Blog · Analysis · May 2026

The Interconnection Queue Becomes AI Governance

The AI buildout is turning the electric grid's interconnection queue into a hidden governance layer: who connects, who waits, who pays, and whose demand becomes infrastructure.

Queue as Institution

The AI race is often described as a contest over models, chips, data, talent, and capital. That list is incomplete. The race also depends on a less glamorous institution: the electric grid interconnection process.

An interconnection queue is where proposed power plants, storage projects, and large facilities move through technical studies before they can connect to the transmission system. The process asks whether the grid can handle the new project, what upgrades are needed, how much those upgrades cost, who pays, and when the project can safely enter service.

That sounds administrative. In the AI era it becomes political infrastructure. A frontier-model company can raise money, reserve GPUs, sign a cloud contract, and announce a campus. But if the power cannot be delivered, the model capacity remains a promise. The queue becomes a machine that converts ambition into time, cost, location, and public burden.

This is not only a data-center story. It is a governance story about allocation. The grid decides which industrial futures can become real soon, which must wait, which require new transmission, which rely on gas, nuclear, solar, storage, or behind-the-meter generation, and which costs get moved onto other customers. AI capacity is being routed through utility planning.

Demand Arrives Faster Than Planning

The demand shock is real enough to move federal planning. The U.S. Department of Energy's 2024 data-center energy report, produced by Lawrence Berkeley National Laboratory, estimated that U.S. data centers used 176 terawatt-hours in 2023, about 4.4 percent of total U.S. electricity. The same report projected 325 to 580 terawatt-hours by 2028, or roughly 6.7 to 12 percent of total U.S. electricity, depending on growth and efficiency assumptions.

The Energy Information Administration has since framed data centers as a driver of a broader reversal in U.S. electricity demand. After years of nearly flat load growth, EIA reported that electricity demand grew about 1.7 percent annually from 2020 to 2025, compared with about 0.1 percent annually from 2005 to 2019. Its 2026 analysis pointed to ERCOT and PJM as regions likely to see the fastest data-center-related demand growth through 2027.

The old planning world assumed demand moved slowly enough for utilities, regulators, generators, and transmission planners to catch up through normal cycles. AI data centers challenge that rhythm. A single campus can resemble a small city load. Multiple campuses can appear in the same region because latency, tax incentives, fiber, land, water, workforce, and existing power infrastructure make some locations more attractive than others.

The result is not a generic "more electricity" problem. It is a timing, geography, and governance problem. Power has to be available in the right place, at the right voltage, with enough transmission, reserves, fuel, cooling, and operational flexibility to serve a load that may want near-continuous reliability.

Generation Waits Too

The queue is already crowded on the supply side. Berkeley Lab's Queued Up 2025 database reported that, at the end of 2024, about 10,300 projects were actively seeking U.S. grid interconnection, representing 1,400 gigawatts of generation and about 890 gigawatts of storage. The dataset covers transmission queues representing roughly 97 percent of installed U.S. generating capacity.

Those numbers do not mean every project will be built. Queue projects often withdraw, change, fail financing, or become uneconomic after network-upgrade costs are assigned. But the scale shows why interconnection has become a bottleneck. The grid is not simply waiting for more ideas. It is trying to study, price, sequence, and physically integrate an enormous pipeline of proposed resources.

FERC Order No. 2023 was designed for this backlog. FERC described the rule as a reform to reduce interconnection delays, improve certainty, and modernize how transmission providers process new generating facilities. It moved the system toward cluster studies and "first-ready, first-served" procedures so speculative projects would be less able to occupy the queue ahead of projects prepared to proceed.

That reform matters for AI governance because data-center demand and clean generation are now racing through the same physical constraints. If AI demand arrives faster than new resources and transmission can connect, the near-term marginal answer may be existing gas plants, delayed retirements, higher capacity prices, or emergency infrastructure. If clean generation waits too long in the queue, AI's public story of innovation can be powered by the slowest parts of the old system.

Large-Load Politics

The interconnection problem is no longer only about generators. Large loads are becoming active grid actors. A hyperscale data center is not like a neighborhood adding a few appliances. It can request hundreds of megawatts or more, seek special tariffs, co-locate near generation, build behind-the-meter resources, sign power-purchase agreements, negotiate curtailment, or ask for dedicated transmission upgrades.

DOE's clean-energy resource guidance for data centers emphasizes the same complexity. Data-center demand can be regionally concentrated, require firm power, and interact with transmission, storage, generation, energy efficiency, demand flexibility, rate design, interconnection, and regulatory reform. DOE points to a portfolio approach: solar, wind, storage, efficiency, nuclear, geothermal, retired coal-site redevelopment, transmission expansion, and demand resources.

