Half of US Data Centers Planned for 2026 Are Being Canceled or Delayed — AI's Power Wall Is Real
Sightline Climate's 2026 Data Center Outlook finds that roughly 50% of planned US data center capacity for 2026 faces delay or cancellation, driven by power grid constraints, equipment shortages, and community opposition. Only 5 GW of the planned 16 GW has entered active construction.
Original sourceThe AI infrastructure buildout has hit a wall — literally. Sightline Climate's 2026 Data Center Outlook, released this week, finds that of roughly 16 gigawatts of planned additional US data center capacity expected to come online in 2026, only about 5 GW has actually entered the construction phase. Somewhere between 30% and 50% of planned projects face delay or outright cancellation.
The primary culprit is power. Twenty-five percent of planned projects haven't even disclosed their powering strategy — a sign that site selection happened before grid connection agreements were in place. The electricity grid simply was not built for concentrations of 100-500 MW loads at individual sites. Transformers, switchgear, and battery backup systems are backordered 18-36 months; the supply chain that makes them runs heavily through Chinese manufacturers now subject to export restrictions.
Alphabet, Amazon, Meta, and Microsoft have collectively committed more than $650 billion in AI infrastructure spending in 2026. The capital exists. The grid capacity does not. Utility interconnection queues in Virginia, Texas, and Georgia — the three dominant US hyperscale markets — are measured in years, not months. Projects that secured land and permits in 2024-2025 are sitting idle waiting for power agreements.
Community opposition has also become an organized force. Data centers consume vast water for cooling, generate constant low-frequency noise, and typically employ few local workers. Local governments in Nevada, Arizona, and Iowa have passed moratoriums or restrictive zoning laws.
The investment implication is counter-intuitive: the companies positioned to benefit most are those with existing power contracts and cooling infrastructure — established hyperscalers — not new entrants. Nuclear power developers, transmission equipment manufacturers, and grid software companies are the surprise winners of the AI buildout.
For AI developers and startups, the near-term effect is rising compute costs. Cloud GPU availability will tighten through mid-2027 as new capacity remains delayed. The inference economics that looked favorable six months ago are being revised upward across the industry.
Panel Takes
The Builder
Developer Perspective
“This is already showing up in spot GPU pricing — H100 spot rates have climbed 35% since January. If you're building inference-heavy products, now is the time to lock in reserved capacity agreements at today's prices. The companies that didn't are going to get squeezed hard in Q3.”
The Skeptic
Reality Check
“The AI labs are projecting compute demand based on exponential adoption curves that history suggests will plateau. If scaling laws continue to slow and demand doesn't hit projections, we get a different version of this story — stranded assets and written-down capex. The grid constraint is real, but the assumption that demand will outpace it is not a given.”
The Futurist
Big Picture
“The power wall forces efficiency innovation — and that's net positive for AI progress. The models of 2028 will be dramatically more compute-efficient than today's precisely because the infrastructure bottleneck makes efficiency economically essential. Every constraint in the history of computing has been solved by the next architecture, not by building more of the same thing.”