US AI Data Center Boom Hits Physical Limits

Google News is highlighting a sharp conflict in America’s AI expansion: demand remains at record levels, yet new data center construction has started declining.

Capacity under construction across eight primary North American markets fell from 6,350.1 megawatts in 2024 to 5,994.4 megawatts at the end of 2025. That was the first annual decline since 2020, according to CBRE. Permitting delays, limited power, equipment backlogs, and shortages of skilled workers are stretching project schedules.

The decline does not show that companies suddenly need fewer AI servers. Vacancy reached a record low while leasing activity increased. The reversal is between the technology industry’s spending commitments and the physical system required to turn that money into operating computing capacity.

Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and specialized cloud providers all depend on that system. Their competitive plans now collide with utilities, grid operators, local governments, equipment manufacturers, and construction contractors that move on different schedules.

The decisive AI infrastructure race is no longer only about buying advanced chips. It is about securing electricity, permits, transformers, electricians, cooling equipment, and community approval before a rival does.

Google News Captures a Construction Slowdown Without a Demand Slowdown

The most important data point is not that construction declined. It is that construction declined while nearly every measure of demand remained unusually strong.

CBRE reported 5,994.4 megawatts of capacity under construction in primary markets at the end of 2025. The comparable figure was 6,350.1 megawatts one year earlier. That represents a decline of about 5.6 percent.

Those figures measure projects actively being built, not every campus mentioned in a press release or filed with a local planning authority. Moving from an announced project to active construction requires control of land, a viable power plan, permits, financing, equipment orders, and contractors.

The demand side moved in the opposite direction. Primary-market supply expanded 36 percent during 2025, while net absorption reached a record 2,497.6 megawatts. Net absorption measures newly occupied capacity after accounting for space that tenants vacated.

The vacancy rate fell to 1.4 percent, another record. Northern Virginia alone absorbed 1,102 megawatts during the year, according to the market fundamentals.

This combination matters because it separates an infrastructure slowdown from an AI demand collapse. Customers are taking available capacity, and much of the capacity still being built has already been committed.

The market is therefore experiencing scarcity, not oversupply. Developers have potential customers, but they cannot consistently deliver suitable buildings on the schedules those customers want.

That distinction can disappear in headlines. A decline in construction can sound like hyperscalers have reconsidered their AI ambitions. The evidence instead points toward delayed execution after an exceptional development surge.

Construction activity had already expanded dramatically. The 6,350.1 megawatts under construction at the end of 2024 was more than twice the 2023 total. It was also almost 14 times the 456.8 megawatts recorded in 2020.

Maintaining that pace would require every supporting layer to expand at a similar rate. The grid would need new generation and transmission. Manufacturers would need more switchgear, transformers, turbines, and cooling systems. Contractors would need enough specialized workers across several regions.

Those layers did not scale as quickly as proposed server capacity.

AI infrastructure has also become harder to build. A traditional cloud facility could distribute computing across relatively modest power blocks. Modern AI campuses often seek hundreds of megawatts at one location, with dense racks and demanding cooling requirements.

A megawatt measures one million watts of electrical power. At campus scale, adding hundreds of megawatts can require new substations, transmission upgrades, and generation agreements rather than a routine utility connection.

Developers are responding by looking beyond established markets such as Northern Virginia. Faster long-distance networks make secondary locations more practical, especially when electricity is available there sooner.

Moving does not remove the problem. It changes its form. Smaller markets can offer land and power, but they may lack the construction workforce, fiber routes, equipment suppliers, or local experience needed for a rapid build.

The Google News narrative is therefore best read as a scheduling warning. AI demand can remain strong while the delivery curve moves to the right.

Electricity Has Become the First Site-Selection Test

A promising data center site is no longer valuable unless its electricity can arrive on a credible schedule.

A hyperscale developer can negotiate for land, revise a building design, or phase a campus. It cannot operate thousands of accelerators without a dependable supply of electricity.

