The First AI Bottleneck Isn’t the Chip - It’s Power

Chips still make the headlines...

Power is beginning to make the decisions.

For the past several years, the AI race has been measured in processors, parameters and computing capacity. Companies competed to secure GPUs. Investors followed semiconductor demand. Every new model announcement seemed to require another warehouse filled with increasingly powerful hardware.

That race is not over.

But ordering thousands of processors and turning them on are two very different things.

The market is discovering that computing capacity cannot grow faster than the physical systems supplying it. A data center may have the capital, customers, chips and land required to move forward. If the electricity is not available at the right location, at the right price and on the right date, the project is not ready.

That is the signal inside the noise.

The first AI bottleneck is no longer only the chip.

It Is Power.

The Number Worth Knowing

A single planned data center can require between 100 and 1,000 megawatts of electricity, approximately the demand of 80,000 to 800,000 average homes, according to the Electric Power Research Institute. EPRI

Sit with that number for a second.

We are not talking about adding a few server racks to an office building. The largest AI campuses are introducing industrial-scale loads into power systems that were planned years before generative AI entered the conversation.

EPRI now projects that data centers could consume between 9% and 17% of U.S. electricity by 2030. Berkeley Lab’s latest range lands in a similar neighborhood: approximately 9.5% to 15.3%, with a central estimate of 11.8%. Berkeley Lab

Globally, the International Energy Agency expects data center electricity consumption to rise from approximately 485 terawatt-hours in 2025 to 950 terawatt-hours by 2030. Electricity use from AI-focused facilities is projected to triple during that period. IEA

That is a lot of new demand arriving very quickly.

And power infrastructure is not known for moving quickly.

What the Market Is Missing

The conversation is often framed as if the country simply needs to generate more electricity.

It does.

But total generation is only one part of the problem.

A gigawatt of electricity somewhere on the grid does not automatically help a data center that needs that gigawatt at a specific site. The power must be generated, transmitted, stepped down, distributed and delivered through equipment capable of handling the load.

That requires transmission lines, substations, transformers, switchgear, utility approvals, engineering studies and interconnection agreements.

It also requires time.

This is why national power forecasts can look manageable while individual projects remain stuck. Electricity may exist within the broader system without being deliverable to the specific parcel where a developer wants to build.

The Better Question Is Not...

Is there enough electricity?

It is:

How much firm power is available at this location and when can it be delivered?

That is a very different infrastructure question.

Power Is a Stack, Not a Plug

AI infrastructure is not one industry.

It is a system of interconnected industries, assets and decisions. Each pillar affects the others, and weakness in one can delay or completely stop a project.

When someone says a site “has power,” I immediately want to know what they mean...

• Does the region have generation capacity?

• Is sufficient transmission available?

• Can the local substation support the load?

• Have the necessary transformers and switchgear been ordered?

• Has the utility committed to a service date?

• Does the project have backup or on-site generation?

• Can the system support the next phase of expansion or only the first building?

Power availability is not a yes-or-no question. It is a stack of infrastructure decisions, and every layer must work.

Generation

New AI demand will require a mixture of generation sources.

Renewables will continue growing because they can be deployed relatively quickly and help companies meet clean-energy goals. Natural gas is returning to the center of the conversation because it can provide scalable, dispatchable power. Nuclear is attracting renewed interest because it offers large amounts of around-the-clock electricity. Battery storage, microgrids and fuel cells are becoming increasingly important for flexibility and resilience.

There is no single technology capable of carrying the entire load.

The winning strategy will be a portfolio.

Transmission and Distribution

Generating electricity is only useful if it can reach the customer.

Transmission projects can take years to plan, permit and construct. Substations must be expanded. Distribution equipment must be secured. Interconnection studies must determine whether the existing system can handle the new load without creating reliability problems elsewhere.

The Department of Energy’s 2026 draft National Transmission Needs Study notes that many resources capable of reducing costs and supporting reliability remain caught in backlogged interconnection queues. U.S. Department of Energy

In other words, some of the generation needed to serve new AI demand is waiting for grid access too.

On-Site Power

The wait for traditional utility service is pushing developers toward behind-the-meter generation, fuel cells, microgrids and other on-site solutions.

This does not necessarily mean data centers are leaving the grid. Many still want a grid connection for reliability, backup power and long-term growth.

But developers are no longer assuming the grid will be the only answer.

Eaton and Siemens Energy, for example, are collaborating on standardized on-site power systems designed to accelerate data center deployment. Their pitch is straightforward: integrate power generation and electrical infrastructure earlier so capacity can come online faster. Eaton

The data center is starting to look less like a building connected to the energy system and more like an energy system with a building attached.

The Market Is Already Responding

Watch what the technology companies are buying.