NERC's 2025 reliability-risk priorities report also treats large loads, including data centers, as part of grid-transformation risk. The report warns that permitting delays, misaligned policies, transmission needs, rising demand, and critical-infrastructure interdependencies can magnify reliability risk when electricity, natural gas, water, transport, and communications systems are planned in separate institutional rooms.

That is the new politics of AI infrastructure. The model is not only inside a data center. It is inside a gas pipeline constraint, a substation upgrade, a transmission study, a rate case, a water permit, a local tax deal, a reliability margin, and a market rule. The interface that the public sees is a chatbot. The institution that makes it possible is a stack of utility decisions.

Who Pays for the Machine

The most important governance question is cost allocation. If a data-center campus requires new transmission, generation, distribution upgrades, reserves, or accelerated planning, who pays?

One answer is the data-center customer. If the load causes the need, the load should bear the cost through tariffs, direct assignment, special contracts, or self-funded upgrades. That protects ordinary customers from subsidizing private compute expansion.

Another answer is socialization. Some grid upgrades benefit more than one customer over time, and utilities often recover infrastructure costs from broad rate bases. A new line or substation built partly for data-center demand may later serve other loads, improve reliability, or support regional growth. Overly narrow cost assignment can also slow useful infrastructure.

The danger is that AI becomes too politically powerful inside this ambiguity. A company can promise jobs, tax revenue, national competitiveness, clean-energy procurement, and technological leadership while the less visible risks move into household bills, reliability margins, local water systems, land-use conflict, or fossil generation. The data center receives the branded future. The public receives the infrastructure externalities.

Capacity-market debates in PJM show why this matters. The region has faced sharp controversy over how forecast data-center load affects capacity prices and procurement. Whatever one thinks of specific auction rules, the institutional lesson is clear: forecasted AI demand can change market prices before the public has a settled view of whether, where, and under what conditions that demand should be served.

AI governance therefore cannot stop at model evaluations. It has to ask whether the physical conditions of model production are being governed in public.

The Governance Standard

A serious interconnection regime for AI-scale loads should answer more than "can this customer connect?"

First, disclose the load shape. Regulators and grid planners need credible information about expected peak demand, ramping behavior, backup generation, water and cooling dependencies, redundancy, curtailability, and staged buildout. A vague megawatt headline is not enough.

Second, separate firm demand from flexible demand. Some AI workloads can move in time or location more easily than hospitals, transit, homes, or emergency services. If data centers claim critical-infrastructure status, they should also show which compute can be curtailed, delayed, shifted, or run on local storage during grid stress.

Third, make cost allocation legible. Ratepayers should know which upgrades are directly assigned to large-load customers, which are broadly socialized, which are subsidized by public incentives, and which depend on future customers who may never arrive.

Fourth, coordinate generation and load queues. A region should not approve large AI loads faster than the credible resource and transmission plan needed to serve them. The queue should link demand growth to energy sufficiency, not merely to connection paperwork.

Fifth, require community evidence before inevitability claims. Local people should see the expected jobs, tax terms, water use, noise, backup generation, land impacts, rate effects, and reliability implications before the project is treated as already decided.

Sixth, align clean-energy claims with physical delivery. A data center's renewable certificate or corporate power-purchase agreement should not be mistaken for proof that the local grid can serve the load cleanly and reliably at the hour of use.

Seventh, preserve public authority over critical infrastructure. Private compute demand should not quietly rewrite utility planning priorities, emergency powers, or market rules without democratic visibility.

The Site Reading

The interconnection queue is a hidden interface. It does not look like an AI system. It does not speak in fluent text or generate images. But it decides what kind of artificial intelligence can become operational reality.

This is recursive infrastructure. Models create demand for compute. Compute creates demand for electricity. Electricity demand reshapes grid planning. Grid planning determines where compute can be built. The resulting capacity changes which models can be trained, deployed, sold, and embedded into institutions. The output of the model economy loops back into the physical institutions that make more model economy possible.

That loop should not be governed by announcement momentum. The public needs a way to distinguish real infrastructure from speculative land banking, clean capacity from accounting claims, flexible compute from immovable load, and private advantage from public necessity.

The queue is where those distinctions can be made concrete. It can become a public memory of proposed demand, required upgrades, assigned costs, reliability findings, withdrawal rates, and community conditions. Or it can become a quiet funnel through which AI companies convert social urgency into grid priority.

The better rule is simple: no model capacity without infrastructure accountability. If AI systems are going to mediate knowledge, labor, government, health, education, and culture, then the power systems behind them must be governable too. The machine's intelligence begins before the model starts. It begins where the wire is allowed to connect.

Sources

Return to Blog