Grid access has become the earliest and most restrictive project test. CBRE says developers increasingly prioritize sites with more than 200 megawatts of available power. In many established markets, grid capacity for existing projects is largely allocated through 2030.

This shortage has emerged as national electricity demand begins growing again. Between 2005 and 2019, U.S. net energy demand increased by only about 0.1 percent annually. From 2020 through 2025, the annual growth rate rose to approximately 1.7 percent.

The U.S. Energy Information Administration expects large computing facilities to remain a major source of rising demand. Its 2026 outlook describes data center server consumption as a central factor behind projected electricity growth.

The agency also examined a scenario in which regional demand rises 50 percent faster than its baseline forecast. Under that scenario, data center growth would place additional pressure on generation in Texas, the Mid-Atlantic, and parts of the Southwest.

The electricity demand outlook illustrates why a signed utility agreement cannot always guarantee a near-term connection. A utility must confirm that generation, transmission lines, substations, and local distribution equipment can support the requested load.

Regional grid operators face another challenge: proposed projects do not always become real loads. Developers may submit multiple connection requests while comparing sites, creating a pipeline much larger than the capacity ultimately built.

If planners assume every proposal will arrive, customers can fund infrastructure that is never needed. If planners discount too many proposals, a genuine project can arrive without enough power waiting for it.

PJM Interconnection faces this problem across a territory serving roughly 67 million people. The region includes Northern Virginia, the largest U.S. data center market.

PJM projects that data center growth can add about 30 gigawatts of demand between 2025 and 2030. One gigawatt equals 1,000 megawatts, so this forecast represents an enormous amount of new load within five years.

PJM’s planning review describes the growth as unprecedented and primarily driven by data centers. Meeting it requires both new generation and transmission investments.

The federal government has begun pushing for clearer connection rules. In June 2026, the Federal Energy Regulatory Commission issued orders to six regional grid operators.

FERC directed each operator to justify or reform its procedures for connecting data centers, factories, and other large electricity users. The commission wants large loads integrated in a timely, orderly, and equitable way.

The FERC action can improve process transparency. It cannot instantly create transformers, power plants, transmission corridors, or public acceptance.

That limitation is driving developers toward on-site generation, also called behind-the-meter power. Such systems serve a facility directly instead of relying entirely on electricity delivered through the public grid.

Options include natural gas generation, fuel cells, solar arrays, batteries, and combinations of those technologies. Some companies are also exploring nuclear power, although most new nuclear proposals will not solve projects facing immediate delivery deadlines.

On-site generation can shorten one dependency while creating others. Developers still need fuel access, equipment, air permits, environmental reviews, and local approval. They may also face higher operating costs or emissions than anticipated.

The power problem therefore cannot be reduced to one slow utility queue. It is a shortage spanning generation, wires, industrial equipment, permitting, and credible forecasts.

Permits Turn Local Politics Into an AI Infrastructure Constraint

The AI build-out depends on local decisions made far from corporate product teams and national technology policy.

A company can announce billions in capital spending at its headquarters. The resulting campus still requires approvals from planning departments, zoning boards, environmental agencies, utilities, and elected officials where construction will occur.

Each authority evaluates a different issue. Zoning officials examine land use. Environmental reviews can cover water, air pollution, wetlands, and backup generation. Transportation departments assess construction traffic. Utilities study the proposed electrical connection.

A project can move through these reviews sequentially, making one delay affect every following contract. A developer may hesitate to order expensive equipment before receiving approval, but waiting can push the project behind other customers in manufacturing queues.

Community concerns are also becoming more organized. Residents have challenged projects over electricity rates, water consumption, noise, diesel generators, land use, and the number of permanent jobs created after construction ends.

These objections do not follow a simple partisan pattern. Communities with different political profiles have questioned whether the local economic benefits justify infrastructure costs and environmental effects.

The conflict is sharper because the benefits and costs appear at different scales. AI companies can use computing capacity nationally or globally. The affected community experiences the construction traffic, transmission lines, water demand, and potential utility impacts locally.