Meta signed a 20-year agreement with Constellation Energy for the output of the Clinton Clean Energy Center in Illinois. The agreement covers 1,121 megawatts of nuclear generation beginning in 2027. Constellation Energy

Microsoft entered its own 20-year agreement supporting the restart of the Crane Clean Energy Center in Pennsylvania. Constellation Energy

These companies are not simply buying electricity.

They are buying certainty.

They want to know that large amounts of reliable power will be available throughout the operating life of their infrastructure. That certainty allows them to plan data center capacity, secure equipment and commit capital with greater confidence.

The power contract is becoming part of the AI strategy.

Follow the Equipment Order

If you want to know which AI projects are becoming real, do not look only at the data center announcement.

Look for the second layer of commitments:

• The turbine reservation.

• The transformer order.

• The substation expansion.

• The utility filing.

• The engineering contract.

• The switchgear manufacturing investment.

GE Vernova reported more than $2 billion in direct data center electrification orders during 2025, more than triple its 2024 volume. The company expects hyperscale data centers to represent approximately 25% of its 2026 orders, up from roughly 10% in 2024. GE Vernova

Eaton announced a new Nebraska manufacturing facility to increase production of medium-voltage switchgear used in data centers, utilities and industrial power systems. Eaton

These may sound like equipment stories.

They are not.

They are early indicators of where physical AI capacity is being prepared.

The Eight-Pillar Connection

Power may be the first constraint, but it does not operate alone.

It changes the value of land because a parcel with a credible utility service date is worth more than acreage waiting for infrastructure.

It changes capital because every month of delay adds carrying costs and pushes revenue further into the future.

It changes policy because regulators must decide who pays for generation, transmission and grid upgrades and how existing utility customers will be protected.

It affects water because both electricity generation and data center cooling depend on local resource conditions.

It affects workforce because the buildout requires engineers, lineworkers, electricians, equipment manufacturers and skilled construction teams.

It affects sustainability because the type of generation used will influence emissions, resilience and community acceptance for decades.

Even connectivity becomes irrelevant if the facility at the end of the fiber cannot be energized.

Power is one pillar of AI infrastructure.

Right now, it is also the gate through which the other seven must pass.

Company Watch:

Everyone is watching NVIDIA ship the processors.

I am watching GE Vernova help build the electrical system required to turn them on.

The company sits across several critical layers of the power stack: generation, gas turbines, grid equipment, electrical systems, consulting and project services. That gives it visibility into where utilities, independent power producers and hyperscalers are preparing for new demand.

GE Vernova is not the only company positioned here.

Eaton, Siemens Energy, Schneider Electric, ABB and Hitachi Energy are supplying the electrical equipment behind the buildout. Quanta Services, Burns & McDonnell and Black & Veatch are helping engineer and construct the grid. Constellation Energy, Vistra, NextEra Energy, Qcells and Bloom Energy are approaching the opportunity from different parts of the generation and energy-solutions market.

These are not companies operating around the edge of the AI story.

They are becoming part of the AI supply chain.

Questions Worth Asking

When a new AI campus is announced, here is what leaders should start asking:

• How much firm power has actually been committed?

• Which utility or generation partner will provide it?

• What is the confirmed service date?

• What transmission and substation upgrades are required?

• Has the necessary long-lead electrical equipment been reserved?

• Will the project use on-site or behind-the-meter generation?

• Who will pay for the grid improvements?

• How will existing utility customers be protected?

• What energy sources will support around-the-clock operations?

• Can the power system support future phases of the campus?

Those answers will tell you far more than the size of the land purchase or the value of the announced investment.

Golden Nugget

An announced megawatt is not an available megawatt.

The number that matters is the utility service date backed by an engineering plan, equipment commitments and a credible path to delivery.

That is the difference between power on paper and power a data center can actually use.

The IOI Take

For years, technology companies could treat electricity as something that appeared when a facility was connected to the grid.

That assumption is disappearing.

Power is becoming a location strategy, a capital strategy, a partnership strategy and a competitive advantage. The companies securing it early will build faster. The communities planning for it intelligently will attract more credible projects. The utilities capable of providing certainty will become some of the most influential economic-development organizations in the country.

This does not mean every proposed data center will be built or that every aggressive load forecast will materialize exactly as announced.

It means the market has reached a point where electricity can determine which projects survive the planning process.

That is a major shift.

One Last Thought

The chip race is easy to see.

It arrives through product launches, earnings calls and photographs of expensive processors.

The power race is quieter.

It is taking place inside utility resource plans, interconnection studies, equipment factories, regulatory filings and long-term energy agreements.

But quiet does not mean unimportant.

If you want to understand where the next generation of AI infrastructure will be built, keep watching the chips.

Then look past them.

The real map is being drawn by pow

Workforce

Policy

Capital

Sustainability