McKinsey Data Center Study Finds Overbuilding Is Not the Biggest Risk

McKinsey has challenged fears of an AI infrastructure glut, despite projecting that US data center IT demand could reach 121 gigawatts by 2030. The McKinsey data center study argues that committed demand remains strong enough to support continued construction. The sharper risk is that developers build projects on paper but cannot connect, power, or complete them where customers need capacity.

That distinction changes the overbuilding debate. A long project pipeline does not necessarily mean the United States will end up with too many operating data centers. Multiple developers sometimes pursue the same limited grid connection, while speculative proposals can remain years from construction. Meanwhile, completed capacity in major markets is being leased almost as quickly as it becomes available.

The result is an infrastructure paradox. The industry can look overbuilt when measured by announcements, yet remain severely under capacity when measured by usable megawatts. Utilities, hyperscalers, colocation operators, equipment suppliers, and local governments must now decide which projects are real enough to support.

The McKinsey Data Center Study Reframes the Buildout

McKinsey sees sustained demand colliding with a delivery system that cannot expand at the same speed.

The firm estimates that data centers will account for approximately 75% of expected US power-demand growth during the next decade. Current construction ambitions would require almost 30 gigawatts of additional power each year. That figure includes about 20 gigawatts of incremental IT demand, plus electricity for cooling, distribution, and reliability.

The firm’s overbuilding analysis does not dismiss the possibility of weaker AI demand. Instead, it tests what the power system would need under different adoption scenarios. Even its lower-demand case requires substantial additions to electricity supply by 2030.

McKinsey’s central argument rests on the quality of current demand. Large hyperscalers have strong balance sheets, substantial operating cash flow, and multiyear infrastructure plans. Many data center investments are also tied to contracted utilization instead of purely speculative occupancy forecasts.

That makes today’s cycle different from a conventional property boom. Developers are not simply building generic warehouses and hoping tenants arrive. Much of the most credible capacity already has a customer commitment, a power strategy, or both.

Demand indicators still contain weaknesses. McKinsey notes that enterprise AI execution remains uneven, even when companies express strong interest. Its survey found that about one-third of respondents were scaling AI, while 39% reported a measurable effect on earnings before interest and taxes.

The same survey found that 64% cited innovation benefits. However, 71% of organizations reported at least one negative implementation outcome. Those figures suggest that AI adoption has moved beyond experimentation, but not every project is producing dependable business value.

Some financing patterns also resemble a bubble. More than half of the new unicorns created during 2025 focused on AI, according to McKinsey. Several reached multibillion-dollar valuations before commercializing a product. Other infrastructure intermediaries are expanding while carrying high leverage or negative cash flow.

Those risks matter because compute demand ultimately depends on viable applications. If AI revenue disappoints, weaker developers and marginal tenants will struggle first. Yet that correction would not automatically eliminate demand from well-capitalized cloud companies or resolve the shortage of deliverable capacity.

The study therefore shifts the central question. Instead of asking whether every announced campus will be needed, planners should ask which campuses can secure power, equipment, customers, permits, and financing simultaneously.

Announced Capacity Is Not the Same as Usable Capacity

The apparent overbuild lives largely in project pipelines, while the capacity shortage exists in operating facilities.

A proposed data center can appear in several databases long before it becomes a functioning site. The developer might control land but lack a utility agreement. Another project may have entered a grid study without financing, equipment orders, permits, or an anchor tenant.

Developers also submit overlapping requests when several locations could serve one requirement. They can then advance whichever site obtains power first. Counting every request as independent demand inflates the apparent construction pipeline.

Early-stage firms have another incentive to reserve access. A developer can secure land or interconnection rights, then sell the project to a larger operator. Some proposals therefore represent options on scarce infrastructure rather than fully financed construction programs.

These practices explain why a large queue can coexist with limited supply. Announcements measure ambition. Operational megawatts measure capacity that customers can actually use.

Commercial property data reinforces that gap. CBRE reported that supply across North America’s primary data center markets grew 33.7% year over year during the first half of 2026. Yet vacancy still fell to a record 1.4%.

The firm’s market findings show 7,481.1 megawatts under construction across those markets, a 24.8% annual increase. More than 80% of that capacity was already preleased. Less than 1,500 megawatts remained available, equal to roughly six months of demand at the current absorption rate.

Northern Virginia illustrates the mismatch. It remained North America’s largest data center market, with 4,496.5 megawatts of inventory. Its vacancy rate fell to 0.2%, even as more than 2,400 megawatts remained under construction.

Atlanta passed Northern Virginia in construction activity, reaching roughly 2,882 megawatts. That shift does not show demand evaporating in Virginia. It shows developers moving toward locations where land, power, and approvals offer a better route to completion.