Faster federal procedures cannot erase that mismatch. A national policy favoring AI infrastructure still passes through state utility rules and local land-use authority.

Permitting also differs widely by jurisdiction. One location may provide a coordinated approval process, while another requires separate reviews by city, county, state, utility, and private landowners.

Developers consequently value administrative certainty alongside electricity and land. A predictable 18-month review can be easier to finance than a nominally shorter process with no reliable decision date.

This changes the competitive geography of AI. Established hubs offer fiber, contractors, and an existing supplier base, but they also face crowded grids and rising opposition. New regions may offer available land and supportive officials, but their infrastructure can be less mature.

Companies are attempting to reduce exposure through phased development. Instead of waiting for an entire campus, they can open one building or power block and add capacity later.

Phasing protects schedules, but it weakens some economies of scale. Separate construction stages can repeat mobilization costs and require temporary utility arrangements. A later phase can still lose its place in a power queue.

Another response is to select locations near existing generation. Co-location places a large customer physically or electrically close to a power plant, potentially reducing pressure on the wider transmission system.

That approach raises difficult allocation questions. If a data center contracts directly with an existing generator, the surrounding grid can lose power it previously expected to serve other customers.

Regulators must then decide which network costs belong to the data center and which remain with households and businesses. These cost questions explain why “bring your own power” is not always a complete answer.

The permit constraint also pressures developers to communicate earlier. Quiet land acquisition once helped companies avoid speculation. A secretive approach now risks generating distrust when residents discover a project late in the process.

Public engagement cannot guarantee approval, especially where electricity or water is already constrained. It can reveal issues before a company commits to a site plan that officials will reject.

For hyperscalers, the lesson is uncomfortable. Money can accelerate engineering and procurement, but it cannot buy unlimited political consent.

Skilled Labor and Equipment Keep Moving Completion Dates

Even a permitted site with an electricity plan cannot open without enough electricians, pipefitters, engineers, and specialized electrical equipment.

Data center construction competes for workers with factories, energy projects, semiconductor plants, hospitals, transportation systems, and conventional commercial development. The overlap is particularly strong for electrical and mechanical trades.

These campuses need workers who can install medium-voltage systems, cooling loops, backup generators, fire suppression, controls, and dense network infrastructure. Commissioning teams must then test how those systems behave under failures and changing loads.

That work is more specialized than assembling a conventional warehouse. A shortage cannot always be solved by adding general laborers or transferring an inexperienced crew.

Associated Builders and Contractors estimates that the U.S. construction industry must attract 349,000 additional workers in 2026 to keep labor supply aligned with demand. That figure covers the wider industry, not data centers alone.

However, the workforce estimate shows the labor market from which data center contractors must recruit. It also excludes the normal hiring needed to replace workers who permanently leave the industry.

The shortage becomes more severe when several campuses rise in the same region. Contractors may have enough staff for one large project but not enough to support overlapping schedules from multiple hyperscalers.

Workers can travel, yet that response adds lodging, transportation, and retention costs. Remote projects can struggle to maintain crews through years of phased construction.

Training programs offer a longer-term solution. Apprenticeships can expand the pool of electricians, plumbers, welders, and other trades while providing paid work experience.

Training still takes time. An apprenticeship cannot produce an experienced high-voltage electrician before a campus scheduled to open next quarter.

Equipment creates a parallel bottleneck. Large transformers, switchgear, generators, turbines, and cooling components require specialized manufacturing. Some items have long production queues and limited substitutes.

A transformer adjusts voltage so electricity can move efficiently and then serve equipment safely. Large custom units are not interchangeable consumer products that developers can buy from any warehouse.

A delayed component can hold back an entire energized section of a campus. The building may look complete from outside while it remains unable to power servers.

This helps explain why satellite or aerial images require careful interpretation. A cleared site does not prove that construction is proceeding on schedule. An apparently finished structure does not prove that its electrical and cooling systems have passed commissioning.

