AI Collides with the Grid: Transformers Missing for Data Centers

The fever for building artificial intelligence data centers in the United States is hitting a less glamorous but much more difficult obstacle than GPUs: electrical infrastructure. Having capital, land, clients, and AI models isn’t enough. Without transformers, electrical panels, batteries, permits, and grid connections, the announced megawatts don’t translate into operational capacity.

This is already evident in the project pipeline. According to estimates gathered by Bloomberg and specialized media, between one-third and half of the data centers planned in the U.S. for 2026 could be delayed or canceled. Of the peak capacity of 12 to 16 GW announced for this year, only about 5 GW are actually under construction. This gap shows how the bottleneck has shifted from financial to physical constraints.

The problem is not announcing gigawatts; it’s energizing them

Over the past two years, major cloud operators, infrastructure funds, and real estate developers have announced a wave of projects to meet AI demand. The logic was simple: more models, more inference, more training, more data centers. But execution has been slower because each project depends on an electrical supply chain that was already strained before the AI boom.

Data centers don’t just need electricity in theory. They require reliable connection, substations, high-capacity transformers, switches, panels, busways, uninterruptible power supplies, batteries, generators, permits, and agreements with utilities. If one piece is delayed, the entire schedule shifts.

Transformers are the clearest example. Before 2020, certain units could be obtained within two years or so. Now, wait times for large units easily surpass 24 months and can approach four years in some markets. POWER Magazine reports data from Wood Mackenzie indicating average power transformer lead times of about 128 weeks, while GSU units take around 144 weeks. pv magazine USA mentions delays of up to four years in a highly tight market.

The economic impact is straightforward. A data center project can have secured financing, land contracts, and committed clients, but if it doesn’t receive electrical equipment on time, it can’t deliver capacity. In an industry where each month of delay can mean millions in missed revenue, the transformer ceases to be just an engineering component and becomes a financial variable.

The grid is not growing at the same pace as capital expenditure

Investment in AI infrastructure now reaches hundreds of billions of dollars. But the electric grid isn’t expanding as quickly as corporate budgets. A GPU may be purchased, even if expensive. A large transformer involves a much slower industrial chain. A transmission line can take years to approve, build, and connect. A substation requires permits, materials, engineering, and coordination with the local utility.

JLL’s late-2025 data center report highlights that grid connection times—averaging four years or more—are changing how hyperscalers and large clients choose locations. The traditional criteria of proximity to markets, fiber, or major hubs remain relevant, but the new primary question is: where is energy available, when can it be connected, and who bears the cost of upgrades?

The EIA projects that electricity demand from data center servers will grow within the U.S. commercial sector. Under the high-demand scenario, independent data center servers could consume 581,000 million kWh by 2050, with electricity use rising from about 7% of commercial consumption in 2025 to between 22% and 33% in 2050.

The International Energy Agency offers another perspective: global data center electricity consumption was around 415 TWh in 2024 and could double to approximately 945 TWh by 2030 in its baseline scenario. This global figure shouldn’t be confused with local pressures. The data center challenge isn’t just total electricity consumption but where and how quickly they need to connect.

This concentration explains why regions like Northern Virginia, Dallas, Phoenix, Georgia, and parts of Texas frequently appear in discussions. A power system may seem solid at the national level but have very specific bottlenecks at nodes where everyone wants to build.

Politics enters the electrical room

The second front is social and political. Local opposition to data centers is no longer insignificant. Communities raise concerns about impacts on utility bills, water use, noise, land, tax benefits, and public resource usage. In some areas, the message is clear: if a facility requires new infrastructure, the costs shouldn’t fall on consumers’ bills.

New York has just sent an important signal. The state Senate approved Bill S10642, the Responsible Data Center Development Act, with 43 votes in favor and 17 against. The bill establishes a one-year moratorium on permits for large data centers—defined as facilities of 20 MW or more—and requires reports on impacts related to electricity, water, land, pollution, and public incentives. It also proposes specific rate classes to ensure these centers shoulder the additional connection and service costs.

If signed, this measure won’t halt all U.S. growth but sets a precedent. Until now, many moratoria were local or municipal. A statewide pause on large projects signals that the debate has moved from city councils to regional legislatures.

For developers and investors, this shifts the calculus. The risk isn’t just about acquiring equipment or grid connection anymore; it’s also about convincing communities and regulators that the data center adds value—through jobs, taxes, services, grid upgrades, clean energy, heat reuse, water efficiency, or local investment commitments.

The industry’s response can’t be just “AI needs more data centers.” That phrase isn’t sufficient when communities fear higher bills or see massive facilities demanding priority over other uses. Social license becomes as vital as contracted capacity.

Efficiency becomes a financial strategy

A key strategic conclusion is uncomfortable for many AI roadmaps: not all announced capacity will arrive on time. Some facilities will be delayed, others will change location, some will be resized, and some projects will be canceled. Planning must account for a computing capacity that may not materialize as expected by planned dates.

The alternative isn’t to stop building but to do so more smartly. If megawatts become scarce, the winners will be those who get the most work done per watt, rack, transformer, and dollar of capex. Efficiency shifts from a sustainability argument to a financial one.

This logic encourages multiple approaches. The first is software: smaller models, compression, quantization, caching, intelligent routing, well-managed MoE, and reducing inference waste. The second is hardware: more efficient accelerators, better-utilized memory, and architectures designed for tokens per watt—not just max FLOPS. The third is infrastructure: modular designs, 800 VDC systems, batteries, local generation, renewables agreements, and grid planning from day one.

This approach can also accelerate geographic shifts. Data centers tend to move toward locations with available energy, clearer permits, better community relations, and the ability to build substations within reasonable timelines. Traditional hubs will remain relevant for connectivity and demand but can’t absorb everything if the grid isn’t aligned.

Capital alone no longer guarantees capacity. This marks a significant difference from earlier cloud phases. Previously, companies with sufficient funds could rent or build more infrastructure within predictable timeframes. Today, they compete for electrical equipment, specialized labor, permits, water, land, connections, and social acceptance.

AI has highlighted a fundamental truth of the digital economy: the virtual world depends on a physical supply chain. Models don’t run on investment announcements but on racks powered by transformers, lines, substations, and backup systems. If that chain breaks, the capacity promise remains just that—a promise.

For companies planning AI products, the key question isn’t whether more data centers will be built—it’s whether they’ll be available where, when, at what cost, and at what density their models depend on. If the answer is uncertain, the backup plan should start earlier: reduce compute costs per unit of work and design AI that doesn’t rely on infinite megawatt expansion.

Frequently asked questions

Why are data centers delayed in the U.S.?

Due to a combination of a lack of electrical equipment, lengthy grid connection times, permitting delays, energy pressures, and local opposition. Large transformers, switchgear, batteries, and substations are among the most cited bottlenecks.

How much capacity is actually under construction?

Estimates from specialized sources place the announced capacity for 2026 between 12 and 16 GW, with only about 5 GW actively being built.

How important are transformers?

They are essential for connecting and adapting the electricity required by a data center. Without transformers and associated equipment, the building can exist but can’t operate at the planned power level.

What does the New York moratorium entail?

The bill approved by the state Senate sets a one-year pause on permits for large data centers of 20 MW or more, along with requirements for public hearings, impact reports on electricity, water, land, pollution, and incentives. It also introduces specific rate classes to allocate connection and infrastructure costs to these centers.