Energy Is the Next AI Bottleneck

Why the AI Infrastructure Boom Is Becoming an Electricity Race

Part 6 of the Growth Solutions KC AI Infrastructure Series

For the first phase of the AI boom, the question was simple:

Who has the chips?

Then the bottlenecks multiplied.

Who has the high‑bandwidth memory?
Who has the advanced packaging?
Who has the servers, networking, cooling, and data‑center capacity?

Now another constraint is moving toward the center of the AI infrastructure story:

Who has the power?

A data center can be filled with the most advanced GPUs in the world, but without reliable electricity those chips are little more than expensive silicon. That is why the next stage of the AI buildout will not be determined by semiconductors alone.

It will also be shaped by generation, transmission, substations, grid interconnections, natural gas, nuclear power, renewable energy, storage, cooling systems, permitting, and the communities where this infrastructure is built.

If chips are the brain of the AI economy, electricity is the bloodstream.

The next bottleneck is power. And unlike a chip shortage, the power problem cannot be solved by simply placing a larger order.

What Matters: AI Has Reached the Grid

For decades, U.S. electricity demand was relatively stable. That is changing.

The Department of Energy reported that data centers consumed roughly 4.4% of U.S. electricity in 2023 and previously projected that share could reach 6.7% to 12% by 2028. More recent federal analysis continues to identify data centers and AI as major drivers of future electricity growth. The Energy Information Administration now describes data‑center load as an emerging dominant driver of long‑term U.S. electricity demand.

Globally, the trend is similar.

The International Energy Agency projects electricity consumption from data centers could roughly double by 2030 to around 945 terawatt‑hours in its base case — far outpacing overall electricity‑demand growth.

These estimates will evolve. AI models may become more efficient. Chip performance per watt will improve. Cooling technologies will advance. Some data‑center projects may never be built.

But the direction is increasingly difficult to ignore. AI is turning computing growth into electricity growth. And electricity operates under very different constraints than software.

Why Power Is a Different Kind of Bottleneck

Software can scale almost instantly. Electric infrastructure cannot. A new AI model can be released overnight. A large power plant cannot. A hyperscaler can order another generation of servers. It cannot instantly create transmission lines, substations, transformers, generating capacity, or a grid connection.

That physical reality creates a mismatch.

AI demand moves at technology speed. Energy infrastructure moves at industrial speed.

New power generation requires financing, equipment, fuel, construction, regulatory approvals, and often years of development. Transmission projects require coordination across multiple jurisdictions. Grid interconnection can become a lengthy process. Large transformers and electrical equipment have their own supply constraints.

And data centers increasingly require power not merely in large quantities, but with extraordinary reliability. AI workloads do not simply need electricity. They need electricity where the data center is located, when the data center needs it, at the scale required, with reliability strong enough to protect billions of dollars of computing infrastructure.

That is a much harder problem.

From the GPU Shortage to the Megawatt Shortage

The early AI infrastructure race centered on compute. Companies competed for GPUs. Then HBM emerged as a critical constraint because processors cannot operate efficiently without enough memory bandwidth. Advanced packaging became another choke point because modern AI systems require tightly integrated compute and memory.

Power is the next layer of the same story. Every bottleneck sits underneath another bottleneck.

• A GPU requires HBM.
• The GPU and HBM require advanced packaging.
• The packaged chips require servers and networking.
• The servers require data centers.
• The data centers require electricity.
• The electricity requires an entire energy system behind it.

The AI stack does not end at the server rack. It reaches into power plants, pipelines, substations, transmission systems, utilities, nuclear reactors, batteries, cooling infrastructure, and local permitting authorities.

AI infrastructure is becoming increasingly interconnected.

There Will Not Be One AI Energy Source

One of the easiest mistakes is assuming the AI power problem has a single solution. It almost certainly does not.

Natural gas offers dispatchable generation and, in some locations, the ability to build power close to data‑center demand.

Nuclear offers reliable, high‑capacity, carbon‑free electricity and is attracting renewed attention from hyperscalers.

