Definition

“Speed to power” names the deployment-timeline competitive advantage that hyperscalers chase when grid interconnection queues stretch beyond the revenue-generation timeline of an AI buildout. Behind-the-meter on-site generation deploys in approximately 18 months; grid interconnection queues run several years. The 18-month-vs-multi-year delta is what flipped hyperscaler power strategy from “negotiate with utility” to “build a private grid.”

Why it matters for the newsletter

Speed to power is the operating constraint that explains the off-grid hyperscaler pivot. It’s adjacent to Chokepoint Control (grid interconnect is the chokepoint being routed around) but distinct: chokepoint control is about who holds the gate; speed to power is about how long the gate takes to open, and what the operator does instead.

The concept also explains why behind-the-meter natural gas (Williams’ $5.1B portfolio) and fuel-cell microgrids (Bloom Energy’s Project Jupiter deal) have both become primary architectures rather than backup architectures.

Evidence & examples

Tensions & counterarguments

  • The Khajuria (Intel) constraint: “Fuel cells have very limited capacity to handle overloading.” Behind-the-meter primary power may be less resilient to demand spikes than grid power. The bet is that AI workload steady-state is so high that overload protection isn’t the dominant design constraint anymore. If that bet is wrong — i.e., if AI training cycles produce unpredictable load spikes that exceed fuel-cell ramp capacity — the speed-to-power strategy has a reliability hole.
  • ESG and carbon-lock-in tension: 15-20-year asset lifetimes for behind-the-meter gas conflict with most climate trajectories’ assumed retirement schedules for fossil-fuel infrastructure.
  • A second tension: speed-to-power strategies bypass the political process that grid-connected hyperscalers participate in. State PUCs, FERC, and ratepayer interests have less leverage over off-grid campuses than over grid-connected ones — which is sometimes an advantage for hyperscalers and sometimes a vulnerability (community opposition shifts to local-permitting layers).

Key sources