Build It Before They Do

Data Centers & Energy By EnergyInnoCore Editor · Published Oct 18, 2026 · 11 min read

I recently had a long conversation with a client who plans to build several gigawatt-scale data center campuses over the next three years. While raising capital for those projects, he and his team had run into a dilemma that I suspect many well-funded developers now share. On one side sits a highly politicized forecast of how much grid capacity will actually be available. On another sits a capital market that, at times, seems willing to fund almost anything labeled "AI infrastructure." And above both sits a group of confident executives, investors and political leaders who describe AI as a race that must be won at almost any cost.

The longer we talked, the more I came to believe that the most important question in AI infrastructure may no longer be technical. It may be psychological. This article looks at the four forces behind that dilemma (grid forecasts, capital, competitive fear and local opposition) and at what a more resilient approach to building might look like.

When "We Need More Compute" Becomes the Strategy

There is no question that AI is advancing quickly. Better models require more computing capability, and computing capability requires GPUs, networking equipment, buildings, cooling systems, transformers, substations, transmission capacity and, increasingly, dedicated power plants. That chain is easy to understand. What is harder to understand is what happens when the fear of falling behind starts to influence how much of it is built, how fast, and how much financial risk everyone is willing to accept.

For most of the technology industry's history, infrastructure followed demand. A company gained customers, usage grew, and more servers and facilities were added. AI is reversing that sequence. Companies are now building enormous amounts of capacity based not only on today's demand but on expectations about what future AI systems might need.

The scale is remarkable. OpenAI's Stargate initiative was announced with the intention of investing up to $500 billion in U.S. AI infrastructure. By September 2025, OpenAI, Oracle and SoftBank said Stargate projects represented nearly 7 GW of planned capacity and more than $400 billion of investment over three years.

In power-industry terms, those numbers are extraordinary. One gigawatt is not simply a large commercial customer; it is comparable to the output of a large conventional power station. Several such loads appearing in the same region within a few years of each other is something most U.S. utilities have never planned for. That is why AI infrastructure has become an energy story, and why the old assumption that the utility will simply deliver electricity when a customer asks for it is becoming unrealistic.

The U.S. Department of Energy's National Transmission Needs Study describes a shift from decades of relatively flat electricity demand toward rapid load growth, driven in part by hyperscale data centers and new industrial facilities. FERC has gone further, in 2026 directing the regional grid operators to justify or reform their rules for connecting very large electricity consumers. The problem is no longer theoretical: the grid was not designed for this.

The Grid Forecast Has Become Political

Utilities have always planned years ahead. They forecast population growth, industrial development, weather, plant retirements and demand, and use those forecasts to justify billions of dollars of investment in generation, substations and transmission. AI has made that process far more complicated.

A utility may receive requests representing several gigawatts of potential data center demand, but it cannot easily tell how much of it is real. A developer may be evaluating several sites at once. A hyperscaler may request capacity before making a final investment decision. Several developers may be competing for the same future tenant. If the utility assumes every request will be built, it may badly overestimate demand. If it assumes only a small share will be built, it risks underbuilding the grid. Neither mistake is cheap.

Once household electricity rates enter the conversation, the issue becomes political very quickly. Communities are asking a reasonable question: why should existing customers pay for transmission lines, substations or generation built mainly to serve very large data centers? Policymakers are responding with proposals to protect ordinary ratepayers, and the hard part is turning general promises into enforceable tariffs and cost-allocation rules. (I covered one federal attempt in our analysis of the Ratepayer Protection Act.)

This is where the AI race collides with local politics. A national leader may see a strategic asset, a governor may see economic development, a hyperscaler may see computing capacity and a utility may see load growth. A homeowner nearby may simply see a higher electricity bill, another transmission line, and a facility that uses a lot of power and water while employing relatively few people after construction. All of those perspectives can be true at once, which is exactly what makes the problem difficult.

Then There Is the Money

Perhaps the most striking feature of the current boom is the amount of capital flowing into it. The largest technology companies have effectively become some of the world's largest infrastructure investors.

• Microsoft said in April 2026 that it expected roughly $190 billion of capital expenditures in calendar 2026, while still expecting to be capacity-constrained through the year.
• Across Alphabet, Amazon, Meta, Microsoft and Oracle, Epoch AI estimated that capital spending has grown at roughly 70% per year since the release of GPT-4 in 2023.
• A Reuters analysis in August 2026 found that the same five companies had committed about $1.09 trillion in future lease payments, much of it tied to data center capacity that had not yet started operating.

A trillion dollars of forward commitments looks less like normal technology expansion and more like an industrial mobilization. The obvious question is whether future AI revenue will justify it. It may. AI could become one of the largest productivity platforms in economic history, transforming medicine, engineering, manufacturing, finance and education. If that happens, today's spending may eventually look conservative.

History offers a caution, though. Capital markets have often identified transformational technologies correctly and overinvested in them at the same time. Railroads transformed America, and railroad investors still lost fortunes. Fiber-optic networks made the modern internet possible, and many telecom carriers still went bankrupt. The internet changed the world, and that did not prevent the dot-com crash. A technology can be revolutionary and overfunded at the same time; those two ideas are not contradictory.

