AI Data Center Crisis 2026: Half of US Projects Cancelled or Delayed

Roughly 50% of US AI data center capacity planned for 2026 has been cancelled or delayed — a structural shock driven by three converging shortages: electrical infrastructure, supply chain components, and local community consent. The damage is concentrated: only about 4-5 GW of the 12 GW announced is genuinely under active construction, with most of the remainder either stuck waiting on grid interconnections or formally withdrawn. The immediate trigger is power. AI campuses now routinely request 500 MW to 1 GW of guaranteed capacity — more than the entire load of mid-sized cities. Utility companies in PJM, ERCOT, and MISO regions have begun rejecting or queuing new requests beyond the 2028-2029 horizon, leaving developers holding pre-leased land they cannot energize. Project Jarvis, a 1 GW, $13.5 billion campus in Port St. Lucie, Florida, was withdrawn in February 2026 after local planning boards refused to certify its water and grid impact studies — exactly the kind of mid-stage cancellation now happening across the country. A second layer is supply chain. Power transformers — the dull, decades-old steel boxes that step voltage down to usable levels — have surged in price by 77% since 2019, and custom units now carry lead times of up to five years. China supplies more than 30% of US transformer and switchgear imports and over 40% of battery imports; every tariff escalation in 2025 directly translated into longer queues for hyperscaler procurement teams.

Transformers are the most boring part of the AI revolution — and the part most likely to break it. A typical hyperscale campus needs dozens of large units, each weighing several hundred tons and requiring 18 to 60 months to manufacture. Domestic US production capacity for grain-oriented electrical steel — the specialty alloy that goes into transformer cores — has not grown since 2018, while global demand has roughly doubled. The practical consequence is queueing. A developer that signs a power purchase agreement today, even with utility cooperation, will not see meaningful transformer delivery before 2029. Several projects originally targeting 2026 commissioning have been quietly slipped to 2028 or 2029 in earnings call footnotes. Memory and storage costs have made the problem worse: DRAM is up roughly five-fold and NVMe storage three-fold since Q1 2025, eroding the unit economics even of fully-permitted builds.

Local opposition to data centers has shifted from scattered NIMBY complaints to organized, state-level political pressure. Tracking databases logged over 100 new conflicts in Q1 2026 alone, and at least 10 states — Louisiana, Michigan, New York, Ohio, Virginia, Georgia, Texas, Indiana, South Carolina, and Tennessee — have introduced moratorium bills or strict siting restrictions in the 2026 legislative session. The complaints follow a consistent script: water consumption (a single 100 MW facility can use up to a million gallons a day for cooling), residential electricity rate hikes (utilities argue infrastructure costs get socialized), and quality-of-life concerns over light, noise, and truck traffic. The Florida-based Project Jarvis cancellation was less about Florida specifically and more about a national pattern — planning boards across the country are increasingly comfortable saying no, and the regulatory cost of saying yes has grown.

Behind the public narrative of US AI dominance sits an awkward physical fact: China supplies more than 40% of US battery imports and roughly 30% of transformer, switchgear, and grid-support imports. Every tariff increase on Chinese industrial goods through 2025 translated directly into longer hyperscaler procurement queues — and into higher final per-megawatt costs. Domestic substitution is happening, but slowly. New US transformer plants announced in 2023-2024 will not reach full output before 2027. The same is true for grain-oriented electrical steel, which is dominated by Japanese and Indian producers. For the AI build-out, this means the 2026-2027 capacity gap is essentially locked in — no amount of additional capital can compress lead times that are physically constrained by global metallurgy.

Even with half the announced US capacity falling away, Alphabet, Amazon, Meta, and Microsoft still plan to spend more than $650 billion combined on AI infrastructure in 2026 — a record. The apparent contradiction makes sense once you look at where the money is actually going: less new ground-up construction, more chip purchases, more existing-site retrofits, more international diversification, and more equity stakes in third-party data center developers like Digital Realty, Equinix, and CoreWeave. For investors, this is a profitability story disguised as a build-out story. Returns on AI capex now depend less on raw gigawatts deployed and more on access to scarce, energized capacity. Operators with already-permitted, already-energized sites command a premium; greenfield announcements are increasingly discounted by the market. Expect the spread between operating AI-grade capacity and announced capex to widen significantly in 2026-2027 — a structural feature, not a temporary disconnect.