~15GW of 2027 AI Compute Capacity Projected to Be Stranded Without Power
Summary
- • Consensus estimate: ~15GW of AI compute built in 2027 won't have power to operate
- • Bottlenecks include grid interconnects, transformers, cooling, permitting, and turbine supply
- • Goldman Sachs, Morgan Stanley, and SemiAnalysis all flag a U.S. AI power crisis by 2027
- • FUNDA estimates North American demand at 35GW vs. deliverable supply of only 16.5–23.4GW
Details
~15GW of 2027 AI Compute Likely to Be Stranded
Consensus across analysts estimates that ~15GW of AI compute capacity produced in 2027 will lack sufficient power infrastructure to operate, representing a major stranded-asset risk for AI hardware investors.
Site-Level Bottlenecks Go Beyond Raw Grid Power
Blockers include grid interconnections, transformers, wiring, liquid-cooling systems, chillers, networking, permitting delays, and turbine supply — meaning power cannot be energized even where grid headroom nominally exists.
North America: 35GW Demand vs. 16.5–23.4GW Deliverable Supply
FUNDA analysis (Aug 27, 2026) estimates 2027 North American AI data-center IT load demand at ~35GW against a deliverable supply of 16.5–23.4GW — a 1.8–2.1× demand-to-supply gap.
Goldman Sachs: U.S. Power Demand to Double to 66GW by 2027
Goldman Sachs projects U.S. data-center power demand rising from 31GW in 2025 to 66GW by 2027, with only 50–60% of scheduled capacity expected to come online on time.
Morgan Stanley: 38GW U.S. AI Power Shortfall Through 2028
Morgan Stanley estimates a 38GW power shortfall for U.S. AI infrastructure through 2028, corroborating the broader multi-analyst consensus on a structural energy gap.
SemiAnalysis: U.S. Grid Headroom Turns Negative in 2027
SemiAnalysis forecasts U.S. grid headroom turning negative by 2027, meaning total AI demand will outpace available power capacity at a national level.
Permitting Delays Hit 78% of 43GW in Pending Projects
Permitting delays are affecting 78% of the 43GW in pending data-center projects; behind-the-meter turbine orders are also constrained through 2031, severely limiting off-grid power options.
Elon Musk X Post on 15GW Shortfall Reaches 6.4M Views
Musk's Aug 29, 2026 post on X citing the ~15GW stranded compute estimate went viral with 6.4M views, bringing the AI infrastructure power gap into mainstream investment discourse.
Details compiled via Grok live web research
Supporting analyst figures and citations sourced via Grok live web research drawing on CryptoBriefing, 24/7 Wall St., and Odaily coverage of the underlying analyst reports.
Research = analyst/institutional estimates; Infrastructure = physical supply-chain and grid constraints; Stat = quantified data point; Context = background framing
What This Means
A convergence of estimates from Goldman Sachs, Morgan Stanley, SemiAnalysis, and FUNDA confirms that the AI infrastructure buildout is racing ahead of North America's ability to power new data centers. Roughly 15GW of AI compute scheduled for 2027 deployment risks being built but unable to operate — a potential stranded-asset crisis for AI hardware and hyperscaler investment. Critically, the bottleneck is not just grid availability but the full stack of site infrastructure: transformers, liquid cooling, permitting, and turbine deliveries. This structural energy gap could significantly slow AI deployment timelines, reshape where hyperscalers build, and accelerate demand for alternative power sources such as small modular reactors and behind-the-meter generation.
Sentiment
Concerned but analytical — infrastructure bottleneck seen as major constraint and investment opportunity
“This is the part nobody talks about. Everyone's racing to build AI compute. Almost nobody is talking about the infrastructure gap behind it. 15GW sitting dark in 2027 isn't a power problem. It's a supply chain, engineering, and timeline problem all at once.”
“Elon says ~15GW of the AI compute built in 2027 can't be turned on in 2027, and that the hard part is transformers, wiring, cooling and chillers, not finding the power. the filings have been itemizing exactly that gap for weeks: all six US grid operators asked FERC to pause its data-center connection rule…”
“Consensus estimate is that ~15GW of AI compute produced in 2027 cannot be turned on in 2027. This is harder than just finding power… There isn't even enough energy or infrastructure to supply the demand. And it can't take priority of the supply of the rest of civilization. Ultimately these data centers will also have to build their own power supply.”
“America’s AI race is running into a bottleneck no GPU can solve: Electricity. Up to half of the U.S. data center projects planned for 2026 face delays or cancellation as power availability, transformers, switchgear and local opposition collide with the AI buildout. This is not a chip problem anymore.”
Split
~80/20 analytical concern vs. opportunity framing (most agree on the bottleneck; split is whether it stalls AI or accelerates behind-the-meter/SMR plays)