The Fear of Being Second

This brings me to what I think is the most underappreciated force in the AI infrastructure boom: fear. Not fear that AI will fail, but fear that a competitor will succeed first.

Imagine running a major hyperscale company. Your advisers say the next generation of models may need dramatically more computing power. Your competitors are buying GPUs, signing long-term power agreements, acquiring sites, negotiating with utilities and exploring nuclear, gas and fuel cells. Your technical team warns that insufficient capacity could slow model development. You could take a cautious approach, but caution carries its own risk: if AI does become the defining platform of the next several decades and your rival builds the infrastructure while you do not, the cost could be existential. From that chair, overbuilding looks far less dangerous than underbuilding.

That creates a powerful incentive. Each hyperscaler can rationally decide that it needs more capacity, and the collective result can still be irrational. It resembles an arms race. If one company plans 2 GW, the next considers 3 GW; once someone announces 5 GW, 1 GW starts to sound small. Over time the unit of conversation changes, from hundreds of megawatts to gigawatts and from billions of dollars to hundreds of billions, and the extraordinary begins to sound normal.

AI Has Also Become a National Race

Corporate competition is only one layer. Governments increasingly describe AI leadership as a matter of national security and economic power. The original Stargate announcement presented the project not simply as commercial infrastructure but as a way to secure American AI leadership and provide strategic capability for the United States and its allies. Federal regulators have similarly linked faster integration of large loads to the innovation economy and national security. (I explored this shift in more depth in AI Is Becoming National Strategy.)

Once infrastructure is tied to national competitiveness, normal financial discipline becomes harder to apply. The question "What is the return on this project?" starts competing with a different one: "What happens if we don't build it?" That second question is much harder to model in a spreadsheet. When China, national security and technological sovereignty enter the discussion, a data center stops being a real estate project connected to a utility and becomes strategic infrastructure, and strategic infrastructure tends to attract capital even when the conventional economics are uncertain.

The Community Sees Something Different

The people who live near these projects do not necessarily share the same sense of urgency. Local resistance has been growing around the country, with disputes over electricity prices, water use, noise, farmland, new transmission lines and the relatively small number of permanent jobs some facilities create.

In Nebraska, for example, residents from very different political backgrounds came together in opposition, and Cass County commissioners approved a pause on new data centers of up to a year; several neighboring counties took similar steps. Comparable conflicts have emerged elsewhere as communities weigh promised investment against concerns about power, water and quality of life.

The national narrative and the local one sound very different. Nationally, the message is "we must win the AI race." At the corporate level it is "we need more compute," and in financial markets it is "we need exposure to AI infrastructure." Locally, the question is simply: "Who is paying for this?" That last question may turn out to be the one that matters most.

What If Everyone Is Right?

It is tempting to divide this debate into believers and skeptics, but I don't think that is useful. The hyperscalers may be right that AI will require vastly more computing capacity. Investors may be right that AI infrastructure is one of the largest opportunities of this generation. Utilities may be right that billions of dollars of new transmission and generation are needed. Communities may be right that they should not automatically carry the financial and environmental burden of infrastructure built for some of the wealthiest companies in history.

And the skeptics may be right about something else: nobody knows precisely how much AI infrastructure the world will need five years from now. That uncertainty is uncomfortable, because the assets being built today have useful lives measured in decades.

The Real Race May Be About Flexibility

Instead of asking only how quickly we can build more AI infrastructure, we should also ask how flexibly we can build it. In practice, that means:

• Flexible loads. Data centers that can reduce or shift consumption when the grid is constrained are easier for utilities to connect and less risky for ratepayers.
• Hybrid power supply. Combining grid electricity with on-site generation and storage allows capacity to be added in stages and re-purposed if demand changes. Our guide to behind-the-meter strategy covers the trade-offs.
• Technology-neutral planning. Fuel cells, gas generation, batteries, renewables and, eventually, small modular reactors evaluated as parts of one system rather than as competing bets.
• Scenario-tested design. Infrastructure and financing structures that remain viable under low, medium and high AI-growth scenarios, not only the most optimistic one.

Most importantly, it means separating urgency from panic. Moving quickly is not the same as assuming that every forecast will come true. AI will probably reshape large parts of the economy, and we should build infrastructure for it. But the quality of that infrastructure, and of the financial structures behind it, may matter as much as the quantity.

How Much Should He Build?

My conversation eventually returned to a simple question: how much should my client build? A few years ago that would have been mainly an engineering question. Today it is an engineering question, a financing question, a political question, a utility question and, increasingly, a social question.

Underneath all of them sits one more. Are we building AI infrastructure because we know what will be needed, or because nobody wants to be the person who discovers, five years from now, that they did not build enough? History tends to remember technological races in two ways: for the technology that changed the world, and for the money lost by people who predicted the future correctly but misjudged the timing. The answer to that question may decide which of those stories the AI infrastructure boom becomes.

EnergyInnoCore Editor

Independent analysis of power, grid and policy for AI data centers. Every article lists its sources. About the editor · Editorial policy

This article is for general information only and is not engineering, legal, financial or investment advice. See the disclaimer.