.5B YTD through June 2026, already exceeding all of 2025). The honest read is that opposition is relocating the buildout (out of contested townships and into Texas, the Dakotas, behind-the-meter campuses) rather than capping it — conversion of NIMBY sentiment into actually-killed projects is real but small (roughly 40 documented defeats against 800+ live projects, per FracTracker).
• The 2026-08-29 Innermost Loop digest ([[2026-08-29-innermost-loop-power-scarcity-anthropic-cursor-win]]) threads the same tension into a live news item: Musk warning that ~15GW of contracted 2027 compute capacity can't actually be switched on for lack of power, prompting SpaceX to build its own turbine-blade factory — the physical constraint is being answered with vertical integration, not with slower buildout plans.

Mapping against Ray Data Co

This constraint is the physical-layer anchor underneath RDCO's power-cycle investing thesis ([[2026-05-17-power-cycle-v1]], updated in [[2026-05-18-power-cycle-v1.1]]), which already tracks ERCOT interconnection-queue growth (5.1x, 68→346GW) as a phase-transition signal, and it is explicitly the reason [[2026-05-17-power-anchors-backfill-notes]] built an annual LBNL interconnection-queue ingestion cron. RDCO's own operational exposure to this constraint is indirect — RDCO consumes compute at retail through Anthropic subscriptions and doesn't procure megawatts directly — but the constraint still matters for two reasons: (1) it's a leading indicator inside the capital-cycle thesis for when frontier-lab compute growth decelerates, which in turn affects how fast frontier model capability (and therefore RDCO's [[2026-05-10-harness-moat-two-layers-portability]] bet) advances; and (2) the moratorium-vs-construction contradiction is a caution against taking any single "AI buildout is stalling" headline at face value — the vault's own dollar figures here (the widely-cited Data Center Watch numbers) have been flagged as methodologically unverifiable and should be qualified, not repeated uncritically, whenever this thesis gets cited externally.