Dallas Fed AI Power Demand Meets a Grid Built for Slower Growth

The Dallas Fed AI power demand analysis identifies a stark timing conflict: data centers take about two years to build, while transmission can require a decade. The United States has generating capacity, capital, and technical options. It does not yet have a coordinated system that can deploy them at AI’s pace.

That distinction matters because the power challenge is often framed as a simple shortage. The Federal Reserve Bank of Dallas reached a more qualified conclusion after bringing together technology, energy, finance, and regulatory leaders. The immediate constraint is not whether America can generate enough electricity in theory. It is whether utilities, grid operators, developers, regulators, and communities can make credible commitments quickly enough.

The findings arrive as electricity consumption resumes growth after years of relative stability. Data centers are concentrating that growth in Texas, Virginia, Arizona, Nevada, and other development hubs. Those clusters create local problems that national generation totals can obscure.

The central contest is therefore AI construction speed against power-system coordination. Hyperscalers can order servers, secure land, and finance facilities much faster than utilities can approve transmission, procure turbines, and complete interconnections. Grid planners must also determine which proposed facilities are real before committing customers to expensive infrastructure.

The Dallas Fed Finds a Coordination Crisis, Not an Energy Dead End

The Dallas Fed’s central finding is that coordination failures now pose a greater obstacle than a lack of viable energy technology.

The bank published its AI power constraints analysis on September 18, 2026. It summarizes discussions from the Powering AI conference held in Dallas on March 4 and 5.

The event included representatives from Google, ERCOT, PJM, Vistra, Goldman Sachs, Morgan Stanley, NRG Energy, and several infrastructure firms. Participants examined generation, grid management, financing, supply chains, and AI adoption.

Their conclusions resisted both easy optimism and predictions of an unavoidable power crisis. Existing plants can produce more electricity, while flexible data center operations can reduce stress during critical periods. Capital is also available for projects backed by credible customers.

However, none of those resources automatically becomes a dependable grid connection. Developers still face shortages of electrical equipment, skilled workers, construction capacity, and approved transmission. Utilities must make long-lived investments while customer forecasts remain unusually uncertain.

The Dallas Fed highlighted a basic timing mismatch. A data center can be completed in roughly two years. A new power plant can require at least twice as long, while major transmission projects can take seven to ten years.

That mismatch changes how planners must think about adequacy. A region might possess enough annual energy while lacking the wires or dispatchable capacity needed at a particular location. It might also approve infrastructure for speculative projects that never become operational.

Texas offers the clearest example. The state combines abundant energy resources, significant land, a relatively fast development environment, and an independent grid. Those advantages have attracted a huge pipeline of data center proposals.

The Dallas Fed reported that ERCOT’s future large-load queue stood at 232 gigawatts through 2030 when discussed by conference participants. By comparison, ERCOT’s record demand reached 91 gigawatts in July 2026.

A queue is not the same as a forecast. It includes early proposals, duplicate development strategies, and projects without final financing or equipment. Yet utilities cannot simply ignore it because some facilities will connect and require dependable service.

The Dallas Fed described the resulting outlook as ranging from “zero to infinity.” That phrase captures the planning problem better than a single national demand projection. The power industry must prepare for substantial growth without treating every announced data center as certain.

This uncertainty explains why the story is not simply about building more generation. Grid planners must identify real demand, assign financial responsibility, and match each project with deliverable power. Every weak commitment consumes engineering time and can distort broader investment decisions.

Dallas Fed AI Power Demand Is Growing Faster Than Grid Timelines

The demand increase is credible, but its speed and geographic concentration make it harder to serve than ordinary national growth.

U.S. electricity consumption remained close to flat for much of the period after 2005. The Energy Information Administration estimates that net energy for load grew only 0.1 percent annually from 2005 through 2019.

The pattern changed after 2020. Demand grew about 1.7 percent per year between 2020 and 2025, according to the agency. Large computing facilities became an important driver in regions with concentrated data center development.

The EIA expects the country to record its strongest four-year electricity-demand expansion since 2000. Its January 2026 forecast called for U.S. electricity use to rise 1 percent during 2026 and another 3 percent in 2027.

Longer-range estimates show why utilities cannot treat the change as a temporary bump. Lawrence Berkeley National Laboratory estimates that data centers could consume 11.8 percent of U.S. electricity in 2030.

Its 2030 estimate spans a wide range, from 9.5 percent to 15.3 percent. The spread reflects different assumptions about hardware shipments, cooling systems, computing efficiency, and facility growth.

