The AI Bottleneck Nobody Is Talking About: Transformers
On paper, the math is simple. The U.S. needs more power for AI. The grid needs to deliver it. Money is available. Everyone agrees on the goal.
But the electricity isn't moving. And the reason has nothing to do with power plants.
The bottleneck isn't generation. It's a piece of equipment that most people have never heard of: the transformer. The U.S. is running out of them. Without transformers, all the AI chips in the world cannot be powered.
The Aging Backbone
Transformers are the physical backbone of the electrical grid. They step voltage up for long-distance transmission and down for local distribution. Every data center needs them. Every substation depends on them. They are the connective tissue between power generation and power consumption.
The average U.S. substation transformer is over 40 years old. Many were installed in the 1960s and 1970s. They were designed for a different era—one with stable demand, predictable growth, and no AI.
In 2020 and 2021, a large power transformer took about a year to deliver. By the first quarter of 2026, generator step-up transformer lead times had surpassed 160 weeks. That's more than three years. For some of the largest units, delivery windows have stretched too. Once 24 to 30 months, they now run three to five years.
Some utilities are now ordering equipment five years in advance. Roseville Electric Utility in California used to procure equipment about a year out. Now it's buying transformers for projects it knows are coming half a decade away. Wood Mackenzie senior analyst Ben Boucher noted: "Equipment availability is becoming the biggest concern for developers as they value time to market so highly".
The Numbers That Explain the Crisis
The scale of the problem is substantial.
By 2030, U.S. data center capacity is expected to reach 110 gigawatts, up from roughly 24 gigawatts today. That is an increase of more than 350%. The sector's share of the electrical equipment market is projected to jump. It goes from just under 2% in 2020 to as much as 40%. The jump happens by the end of the decade.
Transformer demand alone could exceed 9,000 units annually by 2030, up from roughly 1,500 today.
But supply is not keeping pace. Global power transformer supply is running at a 30% deficit. New manufacturing capacity—where it exists at all—won't come online until 2028 at the earliest.
The result is a bottleneck that is already delaying projects. According to Sightline Climate data, the U.S. planned to add about 12 gigawatts of data center capacity in 2026. Only about one-third is under active construction. Nearly half of planned U.S. data center projects are being delayed or canceled. The primary reason is not capital, not permits, not labor. It is electrical equipment—transformers, switchgear, and batteries.
The Interconnection Nightmare
Even when transformers are available, getting power to a data center is a years-long ordeal.
Grid interconnection queues in major U.S. hubs now stretch four to seven years. In some cases, connecting new facilities to the grid can take five to ten years. As one analysis put it: "You can build a data center in two to three years, but it might take at least seven years to build new or upgraded transmission facilities".
The backlog is enormous. Roughly 2.3 terawatts of generation and storage capacity are currently sitting in U.S. interconnection queues—projects waiting for permission to plug into the grid. The time to develop new capacity has stretched in many regions. It went from about two years historically to five to seven years or longer.
Morgan Stanley recently warned that the U.S. faces a data center power deficit of 38 gigawatts between 2026 and 2028. The report noted that capital alone is no longer enough to secure projects.
Why This Is an AI Problem
The transformer shortage is not an energy crisis. It is an AI infrastructure crisis.
AI chips are useless without power. Power is useless without transmission. Transmission is useless without transformers. The entire chain collapses at the weakest link. Right now, that link is a piece of equipment. It takes three to five years to build. Developers are ordering it half a decade in advance.
The conventional narrative about AI infrastructure focuses on chips, data centers, and capital expenditure. But the physical reality is more mundane. You cannot train a frontier model on a transformer that hasn't been delivered yet. You cannot run inference on a data center that isn't connected to the grid.
The U.S. has plenty of electricity generation capacity in aggregate. The problem is not "not enough power." The problem is that power cannot reach where it is needed. The grid was not built for AI. Rebuilding it takes years—years that the AI industry does not have.
What This Means
The transformer shortage is a reminder that AI is not just a software problem. It is a physical infrastructure problem. The chips are the headline. The transformers are the reality.
Developers are responding by buying equipment years in advance, refurbishing older transformers, and sourcing from multiple suppliers. Some are even importing from China. Imports of high-power transformers jumped from fewer than 1,500 units in 2022 to more than 8,000 in 2025. But these are workarounds, not solutions.
The U.S. electrical equipment market for data centers is projected to grow to $65 billion by 2030. It was approximately $20 billion in 2026. But growth in spending does not guarantee growth in supply. Manufacturing capacity takes years to build. Skilled labor is scarce. Raw materials are constrained.
The AI industry has spent billions on chips. It may need to spend billions more on the equipment that delivers the power to run them.
Sources: U.S. Department of Energy; Wood Mackenzie analysis (2026); Sightline Climate / Bloomberg data (2026); Morgan Stanley report (August 20, 2026); Reuters (July 9, 2026); Data Center Knowledge (May 4, 2026); ColoradoBiz (July 9, 2026).
Disclaimer
The information provided in this article is for general informational and educational purposes only. It does not constitute legal, financial, or professional advice. The author and publisher are not responsible for any actions taken based on the content of this article. Readers should consult qualified professionals for advice specific to their situation. All trademarks and references to third-party products, services, or organizations are the property of their respective owners. The performance data and benchmarks discussed are based on specific research studies and may not generalize to all use cases or environments. As of the publication date, the AI landscape continues to evolve rapidly, and readers should verify current information independently.
Limitations
This analysis is based on reporting and public data available as of the article date; figures may be revised as sources update.
Forecasts from third-party analysts can change with market conditions.
Cost and pricing examples are point-in-time estimates; actual rates vary.
Country and company comparisons rely on public reporting, not operational data.
This sector moves fast; timelines and deal terms may be updated later.
Company deals and regulatory rulings may evolve; verify current status.
AI infrastructure is changing quickly; claims can become outdated soon.
Sources
- U.S. Department of Energy
- Wood Mackenzie analysis (2026)
- Sightline Climate / Bloomberg data (2026)
- Morgan Stanley report (August 20, 2026)
- Reuters (July 9, 2026)
- Data Center Knowledge (May 4, 2026)
- ColoradoBiz (July 9, 2026).
The information provided in this article is for general informational and educational purposes only. It does not constitute legal, financial, or professional advice. The author and publisher are not responsible for any actions taken based on the content of this article. Readers should consult qualified professionals for advice specific to their situation. All trademarks and references to third-party products, services, or organizations are the property of their respective owners. The performance data and benchmarks discussed are based on specific research studies and may not generalize to all use cases or environments. As of the publication date, the AI landscape continues to evolve rapidly, and readers should verify current information independently.
Limitations: This analysis is based on reporting and public data available as of the article date; figures may be revised as sources update.; Forecasts from third-party analysts can change with market conditions.; Cost and pricing examples are point-in-time estimates; actual rates vary.; Country and company comparisons rely on public reporting, not operational data.; This sector moves fast; timelines and deal terms may be updated later.; Company deals and regulatory rulings may evolve; verify current status.; AI infrastructure is changing quickly; claims can become outdated soon.