The risk also runs in the opposite direction. Developers sometimes order equipment before final approvals because waiting would make the target date impossible.

Early ordering protects a schedule if the project advances. It creates financial exposure if permits, financing, or power arrangements change.

Companies with several projects can redirect standardized equipment. Highly customized systems are harder to move, especially when regional electrical specifications or facility designs differ.

Supply-chain exposure can also reach outside the United States. Domestic AI policy often focuses on semiconductor manufacturing, yet a functioning campus depends on a much broader industrial base.

Chip availability remains important, but accelerators produce no useful computing output without transformers, substations, cooling, networking, and workers. The constraint has shifted from one scarce component to a connected chain of scarce resources.

This is the core reversal behind the slowdown. Technology companies optimized software deployment around rapid iteration. Physical infrastructure requires sequenced work, certified equipment, inspections, and skilled labor that cannot be compressed indefinitely.

Still, the evidence does not justify claiming that half of all announced capacity will disappear. Development pipelines often contain speculative sites, duplicates, and projects scheduled across several years.

A delay can also mean many things. A campus may open one phase late, reduce its initial power draw, move to another market, or wait for a utility upgrade. None of those outcomes equals permanent cancellation.

The strongest conclusion remains narrower: execution risk is rising, and headline capital commitments no longer provide a reliable measure of computing capacity delivered on time.

What the Next Three Signals Will Reveal

The next stage of the AI data center boom will be judged by energized capacity, not announcements or construction spending alone.

The first signal is the amount of capacity that moves from active construction into operation. CBRE’s next market updates should show whether the 2025 decline was a temporary pause after a record surge or the start of a longer delivery slowdown.

Vacancy and absorption must be read alongside that figure. If construction rebounds while vacancy remains low, demand is still pulling projects forward. If vacancy rises and absorption falls, weaker customer demand becomes a more credible explanation.

The second signal is grid-connection reform. FERC’s orders require regional operators to defend or change how they study large loads. The meaningful result will be clearer milestones, stronger financial commitments, and fewer speculative requests.

Faster review alone is not sufficient. A queue can move quickly and still conclude that a region lacks generation or transmission capacity.

Watch PJM because it combines extraordinary projected demand with the country’s largest concentration of data centers. Its forecasts, connection rules, transmission plans, and capacity-market results will show how costs are distributed.

The third signal is the gap between corporate capital expenditure and operating capacity. Major technology companies continue to commit enormous sums to AI infrastructure, but spending includes chips, networking, land, buildings, and long-term contracts at different stages.

Investors should distinguish money committed from servers installed, energized, and available to customers. A delayed building can preserve reported investment plans while postponing the revenue or productivity expected from that capacity.

Developers will also reveal their confidence through procurement. More long-term power contracts, equipment reservations, contractor partnerships, and on-site generation plans would show that they expect scarcity to persist.

Conversely, widespread deferrals, smaller initial phases, or land sales would weaken the view that this is only a delivery problem.

Government action deserves similar scrutiny. Federal regulators want quicker large-load integration, while communities want protection from higher costs and environmental effects. Those priorities can coexist only if rules clearly assign infrastructure costs and operational risks.

The national AI strategy therefore depends on decisions that sound less glamorous than model releases. Transmission planning, apprenticeship enrollment, transformer production, zoning calendars, and utility rate design now shape available compute.

Google News coverage will continue producing eye-catching stories about individual delays, moratoriums, and massive campuses. Readers should resist treating each item as proof of either an AI collapse or unlimited expansion.

The verified figures describe a market with record occupancy and a smaller active construction pipeline. That is a constraint story until demand indicators say otherwise.

For developers and enterprise buyers, the practical question is no longer whether more AI capacity has been announced. Ask when that capacity will receive power, whether its permits are final, and who will build it. Then track the energy forecast and regional connection decisions against actual openings. Those checks will reveal whether the United States is converting its AI ambitions into usable infrastructure, or accumulating projects that exist mainly in plans.