The range is as important as its midpoint. Building for the highest case risks leaving households and businesses responsible for underused infrastructure. Building for the lowest case risks delayed connections, reliability problems, and higher emergency costs.

The International Energy Agency reaches a similar conclusion from a global model. It expects data center electricity consumption to rise from about 485 terawatt-hours in 2025 to roughly 950 terawatt-hours by 2030.

The agency expects AI-focused facilities to grow much faster than the broader data center category. Its energy demand outlook says data centers will account for nearly half of U.S. electricity-demand growth through 2030.

National percentages still hide the operational challenge. Data centers are not distributed evenly across millions of buildings. Developers cluster them near fiber routes, available land, tax incentives, skilled workers, and existing power infrastructure.

A concentrated campus can therefore transform a local forecast before the national percentage looks alarming. The relevant question for a utility is not only how many terawatt-hours America will consume. It is whether a particular substation can serve a new campus during the hottest or coldest hour.

AI hardware raises another complication. Training clusters place dense electrical loads behind a relatively small number of connections. Cooling, backup systems, and networking equipment add further demand beyond the servers themselves.

Operators also want high availability because expensive accelerators create value only while performing useful work. That commercial logic encourages steady consumption, even when the grid would benefit from temporary reductions.

The Dallas Fed AI power demand assessment does not dispute these growth forecasts. It challenges the assumption that every requested megawatt must operate continuously and receive an immediate conventional connection.

That distinction opens a path forward. If facilities can shift workloads, reduce consumption during emergencies, or support themselves temporarily, the grid can serve more computing with existing assets. The unresolved issue is how reliably those commitments will perform.

Flexible Data Centers Could Turn Load Into a Grid Resource

The fastest near-term response is not a new fleet of power plants, but better use of existing capacity paired with enforceable flexibility.

Traditional cloud facilities often requested about 30 megawatts and took years to reach full utilization. New AI campuses can request much larger connections and seek rapid service from the beginning.

Utilities cannot serve every large request under old assumptions. They increasingly need contracts that define when a facility can consume power and when it must reduce demand. This model treats flexibility as a condition of faster access.

A flexible load is a customer that changes electricity consumption in response to grid conditions or prices. For data centers, that response might include moving computing jobs, delaying nonurgent training, or switching to on-site resources.

Not every AI workload offers equal flexibility. Real-time inference supports live applications and often has strict latency requirements. Model training, testing, batch processing, and some data preparation can offer wider scheduling windows.

Geographic workload shifting provides another option. A company operating facilities in several regions can route suitable jobs toward areas with available electricity. Software orchestration can perform that adjustment without physically moving servers.

The Dallas Fed pointed to Winter Storm Fern in January 2026 as evidence that demand response can matter. Commercial and industrial customers voluntarily curtailed consumption during peak conditions in Texas.

Those reductions helped keep actual demand below forecasts and limited price pressure. The episode did not prove that every data center can operate flexibly. It showed that coordinated load reductions can support the grid during constrained periods.

Existing generation also has room for higher utilization. Conference participants noted that Texas fossil-fuel plants operate below 50 percent annual utilization but can exceed 70 percent. That spare annual capacity creates options, although fuel delivery and peak reliability still matter.

Higher utilization could spread fixed grid costs across more electricity sales. In that case, new data center demand might reduce some per-unit costs. The opposite result remains possible if utilities must build expensive plants and wires for inflexible peaks.

On-site generation can shorten the path to service. Some projects are considering natural gas, batteries, solar generation, or combinations of those technologies. These systems can support normal operations or reduce withdrawals during grid emergencies.

“Bring your own power” does not remove a facility from the wider energy system. On-site gas equipment still needs fuel infrastructure and permits. Batteries store electricity rather than create it, while solar output varies with time and weather.

The more practical model combines several resources. Grid service handles routine operations, batteries cover rapid changes, and dispatchable equipment supports longer constraints. Flexible software then decides which computing jobs can wait or move.

This approach also creates measurement questions. Grid operators need confidence that promised curtailment will occur when called. Contracts require clear baselines, telemetry, penalties, and rules for testing performance.

A data center that claims flexibility but rarely reduces load offers little reliability value. A facility that consistently responds can function more like a grid partner. That difference should affect its position in the interconnection process.