Wind and solar can add substantial generation but often require storage, transmission, or complementary generation to meet around‑the‑clock loads.

Battery storage can help balance supply and demand but is not itself a source of generation.

Existing grids remain essential because data centers rarely operate as isolated islands.

And new technologies — including small modular reactors, geothermal systems, long‑duration storage, and eventually perhaps fusion — may become increasingly important.

The future AI energy system is likely to be plural.

Gas. Nuclear. Renewables. Storage. Grid power. Behind‑the‑meter generation. Efficiency.

The mix will differ by geography, economics, policy, reliability requirements, and speed. For AI infrastructure developers, the question is becoming less ideological and more practical: What combination can deliver dependable power fast enough to keep the compute online?

Big Tech Is Becoming a Major Energy Buyer

Technology companies are not becoming traditional utilities. But energy procurement is becoming part of technology strategy.

Microsoft entered a 20‑year power‑purchase agreement supporting the planned restart of the former Three Mile Island Unit 1, now the Crane Clean Energy Center.

Meta later signed a 20‑year agreement supporting 1,121 megawatts from Constellation’s Clinton nuclear facility in Illinois.

Google has pursued multiple nuclear strategies, including an agreement with Kairos Power for advanced small modular reactors and a separate arrangement supporting the planned restart of Iowa’s Duane Arnold nuclear plant, which could provide more than 600 megawatts to the regional grid.

Amazon has also invested heavily in nuclear technologies, including small modular reactor development.

These moves are significant.

For decades, technology strategy focused on chips, software, networks, and data centers. Now energy contracts are becoming part of competitive strategy.

• Power availability can determine where a data center is built.
• Power cost can influence operating economics.
• Power reliability can affect utilization.
• And speed‑to‑power can determine how quickly billions of dollars of AI hardware begin generating revenue.

Energy strategy is becoming part of AI strategy.

When the Grid Is Too Slow

Another development may be even more revealing. Some data‑center projects are exploring behind‑the‑meter generation — building or securing dedicated power rather than waiting entirely for traditional grid connections.

A recent example is Amazon’s planned Pecos County, Texas, data‑center campus. The project is expected to use new on‑site generation, with Amazon paying the full cost of powering its operations while transitioning toward grid‑connected service as interconnection timelines allow. Reports describe a proposed natural‑gas generation complex capable of reaching several gigawatts if fully developed.

The project illustrates the tradeoffs. Dedicated generation can accelerate access to power and potentially shield existing ratepayers from some infrastructure costs. But large gas‑fired projects raise legitimate questions about emissions, long‑lived fossil‑fuel infrastructure, local environmental impacts, and how corporate climate commitments interact with the urgency of AI expansion.

That tension is important.

The AI power challenge is not simply: How do we generate more electricity?

It is: How do we generate enough reliable electricity quickly, affordably, responsibly, and without shifting unreasonable costs onto everyone else?

Who Pays for the AI Power Boom?

This may become one of the most important public‑policy questions surrounding data centers.

New electricity demand can require new generation.
New generation can require transmission.
Transmission can require substations and grid upgrades.

Those projects cost money. The question is who should pay.

In 2026, the federal government launched a Ratepayer Protection Pledge built around the principle that large AI and data‑center companies should bear the cost of the electricity generation and infrastructure required for their facilities rather than shifting those costs to ordinary households and businesses.

The underlying principle deserves attention regardless of politics.

Data centers can create jobs, tax revenue, infrastructure investment, and economic activity. They may also create enormous new electricity loads. A healthy system should encourage investment while aligning costs and benefits fairly.

That means asking hard questions:

• Who pays for new generation?
• Who pays for transmission?
• Who carries the risk if projected data‑center demand does not materialize?
• Should existing customers subsidize infrastructure built primarily for one large corporate user?
• Can data‑center investment help finance grid improvements that benefit the broader community?

These are no longer theoretical questions. They are becoming part of the economics of AI.

What This Means for Investors

The power bottleneck broadens the AI infrastructure opportunity again. The first investment wave centered heavily on semiconductors. The next wave may increasingly involve the companies that help electricity reach the rack.

