06-reference

stratechery ai on the ballot data center backlash

2026-11-04·reference·source: Stratechery·by Ben Thompson
ai-capexdata-centerscapital-cyclepoliticsnvidia

Why this is in the vault

Thompson argues the 2026 midterms (Democratic landslide, Texas electing a governor who campaigned on a data-center moratorium, nearly every local anti-data-center ballot measure winning) show political opposition — not compute, memory, or power — is now the binding constraint on AI buildout, and lays out ten downstream implications across the AI trade.

The core argument

The constraint on AI dispersion keeps shifting: first compute (solved via ramp), then memory (ramping), then power (being addressed) — and now political permission to build, which Thompson calls harder to solve than the others because it actively blocks solving them. Unlike globalization's factory closures, data-center construction requires local permission, so communities that had no say in offshoring are expressing disapproval the one way they can: blocking permits. He stops short of calling this permanent but treats it as the most likely near-term binding constraint, with ripple effects: the AI trade's recent weakness looks right in hindsight; neoclouds (capital-constrained, lower political sway) lose relative to hyperscalers (more existing/repurposable sites); Nvidia is short-term hurt by fewer builds but long-term helped because scarcity makes tokens/megawatt efficiency more valuable; China's power abundance partially offsets its chip-lag; SpaceX's space-data-center thesis gets more credible as terrestrial siting gets harder; frontier labs benefit because public pressure effectively paces the compute arms race their margins couldn't otherwise sustain; Apple benefits from any slowdown that eases the memory crunch and from edge-inference's rising relative importance; startups lose access to cheap compute while frontier labs consolidate the user relationship; and ordinary users may not get the "free compute" dividend of a Web 2.0-style oversupply crash.

Mapping against Ray Data Co

Direct hit on the founder's Markov capital-cycle investing thesis ([[project_investing_markov_capital_cycle]]): the thesis treats phase shifts in the AI buildout cycle as the source of alpha, and Thompson is naming a new, non-technical phase marker — political permitting — that the backtest-thesis framework hasn't modeled yet (prior phase markers were compute/power/capital availability, not local-ballot risk). It also sharpens the Nvidia-vs-neocloud dispersion already tracked across the Nvidia/SpaceX notes below: Thompson's claim that scarcity raises the value of tokens-per-megawatt efficiency is a testable, falsifiable prediction worth flagging for the next capital-cycle phase-marker pass.

⚠️ Sponsorship

Stratechery carries a standing Stripe sponsorship (per project README); this issue's sponsor section promotes Stripe's own "SaaS Index" research (built from its payments data on 72,000 non-AI SaaS businesses) to argue SaaS revenue growth is accelerating despite the "SaaSpocalypse" stock selloff narrative. Bias implication: Stripe has a direct commercial interest in a bullish SaaS-health narrative, and the data/analysis is Stripe's own, not independently verified by Thompson. Treat the SaaS-growth claim as vendor-sourced, not neutral.

Related