Flexibility will not replace new infrastructure. It can, however, bridge the period before new generation and transmission arrive. It can also reduce the amount of capacity that must be built solely for rare demand peaks.

AI Construction Speed Is Pressuring Utilities and Grid Operators

Utilities must commit billions of dollars years before they know which AI projects will become durable electricity customers.

The power industry usually plans around demand that develops gradually. AI developers are submitting requests at a scale and speed that strain those forecasting methods.

The ERCOT queue illustrates the gap between interest and operational reality. A March 2026 grid overview showed only 6.7 gigawatts of observed large load in 2025. Its broader queue included vastly larger totals for later years.

Many projects had not submitted complete studies. Others remained under review or had not satisfied key connection requirements. Treating every entry as firm demand would produce an implausible construction target.

Yet dismissing the queue would be equally risky. Even a modest conversion rate would produce substantial new load. The challenge is separating serious projects from applications intended to reserve optionality.

Texas responded by increasing the financial requirements attached to large-load applications. Nonrefundable fees force developers to put more capital at risk before occupying scarce planning capacity.

Stronger commitments give utilities better evidence that a customer will appear. Long-term contracts, credit support, deposits, and minimum payments can also protect existing ratepayers if a project is canceled.

This is not simply administrative housekeeping. A new substation, transmission line, or generator can operate for decades. If its intended customer disappears, someone still must cover financing, maintenance, and regulatory costs.

The same pressure appears in PJM, the grid region covering all or parts of 13 states and Washington, D.C. Data center expansion has increased load forecasts while generation development remains slow.

PJM’s 2026 market design analysis describes a physical mismatch between rapid load growth and sluggish supply additions. It says generation can take two or three times longer to build than a hyperscale facility.

The region has improved its interconnection studies and offered agreements to many projects. Conversion remains the problem. PJM reported that about 24 gigawatts with executed agreements had terminated since 2020 before reaching commercial operation.

Those cancellations included 13.5 gigawatts of natural gas projects. The figures show why completing an interconnection study does not guarantee steel, financing, equipment, or an operational plant.

Supply chains add another layer of delay. Large transformers, gas turbines, switchgear, and specialized electrical components require long procurement periods. A project can possess capital and permits but still wait for essential equipment.

Labor presents a similar bottleneck. Data centers and power projects compete for electricians, technicians, network engineers, and specialized construction trades. Training new workers takes time, while experienced crews cannot cover every proposed site.

The Dallas Fed described skilled labor as one of the hardest constraints because it lacks a quick substitute. Companies can pay more or finance training, but neither response instantly creates experienced workers.

These conditions put utilities in an uncomfortable role. They must enable economic development while defending reliability and protecting customers from speculative costs. Moving too slowly can send investment elsewhere, while moving too quickly can create stranded assets.

The Forecasts Still Leave Three Major Risks Unresolved

The optimistic case depends on flexibility, credible demand, and public consent, none of which is guaranteed at the required scale.

The first uncertainty is whether enterprise AI adoption will justify the proposed computing buildout. Conference participants noted that many companies remain stuck in pilot programs.

One speaker estimated that only about 5 percent of corporate AI pilots achieve value at scale. That figure reflects a conference view rather than a comprehensive national measurement. It still captures the commercial risk behind electricity forecasts.

Hyperscalers can finance large investments from substantial operating cash flow. The Dallas Fed reported planned annual capital spending of $700 billion to $900 billion during 2026 and 2027. Amazon alone planned a $220 billion capital budget.

Large balance sheets reduce financing risk, but they do not guarantee profitable demand. If customers fail to expand AI use, some announced campuses could be delayed, resized, or canceled.

The second risk is reliability. The North American Electric Reliability Corporation found that 13 of 23 assessed areas face resource adequacy challenges during the next decade.

Its reliability assessment warns that new resources and transmission are not keeping pace with data centers and other large loads. Extreme weather can expose that gap before annual energy totals look problematic.

Flexible operation helps only when companies accept meaningful interruptions. Some facilities may offer limited reductions because customer contracts require continuous service. Others may keep backup generators for emergencies but face restrictions on routine operation.

The third risk is political resistance. Communities are raising concerns about electricity bills, noise, water consumption, pollution, land use, and limited transparency.

Developers increasingly target unincorporated or rural areas where land and power are more available. Many AI workloads tolerate greater distance from population centers because latency matters less for training than for real-time services.

That shift can reduce pressure in established urban hubs. It can also transfer costs and infrastructure disputes to communities with less planning capacity.

