What is So Hard About Behind-The-Meter Power For Datacenters? Part 1
Source
Ellie Holbrook
Author
Ellie Holbrook
Date
Terms in this piece · Glossary
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
Why it matters
Quantifies how much AI compute capacity is now genuinely locked into behind-the-meter power (75GW of binding orders, ~20GW added in Q2 2026 alone), a hard constraint on how fast new inferenceRunning a trained model to get answers — the phase where AI is actually used, as opposed to trained.Full definition → and training capacity can actually come online.
Key quotes
“Paying 2x more money or accepting 30% lower efficiency for faster speed of deployment becomes a no-brainer.”
“Anthropic and its peers can pay back the value of a power plant in 20 days of inference revenue.”
“A turbine order and a nearby pipeline do not make a project.”
articleSpaceX 10GW in 2027 – Why It’s Real, Will Drive $300B ARR for SpaceX, and Why Microsoft Will Be the Largest OfftakerJeremie Eliahou Ontiveros
articleBuild Ai In AmericaAnthropic News
articleAmazon’s AI Resurgence: AWS & Anthropic’s Multi-Gigawatt Trainium ExpansionJeremie Eliahou Ontiveros