Power as the Binding Data Center Constraint

Introduction

For most of commercial real estate, the scarce input is well-located land. For data centers, it is electricity, and the difference reshapes how the entire asset class is developed, valued, and financed. A site can have land, fiber, capital, and a signed tenant, and still be worthless for years if the local utility cannot deliver power. That is no longer a hypothetical edge case. Grid access has overtaken land and permitting as the primary constraint on data center construction, and the practical question a developer now asks first is not "can I build here" but "how many years until this site is energized." The broader demand picture behind that pressure is laid out in the data center demand surge, but the supply side comes down to one word: power.

The core problem is that demand for new electricity is growing far faster than the grid can add and connect generation. The interconnection process, by which a new power plant or large load is studied and approved to connect, has become a multi-year bottleneck. In PJM, the largest United States grid operator, the timeline from interconnection request to commercial operation has stretched from under two years in 2008 to more than eight years by 2025, as the operator worked through a backlog exceeding 170 gigawatts of generation requests. Nationally, the average queue time has roughly doubled to about five years.

Time-to-Power

The elapsed time between securing a data center site and the point at which sufficient, reliable utility power is actually energized and available. It has become the single most important development variable in the sector, often determining a project's value more than land cost or construction timeline.

The capacity math is now openly stressed. PJM came up roughly 6,625 megawatts short in its December 2025 capacity auction, with capacity prices hitting a record $333.44 per megawatt-day, and the operator expects the summer of 2027 to be the first season it lacks sufficient capacity, with data center growth outpacing new generation by a factor of two to one. In Texas, ERCOT saw large-load interconnection requests rocket to more than 230 gigawatts in 2025, nearly four times the level a year earlier. