The next 1,000x of AI inference won't fit in giant data centers: 41% runs below 20MW
antavedissian · x · 2026-11-12
Anthony Avedissian's long-form analysis of where the next 1,000x of AI inference will run. Inference will be 1,000x today's volume in five years, but giant data centers won't hold it all: their model suggests that in 2031, roughly a third of tokens still run on hyperscale campuses, a quarter on regional sites, and 41% below the 20 MW line — on phones, laptops, robots, office racks, and small metro sites next to solar farms.
The key economics: a small site costs 1.3–3x more per token than a giant campus because its GPUs sit idle when nobody nearby needs them; software that routes work to whichever site is free closes most of that gap — which the author sees as where the next set of AI infrastructure companies gets built.
Context: US midterms saw 28 of 41 local data-center measures across 10 states pass to limit development, yet hyperscaler capex tracks above $800B this year and Amazon says 2027 will be "meaningfully more" — the fight just moves to statehouses, utility commissions, and town halls.
Related event: Analysis: 41% of AI inference to run on sub-20MW sites by 2031(5 posts)→
More from Infra
- Polymarket prices 28% odds of a state data center moratorium; Torus raises $500M — Polymarket · 2026-11-12
- AWS-wrapper cloud startup shuts down just one month after seed round — thedealdirector · 2026-11-12
- Utah startup Torus raises $500M for shipping container-sized AI data centers deployable in 4 months — Polymarket · 2026-11-12
- Greylock reading group on AI accelerator design: more transistors no longer mean more performance — _shreya_s · 2026-11-12
- Dennard Scaling Is Dead: Automated Chip Design Could Shrink Hardware Cycles — _shreya_s · 2026-11-12
- Two smuggling scandals put Nvidia's China denials to the test — fortune · 2026-11-12