All research

Energy · Pillar guide · 2026-06-25

AI's energy bottleneck: the race for delivered power

A GPU reservation is only a promise. Capacity becomes real when turbine slots, transformers, interconnection studies, switchgear, cooling loops and crews converge at the same site.

, Founder and publisher, AI Bottlenecks

The AI buildout keeps producing numbers too large to feel physical. One gigawatt here, five gigawatts there, a campus larger than a city load, a rack that wants more power than a small office floor. The announcements become easier to understand when you walk the chain backward. A gigawatt is a turbine reservation, a gas lateral, a transformer core, a substation bay, a protection scheme, a queue position, a cooling plant, a crew schedule and a utility planner willing to sign an energization date.

That date is the whole game. A GPU can be ordered in one fiscal year and delivered in the next. A data center can be financed before the steel is in the ground. The power system lives on a different clock. The grid has to study the load, connect it, protect it, serve it during summer peak and survive the failure modes created by the load itself.

The useful question is not whether the world can generate enough electrons. It is where firm megawatts can be delivered, at what node, through which equipment and by which year. The constraint has moved from the chip order to the energized rack.

What is the AI data center power bottleneck?

The bottleneck is the conversion of announced AI capacity into usable, reliable electrical load. It starts outside the building with generation and grid interconnection. It continues through transformers, switchgear, substations and transmission upgrades. It ends inside the building, where high-density racks force new power-distribution and cooling designs. Every layer can stop the project. Every layer has its own supplier base, lead time and approval process.

The scale is real. The International Energy Agency estimates global data-center electricity consumption at about 415 TWh in 2024 and projects roughly 945 TWh by 2030 in its base case. The IEA also expects the United States and China to account for nearly 80 percent of global data-center electricity-consumption growth through 2030. In the United States, a Department of Energy report backed by Lawrence Berkeley National Laboratory estimated that data centers used about 4.4 percent of U.S. electricity in 2023 and could reach roughly 6.7 percent to 12 percent by 2028.

Those percentages can sound manageable at global scale. The local version is what bites. Data centers cluster near fiber, land, tax treatment, customers and power markets. Northern Virginia, ERCOT, PJM, Dublin, London, Frankfurt, Malaysia, Tokyo and parts of the Middle East do not experience global electricity demand. They experience a queue of specific campuses asking for specific blocks of firm load.

In an April 2026 update, ERCOT said it was tracking about 410 GW of large loads seeking interconnection, with roughly 87 percent of that load coming from data centers. That does not mean 410 GW will be built. It means the planning system has been flooded by a class of demand that looks more like industrial siting than ordinary load growth. The request itself becomes a bottleneck.

The chain from fuel to rack

A modern AI campus is a chain of dependencies. The fuel has to exist. The generation equipment has to be available. Transformers have to change voltage. The interconnection study has to clear. The transmission system has to absorb the new load. The substation has to be built. The campus electrical room has to protect and distribute power. The cooling system has to remove the heat. The rack architecture has to survive density that keeps climbing.

Generation supplies energy. Transmission and substations move it. Transformers change voltage. Switchgear protects and routes it. Campus distribution carries it to power electronics, cooling and racks. Field crews install, test and commission the complete system.

A single missing part can freeze the whole project. That is why delivered power is a better bottleneck than a slogan. It has parts.

Gate one: gas turbines and firm power

AI load is unforgiving. Training clusters, inference campuses and agentic workloads cannot be treated like a flexible home charger. They need firm power, fast response and high reliability. Renewables and batteries help, especially where the location is flexible, but many campuses still need dispatchable generation behind or near the meter.

The heavy-duty gas-turbine market is narrow. GE Vernova, Siemens Energy and Mitsubishi Heavy Industries sit at the center of it. Their reservation books have become a calendar for AI power.

GE Vernova said in Q1 2026 that demand was accelerating across Power and Electrification, backlog grew by more than $13 billion quarter over quarter, and it expected at least 110 GW of combined gas-turbine backlog and slot reservations by the end of 2026. Its Electrification segment booked $2.4 billion of data-center equipment orders in the same quarter, more than all of 2025.

Siemens Energy reported a similar pattern in Q1 FY2026. Gas Services booked 102 gas turbines and converted 12 GW from reservation agreements while adding another 12 GW of new reservations. Grid Technologies also reported strong order growth, including several U.S. data-center-related orders in the high triple-digit-million-euro range.

The point is not that gas is the only answer. It is that dispatchable generation equipment is now part of the AI supply chain. A turbine slot can gate a campus the way a packaging slot can gate a GPU ramp.

Gate two: transformers, GOES and voltage conversion

Transformers are the least glamorous object in the story and possibly the most painful one. Power has to move at one voltage and be used at another. Every campus needs voltage conversion. Every new generation project needs it too. Transformers sit at the crossing point between generation, transmission, distribution and load, so demand arrives from AI data centers, renewables, grid hardening, electrification, manufacturing and ordinary replacement at the same time.

