The AI boom's next bottleneck is the power grid's plumbing

For two years, the story of AI and electricity has been told through one machine: the gas turbine. Data centres need enormous amounts of power, power plants need turbines, and turbine makers have been sold out for years ahead. That story is still true. But on October 28, the biggest turbine maker in the world did something that suggests the shortage is moving.

GE Vernova reported that its gas-turbine backlog grew much more slowly than in the first half of the year. On the same day it agreed to spend $9 billion on a company that makes none of those turbines. Instead, Trench Group makes small but essential parts for high-voltage grid equipment. When the market leader puts that much money into the plumbing of the grid rather than into more power plants, it is worth asking why.

This article looks at the early signals that the AI power crunch is shifting from making electricity to moving and connecting it, and why that could matter well beyond the tech industry.

The turbine boom is slowing down, not ending

GE Vernova measures demand for its gas turbines in gigawatts of orders plus "slot reservations", where customers reserve a place in the production queue. That figure rose from 116 to 122 gigawatts in the third quarter of 2026. It is still growing, but the earlier quarters had seen double-digit gains, and management guided to slower growth in 2027.

Scarcity hasn't disappeared. According to Utility Dive, turbine prices this year are tracking about 50% higher, and customers can now reserve slots for delivery in 2032. About a fifth of the demand comes from data centres. But the growth curve is flattening.

The $9 billion deal points to where GE Vernova sees the next squeeze. Trench Group makes bushings, instrument transformers and coils. A bushing is the insulated passage that lets a high-voltage cable safely enter a transformer or a switch. It is not exciting, but without it nothing connects. Trench has about $5 billion of orders on its books, and its new plant in North Carolina will be the first US maker of the bushings needed for the very highest transmission voltages.

The shortage is spreading into the parts inside the transformer

The transformer shortage is not new. What is new is that companies are now investing specifically in the components inside transformers.

In September, Hitachi Energy announced a $528 million transformer factory in Mississippi, more than twice the size of its existing US plant. Construction starts in late 2026, and production is scheduled for 2029. A month later it committed 479 million Canadian dollars to a factory in Ontario that will make transformer insulation, the pressed board and laminates that keep the high-voltage parts from short-circuiting. It is due by 2028. Siemens Energy, another major supplier, reported a record €51 billion backlog in its grid business at the end of June, with orders up almost 28%.

Two things stand out. First, most of the new capacity arrives in 2028 and 2029, so the shortage looks structural for at least two more years. Second, the big manufacturers are now building supply chains for sub-components separately. That usually happens when those parts, not the finished product, have become the thing holding everything up.

America's biggest power market is short and can't respond

On the demand side, the clearest signal comes from PJM, the grid operator for a large part of the eastern United States.

Each year PJM runs an auction where power producers promise to be available in a future year, and get paid for that promise. In July, the auction for 2028/29 cleared at $325 per megawatt-day, the maximum the rules allow. It was the third year in a row that the auction hit its cap, compared with under $29 just a few years earlier, for 2024/25. Even at the cap, the auction came in 6,831 megawatts short of what PJM says it needs for reliability, and only about 525 megawatts of new generation and upgrades cleared. The total bill was $16.4 billion. PJM's report says the requirement rose largely because of new large electricity users.

When prices are capped and new supply still doesn't show up, the market is telling you something is physically stuck. The next auction, for 2029/30, is planned for December 2026.

The rules for "large loads" are being written now

Regulators are trying to decide what to do about customers who suddenly need as much power as a small city.

In October, the US federal energy regulator FERC accepted part of PJM's proposal: a formal definition of a "large load" and a registry to track them. It rejected two interim fixes, including one that would have let data centres bring their own new power plants as a condition of connecting, citing gaps in authority and justification. Another US grid operator, MISO, has filed rules that treat anything over 50 megawatts at one substation as a large load and create a separate tier for "computational load" above 25 megawatts of IT equipment. Those customers would need real-time monitoring, separate metering of any on-site generation, and constantly updated forecasts of their consumption. MISO asked for the rules to take effect on December 4, 2026.

