We Can Serve You, But Not Yet: The Grid Queue Holding Back AI

Data Centers & Energy By EnergyInnoCore Editor · Published Oct 20, 2026 · Updated Oct 21, 2026 · 10 min read

For the past few years, most conversations about artificial intelligence have revolved around chips. Who has enough GPUs? How fast can Nvidia make them? Which company can build the biggest cluster, and how much will Microsoft, Google, Meta, Amazon and OpenAI spend? Those questions still matter, but another problem is moving to the front of the line: electricity. More precisely, in many parts of the United States the issue is not a lack of generation as such, but getting a very large new data center connected to the power system at all.

A company may have the land, the financing, the servers and cooling equipment on order, and construction crews ready to start. Then the utility says something along the lines of: "We can serve you, but not yet." That "not yet" can mean several years. For developers moving at technology-industry speed, several years is an eternity, and the gap between the pace of AI and the pace of electric infrastructure is becoming one of the most important and least understood constraints on the AI boom. This article explains how the large-load interconnection process works, why it has become a bottleneck, and what is being done about it.

What Is an Interconnection Queue?

The concept is simpler than it sounds. Suppose a company wants to build a 500 MW AI data center, a load comparable to a sizable city. The developer cannot simply ask the utility for a bigger meter. Before the facility can connect, engineers must determine whether the surrounding system can actually handle another 500 MW. Typical questions include:

• Can the existing substation handle the load, and are the transformers large enough?
• Will nearby transmission lines become overloaded?
• What happens if a major line unexpectedly trips offline?
• Will the new load cause reliability problems elsewhere on the system?
• If upgrades are needed, who will build them, and who will pay?

Answering those questions takes engineering studies, planning decisions, regulatory approvals and often major construction. Requests are studied in order, or in groups, which is why they form a "queue." Utilities have always served large industrial customers such as steel mills, chemical plants, aluminum smelters and semiconductor fabs. But AI data centers are arriving at a scale and speed the existing process was never designed to handle.

The Energy Systems Integration Group put it directly in a 2026 report: the rapid growth of large-load requests from data centers and advanced manufacturing has exposed gaps in interconnection processes across U.S. utilities and grid operators. Federal regulators reached a similar conclusion. In June 2026, FERC ordered six regional grid operators to justify or reform their rules for connecting data centers and other very large users, framing the issue around improving "speed to power" while protecting existing customers. When federal regulators start using that phrase, it is no longer a niche planning issue. It is an economic development issue.

AI Moves at One Speed; the Grid Moves at Another

An AI data center is essentially a technology project sitting on top of heavy industrial infrastructure. The technology side moves extremely fast: new GPU generations arrive every year or two, AI models change in months, and multibillion-dollar investment decisions are made quickly. The grid runs on a very different clock. High-voltage transmission projects need years of planning, permitting, engineering, land acquisition and construction. Large transformers and other specialized equipment cannot be pulled off a shelf. New substations must be designed and built, and sometimes new generation must be developed as well. We are trying to connect one of the fastest-moving industries in history to infrastructure that has traditionally been planned decades at a time.

Texas illustrates the scale. ERCOT, which operates most of the state's grid, reported in late 2025 that it was tracking about 226 GW of large-load interconnection requests, up from 63 GW a year earlier, with roughly 73% tied to data centers. By August 2026 the governor's office put total requests at more than 474 GW. Not all of those projects will be built (far from it), but the numbers show how dramatically the planning environment has changed. Utilities now have to work out which projects are real, which are speculative, and how much infrastructure to build before they know the answer.

There Is Another Queue Behind the Queue

Data centers are waiting to connect as loads. At the same time, thousands of proposed power plants and battery projects are waiting to connect as generation, and the two problems are linked. Berkeley Lab's queue data show that at the end of 2025 about 8,200 generation and storage projects were actively seeking U.S. grid interconnection, representing roughly 1,312 GW of generation and 749 GW of storage. Many will never be built; projects routinely drop out because of financing, permitting, upgrade costs or changing economics. (We explain ways to get more out of existing lines in The Hidden Capacity Already on the Grid.)

Still, the irony is hard to miss. On one side, AI developers are saying "we need more electricity." On the other, generation developers are saying "we are trying to build it." In between sits the transmission and interconnection system. That is why "America needs more power plants" does not fully describe the AI energy problem. The country needs generation, but it also needs the infrastructure and processes that let generation and demand connect. A gigawatt of available power in the wrong place, or stuck behind transmission constraints, does little for a gigawatt-scale data center that wants to start operating next year. Electricity has to be available where it is needed, when it is needed, through infrastructure capable of delivering it.