Preleasing provides a stronger demand signal than a press release. A customer that reserves future capacity makes a contractual commitment before the facility opens. Those agreements vary in strength, but they offer more evidence than an uncommitted development plan.

They also limit the risk of indiscriminate construction. Lenders and investors increasingly examine tenant credit, lease terms, power availability, and construction milestones before releasing capital. Projects without those foundations face higher financing barriers.

Still, preleasing does not eliminate risk. A financially weak tenant can fail, and a heavily leveraged operator can encounter refinancing problems. Customer concentration can expose a facility to one company’s changing AI strategy.

Construction can also outrun equipment installation. A building shell does not deliver useful compute without servers, networking hardware, cooling systems, transformers, and energized connections. A project can look physically advanced while remaining months from service.

This is why “overbuilt” and “under capacity” can both describe the same market. The industry has accumulated more proposed projects than it will complete. At the same time, customers face very little immediately available space in the locations they prefer.

Power Has Become the Real Capacity Gate

The decisive constraint is no longer whether developers want to build, but whether the electricity system can support their schedules.

McKinsey expects US data center IT demand to grow around 27% annually through 2030. Its modeled range reaches between 121 and 137 gigawatts of incremental IT demand by the end of the decade.

Serving that load requires more generating capacity than the IT figure alone suggests. Data centers also need cooling, power conversion, redundancy, and other supporting systems. Generation resources cannot all operate continuously at their stated nameplate capacity.

McKinsey estimates that the United States has roughly 40 gigawatts of dispatchable headroom, meaning generating capacity above steady demand. About 100 gigawatts of additional capacity has been committed for construction.

Those figures initially sound sufficient. However, coal and older steam-gas plant retirements are expected to remove between 50 and 75 gigawatts by 2030. After projected demand growth and retirements, McKinsey identifies a national capacity need of approximately 30 to 55 gigawatts.

National totals also conceal local shortages. Data center demand is concentrating in Northern Virginia, Phoenix, Louisiana, Texas, and emerging Midwestern markets. Power cannot be moved freely between every region, especially when transmission lines are congested or absent.

Building major transmission infrastructure can take more than a decade. Data center schedules are usually much shorter. A hyperscaler purchasing accelerators today cannot assume that a proposed power line will arrive in time for its next deployment cycle.

The International Energy Agency sees the same collision globally. Its updated energy outlook projects data center electricity consumption rising from 485 terawatt-hours in 2025 to 950 terawatt-hours in 2030. AI-focused facilities grow considerably faster than the overall market.

The IEA also found that electricity consumption at AI-focused data centers increased 50% during 2025. That recorded growth offers firmer evidence than distant construction announcements. It shows that installed AI hardware is already increasing load.

Server density raises the delivery challenge. According to the IEA, AI server power density increased elevenfold between 2020 and 2025. It expects another fourfold increase by 2027. Higher density concentrates more electrical and cooling demand inside each facility.

That trend can improve compute output per building, but it creates new engineering requirements. Operators need high-capacity electrical distribution, liquid cooling, suitable transformers, and equipment that can handle rapid fluctuations in AI workloads.

Equipment production cannot adjust instantly. Transformers, switchgear, turbines, and specialized cooling systems have long manufacturing lead times. Suppliers must expand factories before knowing which announced projects will reach construction.

Grid planners face a similar forecasting problem. They must prepare for large loads before facilities begin operating, yet not every request represents a distinct project. Overestimating demand can burden ratepayers with unnecessary infrastructure. Underestimating it can delay investment and threaten reliability.

The McKinsey data center study therefore describes more than an electricity shortage. It identifies a coordination failure among industries operating on different timelines. AI companies plan in quarters, data centers develop over years, and major grid projects can take a decade.

Hyperscalers Can Build Faster Than Utilities Can Respond

The pressure falls on utilities and regional grid operators because hyperscaler spending can move faster than regulated infrastructure planning.

McKinsey says hyperscaler capital spending increased approximately 50% annually from 2022 through 2026. The largest cloud companies can finance campuses, reserve chips, sign power agreements, and shift workloads between regions at a scale unavailable to most developers.

Utilities cannot respond with equal speed. They must forecast demand, seek regulatory approval, study system effects, obtain land rights, purchase equipment, and recover costs through approved structures. Each step can create delays or disputes.

Regional grid operators also need better evidence that proposed loads will appear. PJM, which covers much of the Mid-Atlantic, has emphasized the danger of double-counting data center requests. Several utilities can receive inquiries related to the same underlying project.

This creates a difficult choice. If grid planners assume every request will become real, they risk buying excess capacity. If they discount too many requests, they can enter a delivery year without enough reliable supply.