• Utilities.
• Independent power producers.
• Natural‑gas infrastructure.
• Nuclear operators.
• Electrical‑equipment manufacturers.
• Transmission and grid technology.
• Transformers and substations.
• Backup and distributed generation.
• Energy storage.
• Cooling systems.
• Data‑center infrastructure.

But the same discipline from Part 5 still applies.

A powerful theme does not automatically make every company inside that theme a good investment.

• Utilities face regulatory constraints.
• Power producers face commodity cycles.
• Nuclear projects carry execution and permitting risk.
• Infrastructure companies can become overvalued.
• Large generation projects can be canceled or delayed.
• Data‑center forecasts can prove too optimistic.
• Technology can improve faster than expected.

Investors should therefore look beyond the headline.

Which companies control scarce, difficult‑to‑replicate infrastructure that AI demand increasingly requires — and can they convert that position into durable cash flow?

That is the same investor lens applied to a different bottleneck.

What Could Break the Power Thesis?

The energy thesis should not be treated as inevitable.

• AI efficiency could improve substantially.
• Future chips may process more work using less electricity.
• Models could become less computationally intensive.
• Workloads could migrate toward smaller models or edge devices.
• Hyperscaler capex could slow.
• Data‑center projects could be canceled.
• New generation could arrive faster than expected.
• Grid reforms could unlock capacity currently trapped by inefficient processes.

These developments could reduce pressure. But efficiency alone does not guarantee lower total electricity demand. If AI becomes cheaper and more useful, people may simply use much more of it.

The central variable is therefore not merely energy consumed per AI task. It is the interaction between efficiency and total usage.

That relationship will help determine whether today’s power shortage becomes a temporary constraint or a defining infrastructure cycle.

What to Watch Next

Several signals will shape the energy thesis.

Data‑center electricity forecasts. Are projected loads continuing to rise?

Hyperscaler power agreements. Are companies securing more nuclear, gas, renewable, storage, or behind‑the‑meter generation?

Grid‑interconnection timelines. Are projects connecting faster or slowing down?

Utility capital spending. Are utilities expanding generation and transmission?

Nuclear milestones. Do restarts and small modular reactors move from announcements to operating assets?

Natural‑gas infrastructure. Does gas become the principal near‑term bridge for AI load growth?

Rate structures and regulation. Who ultimately pays for the buildout?

Community response. Are local governments welcoming or resisting data‑center expansion?

These signals will determine whether energy becomes merely another input into AI infrastructure — or its most important limiting factor.

What Happens Next

The first phase of the AI revolution was about models. The second phase became about chips. Then came memory, packaging, networking, servers, and data centers. Now the story is reaching the electrical grid.

That does not mean semiconductors matter less. It means the infrastructure stack is expanding. And the deeper the stack becomes, the clearer the central lesson of this series becomes:

Artificial intelligence is not weightless.

It requires enormous amounts of physical infrastructure. The AI race is becoming an energy race.

The companies and nations that can build that infrastructure — and the communities that can support it — will have an advantage.

And that opens another set of questions we will explore next:

• Who pays for the power boom?
• Can nuclear scale fast enough?
• What role will natural gas play?
• Can small modular reactors move from promise to reality?
• Can America’s grid and transmission system keep up?
• How should policymakers balance growth, affordability, reliability, and local impact?

Because the next AI bottleneck is not hidden inside a chip. It is flowing through the wires. If chips are the brain of the AI economy, electricity is the bloodstream.

Without enough of it, the AI infrastructure revolution stops.

— Matt Cucinotta | Growth Solutions KC | Inspire · Inform · Ignite

Source and methodology note: This article draws on current analysis and reporting from the U.S. Department of Energy, U.S. Energy Information Administration, International Energy Agency, company announcements, and energy‑industry sources. Forecasts for future data‑center electricity demand vary materially and should be treated as scenarios rather than guarantees. Company projects discussed remain subject to development, permitting, regulatory, financing, and execution risk.