English
Nearly half of all US AI data center capacity planned for 2026 has been delayed or canceled, creating a 7 gigawatt (GW) shortfall between announced capacity and active construction. According to multiple independent reports from JLL, Morgan Stanley, and industry analysts published in April 2026, grid-power constraints have overtaken semiconductor supply as the primary bottleneck for AI expansion, marking the defining infrastructure crisis of the year. Of the roughly 12 GW of data center capacity announced for 2026, only about 5 GW is under active construction, with the remainder stalled by electrical infrastructure shortages, utility interconnection delays, and tariff impacts on critical components.
The New Bottleneck: From Silicon to Substations
For the past two years, the AI boom has been constrained by GPU availability and semiconductor fabrication capacity. That dynamic has now shifted. The four largest hyperscalers — Alphabet, Amazon, Meta, and Microsoft — have committed approximately $650 billion in AI infrastructure spending for 2026, the largest single-year corporate investment cycle in history. Yet capital is no longer the binding constraint. The bottleneck has moved to the physical layer: high-voltage transformers, switchgear, circuit breakers, and utility interconnection approvals.
Lead times for large power transformers have stretched to 128 weeks (2.5 years) on average, with generator step-up units reaching 144 weeks, according to FluxCo's February 2026 Transformer Lead Times Tracker. For specialized units, procurement timelines approach 36 to 48 months. These components represent less than 10% of total data center construction costs but now dictate the entire project timeline. The AI data center power crisis has made transformer supply the single most critical factor in project viability.
Meanwhile, interconnection queue wait times have ballooned. In PJM, the largest US wholesale electricity market, average interconnection timelines now exceed five years, with over 2,600 projects totaling 260+ GW stuck in the queue. FERC Order 2023 has helped modestly, but the sheer volume of generation and storage projects — over 2.2 terawatts in queues nationwide — means that only a fraction reach commercial operation. The average time from interconnection request to operation has risen from under two years in 2008 to nearly five years today.
The 7 GW Gap: By the Numbers
The scale of the disconnect between ambition and reality is stark. For 2026, 12 GW of data center capacity was announced across roughly 140 US projects. Of that, only 5 GW — about 42% — is actually under construction. The remaining 7 GW faces delays or outright cancellation. Looking ahead to 2027, the picture worsens: 21.5 GW has been announced, but only 6.3 GW has broken ground. Analysts warn that the actual energization rate for 2026 may end up closer to 20% by year-end.
This gap has profound implications for the global AI race. Hyperscalers are spending heavily on GPU clusters, custom chips, and networking equipment, but those investments cannot generate returns without energized data centers. The US electrical grid capacity constraints are effectively capping the pace of AI infrastructure deployment, regardless of how much capital is available.
Tariffs and Supply Chain Disruptions
Compounding the equipment shortage, the US imports over 40% of its high-power transformers from China. New 50% tariffs on copper — a core transformer component — have raised costs without creating immediate domestic alternatives. While about $2 billion in new North American transformer manufacturing capacity is scheduled to come online by 2026-2027, that is insufficient to close the gap. The US currently meets only about 20% of its large power transformer demand through domestic production.
Who Wins and Who Loses in the Grid-Constrained Era
The crisis is reshaping market power dynamics. Companies that control physical assets — energized power, interconnection rights, permitted sites, and electrical equipment — have become the pricing-makers in this market. Data center tenants, particularly AI cloud companies and smaller colocation providers, are becoming pricing-takers, facing rising costs and delays stretching into 2027-2028.
Hyperscalers with pre-negotiated supply agreements and dedicated utility relationships are partially insulated. Amazon, for instance, has secured power purchase agreements for multiple gigawatts of capacity, while Microsoft has signed a deal to restart Three Mile Island and invested in nuclear capacity. Meta's Prometheus AI campus and Ohio superclusters represent gigawatt-scale projects that move forward thanks to early grid engagement. But smaller players face existential pressure.