Northern Virginia Shows What Concentrated Demand Does

Northern Virginia has long been the center of the global data center industry, and conventional cloud infrastructure there is now being layered with enormous AI demand. The surrounding PJM grid, which serves about 67 million people across all or parts of 13 states and Washington, D.C., is feeling the pressure. PJM's planning suggests large new loads, including data centers, could add as much as 70 GW of demand by 2038, while the region is already dealing with plant retirements, transmission constraints and rising capacity needs.

The physical grid is showing signs of strain. According to Reuters, PJM's transmission congestion costs reached about $6 billion in the first half of 2026, up 43% from a year earlier, with Northern Virginia among the most affected areas. That does not mean data centers alone caused the increase; weather, outages, system operations and other load growth all play a role. But it shows that when demand concentrates in one place, generating enough electricity somewhere in the region is not sufficient. The transmission network has to deliver it into that load pocket, which is why wires, transformers and substations have become as strategically important as GPUs. (Our article on why Virginia and Texas keep winning data centers looks at how this shapes site selection.)

The Queue Is Also a Credibility Problem

Utilities often do not know which proposed data centers will actually be built. Several developers may each request 500 MW; one developer may be evaluating multiple sites for the same project; some projects are financed while others are early-stage ideas. Should the utility plan billions of dollars of new infrastructure assuming all of them will materialize? If it does and half disappear, existing customers could be left paying for unnecessary assets. If it waits for certainty, legitimate projects lose years. Utilities want better information from developers before committing; developers want firm commitments from utilities before committing billions to a site. Each side wants certainty from the other first.

Texas offers an example of the response. In August 2026, Governor Greg Abbott directed state regulators and ERCOT to audit every data center in the interconnection process before any could move forward, citing the sheer volume of requests. More broadly, interconnection processes are likely to require stricter evidence that a project is ready: deposits, financial commitments, site control and milestone deadlines that separate serious projects from speculative reservations. Getting into the queue may become harder, but getting real projects through it could become faster.

Can the Process Be Fixed?

There is no single solution, because the bottleneck is not one thing. It involves engineering, permitting, transmission construction, forecasting, generation availability and an interconnection process designed for another era. In June 2026, Berkeley Lab published a report identifying more than 40 potential solutions for accelerating large-load connections, grouped into load forecasting, interconnection, resource planning and procurement, markets and operations, and cost allocation. The Department of Energy's earlier interconnection roadmap laid out similar reforms for generator queues.

Grid operators are changing their processes too. PJM has moved away from a strict first-come, first-served queue toward a process that prioritizes projects that are ready to proceed, and says it worked through a large backlog of generation requests between 2023 and 2025. But process reform only goes so far. At some point America still has to build transmission lines, substations, transformers, power plants, battery systems and, increasingly, power infrastructure located right beside data centers.

What If AI Companies Decide They Can't Wait?

If waiting five or six years for grid capacity threatens a multibillion-dollar AI investment, the economics change, and building your own power plant no longer sounds extreme. That is why "bring your own generation" is moving from an unusual idea to a mainstream strategy. PJM has supported a voluntary bring-your-own-new-generation approach for large customers that want to reduce their curtailment risk and get power sooner. Options include gas turbines, reciprocating engines, fuel cells, batteries and microgrids, and over the longer term perhaps nuclear power and small modular reactors (see When the Power Won't Show Up, Do You Build Your Own?).

That could fundamentally change the relationship between technology companies and utilities. Traditionally a data center was simply a customer of the power system. The next generation of hyperscale campuses may look more like miniature utilities: producing electricity, storing it, managing their own microgrids and interacting with the regional grid when needed.

The Real AI Race May Be a Race for Megawatts

For years the AI infrastructure race has been measured in GPUs. Increasingly it may be measured in megawatts, and the companies that secure reliable electricity first may gain an advantage as meaningful as those that secure the newest chips. That makes grid interconnection much bigger than a utility paperwork problem; it is part of AI strategy.

The United States has plenty of proposed generation, enormous private capital for AI infrastructure and technology companies willing to spend hundreds of billions of dollars. What it does not always have is a fast way to connect those pieces. That may be the central irony of the AI boom: one of the most advanced technologies ever created could be limited by some very familiar infrastructure, namely power lines, substations and transformers. The next phase of the AI race will not be won in the data center alone. It may be won or lost at the grid connection.

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.