Generation queues provide a warning about treating applications as completed infrastructure. A Berkeley Lab analysis counted about 1,400 gigawatts of proposed US generation and 890 gigawatts of storage at the end of 2024.

Only 13% of capacity entering interconnection queues from 2000 through 2019 had reached commercial operation by the end of 2024. Another 77% had withdrawn, while 10% remained active. The median development period for completed projects had also grown beyond four years.

These queues concern power generation rather than data center loads, but the lesson transfers. An application represents intent, not delivery. Project quality, financing, permitting, equipment, and interconnection progress determine which proposals become operating assets.

Utilities are responding through new tariffs and contract structures. Some want large-load customers to provide stronger financial security or accept minimum payment obligations. Those measures can protect other customers when a planned data center never materializes.

Hyperscalers are also pursuing direct power agreements. Nuclear restarts, renewable contracts, gas generation, batteries, and fuel cells can supplement conventional utility service. These arrangements give large buyers more control, but they do not remove transmission, permitting, or equipment constraints.

Smaller operators face greater pressure. They cannot always fund generation alongside a data center, and they carry less negotiating leverage with suppliers. They may depend on colocation providers or choose secondary markets with more accessible power.

Enterprise buyers should therefore treat location as a reliability variable. A capacity reservation in a constrained region carries different delivery risk from an operating deployment. Contract reviews must distinguish scheduled completion, energized capacity, and contingency arrangements.

Cloud customers may feel the effects indirectly. Capacity constraints can influence where new AI instances become available, how quickly reserved clusters start, and whether providers impose longer commitments. Geographic flexibility can become as important as the choice of accelerator.

The competitive divide will not simply separate companies that spend the most from those that spend less. It will separate operators that control a credible power path from those holding construction plans without deliverable energy.

On-Site Power Solves Speed but Creates New Risks

On-site generation can shorten the wait for electricity, but it transfers fuel, equipment, emissions, and reliability risks to the data center.

McKinsey’s 2025 survey of power-industry leaders found that 65% expected to deploy some form of on-site power. The survey included 145 respondents, with 46 US experts representing hyperscalers, colocation operators, and utilities.

Nearly 60% expected permanent on-site generation by 2030, even after grid service became available. Among respondents planning on-site power, 64% expected to use natural gas. That suggests behind-the-meter generation is becoming part of normal data center design.

Gas engines and turbines can provide firm output without waiting for a major transmission expansion. Fuel cells and batteries can support reliability or bridge limited grid service. Solar paired with storage can contribute energy where space and operating conditions allow.

However, self-generation does not make infrastructure constraints disappear. Gas turbines are subject to manufacturing backlogs. A site may also need new pipeline capacity, redundant fuel access, air permits, and specialized maintenance.

The IEA estimates that reliable on-site gas systems can require 30% to 70% more generating capacity than the data center’s expected load. Operators need that margin because equipment can fail or undergo maintenance. Variable AI workloads add another layer of complexity.

The agency projects that between 15 and 27 gigawatts of on-site natural-gas generation might serve data centers by 2030, mostly in the United States. It also cautions that only around one-fifth of tracked projects had begun land clearing or construction when it assessed them.

Batteries offer speed and flexibility, but duration matters. Short-duration storage can smooth power swings and cover brief interruptions. It cannot independently supply a large AI campus through an extended grid outage.

A virtual power plant, which coordinates distributed energy resources through software, can help shift demand or supply electricity during constrained periods. Yet it depends on market rules, communications, customer participation, and the availability of those distributed resources.

Data centers could also adjust workloads. Some training tasks can move between times or regions, reducing demand during grid stress. McKinsey notes that operators have shown limited willingness to curtail valuable AI computing, because compute revenue can outweigh electricity costs.

Inference workloads create a different challenge. Inference means running a trained model to answer requests or perform tasks. Customer-facing services often require low latency and continuous availability, leaving less room to pause consumption.

Environmental tradeoffs will become harder to avoid. Extending coal plants or adding on-site gas can help meet near-term demand, but both can increase emissions. Those decisions can conflict with corporate clean-energy targets and community expectations.

Water, noise, land use, and electricity affordability can also trigger local opposition. A technically viable project can still fail if residents or regulators conclude that its public costs exceed its benefits.

The McKinsey data center study does not establish one universal power solution. It points toward mixed portfolios shaped by local conditions. Grid service, gas generation, solar, storage, fuel cells, and workload flexibility will play different roles across markets.

The risk is that temporary measures become permanent without adequate scrutiny. A gas plant installed to bridge a grid delay can operate for decades. Utilities and regulators must therefore evaluate speed alongside long-term costs, emissions, and stranded-asset exposure.

What the Overbuild Case Still Gets Right

Durable aggregate demand does not guarantee that every developer, tenant, region, or financing structure will succeed.