The data center interconnection queue crisis has created a secondary market for interconnection rights, with permitted sites trading at significant premiums. Some developers are now acquiring existing industrial facilities specifically for their grid connections, a strategy known as 'brownfield data center development.'
Implications for Energy Markets and the AI Race
The grid bottleneck has direct consequences for US competitiveness in AI. While China continues to build out its domestic AI infrastructure with state-backed grid investments, US projects face fragmented utility approval processes and local opposition. Between May 2024 and June 2025, billions of dollars in data center projects were halted or delayed by community resistance movements in the US, particularly in Virginia, Ohio, and Arizona.
Global data center electricity consumption reached approximately 415 TWh in 2024, about 1.5% of worldwide demand, and the IEA projects it could double by 2030. JLL's 2026 Global Data Center Outlook describes the industry entering a potential $3 trillion infrastructure supercycle, with global capacity expected to nearly double to 200 GW by 2030. But that projection depends on energy infrastructure expansion keeping pace — a premise that the 7 GW gap calls into question.
The AI infrastructure spending ROI debate is intensifying. Morgan Stanley projects Amazon could post negative free cash flow of nearly $17 billion in 2026. Investors have begun punishing companies with murky ROI timelines while rewarding those showing clear revenue inflection. The bull case argues that compute demand remains genuinely supply-constrained, while the bear case warns of dot-com style overbuilding.
Expert Perspectives
'The bottleneck has shifted from capital to physical assets,' said a senior analyst at JLL who worked on the 2026 Global Data Center Outlook. 'You can throw billions of dollars at this problem, but if you can't get a transformer delivered or a utility interconnection approved, that money sits idle. The companies that control the physical layer — utilities, permitted site owners, electrical OEMs — are now the pricing-makers in this market.'
A Morgan Stanley infrastructure analyst added: 'We are seeing a structural transformation. The 7 GW gap is not a temporary blip; it reflects a fundamental mismatch between the pace of AI infrastructure demand and the capacity of the electrical grid to support it. This will take years to resolve, even with accelerated manufacturing investments.'
FAQ
What is the 7 GW gap in US data center capacity?
The 7 GW gap refers to the difference between the 12 GW of AI data center capacity announced for 2026 and the roughly 5 GW actually under active construction. The remaining 7 GW has been delayed or canceled due to grid infrastructure constraints.
Why are data center projects being delayed in 2026?
The primary cause is electrical infrastructure bottlenecks: high-voltage transformers and switchgear face 36-48 month lead times, utility interconnection approvals take up to five years, and tariffs on Chinese-made components have raised costs. GPU supply is no longer the main constraint.
How much are hyperscalers spending on AI infrastructure in 2026?
The four largest hyperscalers — Alphabet, Amazon, Meta, and Microsoft — have committed approximately $650 billion in combined AI infrastructure spending for 2026, the largest single-year corporate investment cycle in history.
What is being done to address the transformer shortage?
About $2 billion in new North American transformer manufacturing capacity is scheduled to come online by 2026-2027. Some developers are using refurbished units (1-6 week lead times) or broker inventory pools, but these are stopgap measures. Policy efforts include tariff adjustments and FERC interconnection reforms.
How does the grid bottleneck affect the global AI race?
The US grid constraints are capping the pace of AI infrastructure deployment, potentially ceding ground to China, which has state-backed grid investments. The gap between announced and constructed projects is widening, threatening US competitiveness in AI.
Conclusion: The Infrastructure Supercycle Hits Reality
The 7 GW gap represents a critical inflection point for the AI industry. After years of focusing on chip supply and software capabilities, the bottleneck has shifted to the physical world of transformers, substations, and utility poles. The $650 billion hyperscaler commitment is real, but converting that capital into energized megawatts will require years of grid investment, regulatory reform, and manufacturing expansion. For 2026, the defining infrastructure crisis is not about silicon — it's about steel, copper, and the grid that powers it all.
Follow Discussion