Public opposition can delay permits, restrict water access, or stop construction altogether. Conference participants said communities have blocked billions of dollars in proposed data center investment.

Developers cannot resolve those conflicts with national economic arguments alone. Residents want specific information about electricity use, rate impacts, tax arrangements, water needs, backup generation, and local employment.

Promises about jobs require particular care. Data centers create significant construction activity but employ fewer permanent workers than many industrial sites of comparable size. Communities may judge tax incentives differently once that distinction becomes clear.

Power costs remain similarly uncertain. New loads can improve utilization and share fixed costs across more sales. They can also require expensive generation and transmission that increases bills.

Lawrence Berkeley National Laboratory’s work on large-load rate design identifies both insufficient supply and underused investments as material risks. Regulators must decide how much financial responsibility belongs with developers rather than existing customers.

Environmental tradeoffs also complicate the near-term response. The EIA’s high-demand scenario finds that faster data center growth would increase natural gas generation substantially through 2027.

Under its baseline, U.S. gas generation rises 1.7 percent from 2025 through 2027. Under higher demand growth, the increase reaches 7.3 percent.

Coal generation still declines in both cases, but the decline becomes smaller when demand grows faster. Renewable generation continues expanding, yet dispatchable resources cover periods when wind and solar output cannot meet concentrated loads.

That result does not mean AI requires a permanently fossil-heavy grid. It shows that deployment speed can favor resources already available or relatively fast to build. Nuclear and geothermal projects generally have longer development timelines.

The Dallas Fed AI power demand conclusion therefore rests on conditions, not certainty. America has multiple technical routes to supply AI. It has not proved that those routes can arrive together, on schedule, with costs allocated fairly.

Three Signals Will Show Whether the Grid Can Keep Pace

The next phase will be decided by queue conversion, verified load flexibility, and the pace of physical infrastructure delivery.

The first signal is the share of large-load applications that become financially committed projects. Queue totals alone create attention but reveal little about near-term electricity use.

Watch for deposits, executed service agreements, completed engineering studies, land control, and firm equipment orders. These milestones show that developers have moved beyond reserving an option.

ERCOT’s higher application fees should provide an early test. If the queue shrinks while committed projects advance, planners will gain a more credible demand forecast. If speculative requests persist, uncertainty will continue delaying rational investment.

The second signal is measured curtailment from operating data centers. Companies increasingly describe workload shifting and demand response as solutions. Grid operators need performance data from actual stress events.

The relevant measures include megawatts reduced, response speed, duration, and recovery behavior. Contracts should also reveal whether reductions are mandatory, compensated, or merely voluntary.

Repeated performance would strengthen the Dallas Fed’s coordination thesis. It would show that computing demand can adapt while slower infrastructure catches up. Weak performance would shift the problem back toward new firm generation.

The third signal is whether approved generation and transmission become operational on schedule. Completed studies and signed agreements are intermediate steps, not electricity.

Track transformer deliveries, turbine orders, construction starts, transmission approvals, and commercial operation dates. Delays across those categories would widen the timing gap even if forecasts become more accurate.

The same test applies to on-site power. Announced gas, nuclear, geothermal, solar, and battery projects matter only when they obtain permits, equipment, fuel access, and operational status.

For developers, the strategic implication is direct. Power availability now influences site selection, facility design, workload architecture, and launch schedules. Electricity planning can no longer begin after land acquisition.

Utilities face a different decision. They need stronger customer commitments without blocking credible projects through slow or inconsistent processes. Transparent standards can reduce both speculation and arbitrary treatment.

Regulators must decide who pays when forecasts prove wrong. Rate structures should protect households and existing businesses while recognizing the potential benefits of higher grid utilization.

Communities also need earlier access to project details. Developers that disclose water, energy, emissions, jobs, and local infrastructure requirements can address concerns before they harden into opposition.

The U.S. power industry can support substantial AI growth, but technical capacity does not guarantee timely delivery. The decisive question is whether institutions can turn uncertain proposals into enforceable commitments.

Readers should watch those three signals rather than another headline-sized queue number. Are speculative requests becoming financed projects? Are flexible facilities reducing load during real grid stress? Are promised plants and transmission lines reaching operation?

If the answers improve together, the Dallas Fed AI power demand assessment will look justified. If any one fails, data center schedules, electricity costs, or reliability will absorb the consequence. The grid race will be won through measurable execution, not ever-larger announcements.