The strongest case against overbuilding comes from contracted capacity, low vacancy, and rising electricity consumption. The strongest skeptical case focuses on what happens inside those aggregate numbers.

AI demand forecasts depend on continued improvements in models and sustained use of compute-intensive services. Efficiency gains could reduce the power required for each task. Businesses could also become more selective after early deployments fail to produce adequate returns.

Efficiency does not always reduce total consumption. Cheaper inference can encourage developers to place AI features in more products, increasing the number of requests. This rebound effect makes the net result difficult to forecast.

McKinsey expects growth to slow after 2030. Chip demand could become less intense, adoption might mature, and operators may replace existing hardware with more efficient systems. That transition could weaken demand for new buildings while supporting upgrades inside established facilities.

EPRI’s 2026 projections show how wide the uncertainty remains. It estimates US data centers used between 177 and 192 terawatt-hours in 2024. Its 2030 scenarios range from 383 to 793 terawatt-hours.

That spread is too large to support one deterministic construction plan. Utilities must invest before they know whether demand follows the low, medium, or high path. Developers face the same uncertainty when committing capital to projects with long payback periods.

The risk also differs across market participants. A hyperscaler can redirect servers, slow one campus, or absorb unused capacity elsewhere in its network. A speculative developer with one site and significant debt has fewer options.

Infrastructure intermediaries sit between those extremes. They may hold long-term power obligations while relying on customers whose credit quality is weaker than a major cloud provider’s. A mismatch between lease revenue and energy commitments can become expensive.

Geography creates another source of stranded assets. National demand can grow while a particular site fails. A campus might lose because its grid connection arrives late, local rules change, or customers prefer another region with lower latency.

Technology design can also strand capacity. Facilities built for conventional air-cooled servers may require extensive retrofits for higher-density AI systems. Electrical distribution and cooling architecture can limit which hardware a site supports.

This is why the overbuild concern remains reasonable, even if the broad market stays tight. Some projects are likely to disappear from pipelines. Others will open later than announced, change ownership, or operate below their planned scale.

The mistake is treating those failures as proof that aggregate AI infrastructure demand collapsed. A healthier interpretation is that scarce power and capital are filtering projects. The buildout can experience severe local losses while remaining undersupplied overall.

Investors and enterprise customers should examine contracted utilization, tenant quality, energized capacity, equipment readiness, and power-delivery dates. Announced megawatts alone reveal little about project durability.

Three Signals Will Decide Whether the Shortage Persists

The next stage of the AI data center buildout will be determined by delivered capacity, verified load, and enterprise AI economics.

The first signal is the gap between construction and energized supply. CBRE’s next market updates should show whether record construction raises vacancy or disappears into preleased demand. Rising supply paired with persistently low vacancy would strengthen McKinsey’s under-capacity thesis.

A sharp increase in available inventory would weaken it. That change would matter most if it appeared across several major markets rather than one delayed campus. Watch operating capacity and net absorption, not project announcements alone.

The second signal is how utilities vet and serve large-load requests. Stronger deposits, minimum-payment requirements, and clearer milestones would help remove duplicated or speculative projects from forecasts. They would also reveal which customers are willing to make binding commitments.

Grid additions matter just as much. More generation entering commercial operation, faster transmission approvals, and measurable interconnection progress would reduce the risk of local shortages. Continued queue delays would reinforce the view that delivery, not demand, defines the bottleneck.

The third signal is enterprise AI value. Revenue growth, scaled deployments, and measurable productivity gains would support sustained compute consumption beyond hyperscaler training programs. Repeated implementation failures or shrinking AI budgets would weaken the demand case.

Enterprise results deserve more weight than model announcements. A new benchmark leader can trigger attention, but recurring workloads pay for infrastructure. Data center demand becomes more durable when AI systems enter customer service, software development, research, operations, and other continuing processes.

Readers should also distinguish training from inference. Training demand can arrive in large, concentrated cycles. Inference grows with recurring product usage and can spread across more locations. A shift toward inference would change facility design without necessarily reducing aggregate power needs.

For technology buyers, the practical question is no longer whether an AI infrastructure boom exists. It is whether a provider controls usable capacity in the right region, with credible power and completion dates.

Ask vendors which capacity is operational, which is merely contracted, and what happens if a grid connection slips. Examine geographic alternatives and workload portability before relying on one future campus.

The McKinsey data center study ultimately presents a more complicated picture than either a bubble or a permanent shortage. The United States has an enormous development pipeline, but only part of it will become usable infrastructure.

Demand can remain strong while individual projects fail. Construction can set records while customers struggle to obtain capacity. The decisive test is whether promised megawatts become powered computing systems before AI economics or financing conditions change.