These rules sound technical, but they create a quiet, recurring need at every large AI site: meters, monitoring systems and forecasting software. Rules that pass tend to stick, which makes this a more certain kind of demand than the headline-grabbing schemes.

Europe is starting to ration

In Europe the shift is even more direct: from first-come, first-served queues towards deciding politically who gets connected.

In the Netherlands, grid operators told the public broadcaster NOS in October that data centres account for about 15 gigawatts of connection requests, nearly a third of the entire waiting list. They asked the government to decide where and how many more data centres the country should accept, with conditions on energy, water and reporting. The Dutch Data Center Association pushed back, saying requests say little about what actually gets built. DC Market Insights calculated that 15 gigawatts is about 26 times what Dutch data centres consume on average today.

In the UK, the October 28 Budget introduced a "data centre covenant". Sites that want government help to get a grid connection must pass five tests covering energy costs, community investment, environmental standards, skilled jobs and national security, and agree a local benefit plan. The government's target is 6 gigawatts of AI data-centre capacity by 2030, against about 2 gigawatts today. Ireland already required new data centres to bring matching generation from late 2025.

For anyone building data centres, this changes the game. Where you can connect, and how little strain you put on the grid, may matter as much as how cheap the power is.

The contrarian signal from the research labs

The earliest signal in this story comes from two research papers, and it cuts against a popular idea.

The popular idea is that data centres can get connected faster if they are "flexible": pausing or shifting computing work when the grid is stressed. A paper posted on arXiv in October by Fatima and Li modelled a constrained grid connection and found something different. Making the data centre flexible added only about 2–5% more usable capacity. Changing how the utility switches and reconfigures its own network added about 42–45%. Doing both together raised the feasible data-centre size by roughly half.

A second paper, by Khanal, Dvorkin and colleagues, modelled the PJM region and found that moving computing work between locations cut system costs by 6% in 2028 and 19% in 2038, much more than simply switching work off at peak times.

Both papers are early and based on models, not real-world trials, and the first uses a single test network. But if they hold up, the overlooked opportunity may be grid automation software sold to utilities rather than flexibility tools sold to data centres.

A related lab-to-field step: in August, the New York Power Authority and the research institute EPRI tested a 1-megawatt solid-state transformer, an electronic alternative to the traditional heavy design, built at North Carolina State University. It ran at about 98% efficiency. A startup called Heron Power, which raised $140 million in February, plans a factory to start production in 2027. As the lab's lead researcher told MIT Technology Review, these devices cannot yet replace the largest high-voltage units.

What could go wrong with this story

• Turbine scarcity may simply be peaking. Part of GE Vernova's slowdown is reservations turning into firm orders, and analysts at William Blair noted no customers giving up their slots.
• Queues overstate real demand. The Dutch 15-gigawatt figure is requests, not buildings.
• The spending could pause. Alphabet, Microsoft and Meta spent about $51 billion, $45 billion and $35 billion respectively on capital projects in the September quarter. Fortune reported that Amazon's trailing free cash flow swung from about +$15 billion to about −$42 billion. If spending slows, today's equipment shortage could turn into spare capacity a few years later.
• The research is early. The flexibility papers are models, not field evidence.

What to watch next

• PJM's December auction for 2029/30: a fourth clear at the cap would confirm the market still cannot respond.
• MISO's large-load rules, requested to take effect on December 4, 2026, the first with a separate tier for computing loads.
• The Dutch government's data-centre plan, due this autumn, which could set national limits or conditions.
• The UK consultation on data-centre planning rules, open until January 8, 2027, which will define what counts as "strategic".
• The first commercial order for a solid-state transformer from a utility or a big tech company. It would mark the step from lab curiosity to real product.

The power plant was the obvious bottleneck. The quieter one is the copper, porcelain and paperwork that connects it to everything else, and that is where the money has started to move.