Kyle Harrison
article

What Is So Hard About Behind-The-Meter Power for Datacenters Part 1

Ellie Holbrook, Robert Boswall, Jeremie Eliahou Ontiveros, Nicolas Bontigui, Dylan Patel September 10, 2026 View original ↗

What Is So Hard About Behind-The-Meter Power for Datacenters Part 1

SemiAnalysis, September 10, 2026, by Ellie Holbrook, Robert Boswall, Jeremie Eliahou Ontiveros, Nicolas Bontigui, and Dylan Patel. Subtitle: “Dumb Science Experiments vs. Money Printing Machines.”

The scale of the market

The numbers are staggering even for a sector accustomed to large figures:

  • 75 GW of firm binding orders for BTM power currently tracked by SemiAnalysis
  • ~20 GW ordered in Q2 2026 alone
  • 22 OEM manufacturers now carrying multi-hundred-MW orders

The economics that are driving it

The key insight is the inference economics at scale:

Inference API revenue can yield $100 billion per GW per year at 90%+ gross margins.

At those unit economics, paying $1–3B to build a gigawatt of dedicated power infrastructure is easily justified. The hyperscalers and AI labs are not waiting for utility interconnection queues that run 5–8 years — they are building their own generation assets that connect directly to the datacenter load (“behind the meter”) and bypass the grid entirely.

Specific deals

  • Microsoft — 5 GW commitment (BTM/dedicated generation across multiple sites)
  • Google — 930 MW of aeroderivative gas turbines (airline-engine-derived, fast-start turbines) + 900 MW Bloom Energy fuel cells
  • Anthropic + Meta — partnering with Enchanted Rock for BTM microgrid solutions
  • OpenAI — 1.4 GW campus in Shackelford County, TX

What makes BTM hard

The piece is Part 1 of what will presumably be a multi-part series. The structural challenges include:

  • Interconnection is bypassed but not eliminated — BTM power still requires gas pipeline capacity, permits, land, cooling, and distribution within the campus.
  • Aeroderivative turbines (the fast-start gas turbines used by Google and others) have long lead times from a limited supplier base; GE Aerospace and Siemens Energy are dominant, and they are sold out.
  • Fuel cell deployments (like Bloom Energy’s 900 MW with Google) depend on H₂ or NG supply infrastructure.
  • OEM manufacturing capacity — 22 manufacturers with multi-hundred-MW orders means the supply chain is stretched.

Connections