Wood Mackenzie put average transformer lead times around 120 weeks in 2024, up from about 50 weeks in 2021. It estimated a wider range of roughly 80 to 210 weeks for large substation and generator step-up transformers. The Department of Energy's large-power-transformer work points to the material choke underneath the equipment choke: grain-oriented electrical steel, the specialized steel used in transformer cores.

The transformer layer is best read as its own manufacturing system. Steel and copper matter, but so do winding capacity, insulation, drying, factory build slots, high-voltage test, transport and site engineering. A power plant without a generator step-up transformer is stranded. A data center without a substation transformer is a shell.

The dedicated transformer bottleneck deep dive follows that chain from material to energization.

Gate three: interconnection queues and tariff design

Interconnection is where electrical engineering becomes queue theory. A grid operator has to study how a new load changes power flows, fault current, voltage stability, protection settings and upgrade requirements. The study has to decide who pays for network work and what conditions apply before service begins.

Large-load requests now arrive faster and at larger scale than many planning processes were designed to handle. A queue can contain speculative projects, duplicate requests and real campuses competing for the same local capacity. Reform has to remove weak requests without letting a very large customer shift unacceptable cost or reliability risk to other users.

FERC opened targeted action around large-load integration in June 2026 to push regional operators toward clearer connection rules. The policy is part of the physical stack. A transformer and turbine can exist while the tariff, upgrade allocation or study remains unresolved.

Gate four: campus electrical equipment

The substation is not the end of the chain. Medium-voltage switchgear, breakers, busway, uninterruptible power systems, power shelves and protection controls have to carry power through the campus. Racks that climb toward hundreds of kilowatts change the size and shape of this equipment.

The campus layer rewards suppliers that can ship integrated systems, meet project schedules and commission them in the field. Backlog is useful evidence only when it converts into recognized revenue, accepted equipment and a working site. Late-stage project risk often appears in test windows, cable termination, controls integration and scarce electrical labor.

Gate five: power becomes heat

Almost every watt entering an AI rack leaves as heat. Higher rack density therefore ties the electrical and thermal systems together. Liquid cooling, pumps, cold plates, coolant-distribution units, chillers and heat rejection become part of the power-delivery question.

A campus can secure its utility service and still miss its compute schedule if the cooling design cannot support the selected rack. Cooling also consumes power, water, space and maintenance. The right metric is sustained useful compute under the expected environmental and service conditions, not a peak rack rating.

Behind the meter changes the route, not the hardware

When grid delivery is late or uncertain, developers can move generation on or near the campus side of the utility meter. That can shorten one queue, but it does not make the physical system disappear.

The project still needs turbines or engines, fuel delivery, transformers, switchgear, protection, permits, cooling and field crews. It may add batteries, grid-service controls and a more complicated operating agreement. The workaround reroutes the bottleneck through much of the same equipment.

800 VDC helps inside the building

Higher-voltage direct-current architectures can reduce conversion stages, current and copper bulk as rack power rises. They can make the electrical path inside the campus more efficient and compact.

They do not remove the upstream problem. Generation, interconnection, transformers and site delivery still have to exist. An architecture improvement can loosen one constraint and tighten another by shifting demand toward new power electronics, protection devices, connectors and service practices.

Nuclear belongs on the long clock

Nuclear can supply large blocks of firm, low-carbon power. Restarts, uprates and life extensions can matter. New reactor programs and small modular reactors may matter later.

The near-term bottleneck is less forgiving. Licensing, fuel, siting, supply chains, construction and public acceptance usually operate on a longer clock than the current campus buildout. Nuclear can be a destination without being the bridge that energizes the next wave.

What would prove the energy thesis wrong?

A useful thesis needs failure modes. The energy bottleneck weakens if data-center load forecasts roll over, transformer lead times compress because qualified output rises, turbine reservation books clear, interconnection reform makes queue position cheap, rack efficiency improves faster than campus scale, or hyperscalers accept a slower growth curve.

It also weakens if valuation outruns the physical cycle. A real bottleneck can still become a bad trade.

What to watch now

Watch lead times, not adjectives. Watch reservations, not ambition. Watch queue rules, not conference slogans. Watch whether data-center orders are appearing in the parts of the supply chain that actually touch power delivery.

The shape of it

The new scarce objects are old objects: turbines, transformers, steel, cables, breakers, substations, skilled crews, permits, cooling loops and utility studies. They do not care how fast the model roadmap moves.

This is not electricity in the abstract. It is a map of friction. Each layer has a physical object, a manufacturing base, a queue, a price signal and a calendar. The companies that turn paper megawatts into energized racks get paid before the model does anything useful with the power.

A GPU order says someone wants compute. A delivered megawatt says the factory can open.

Sources

This research is educational and is not investment advice.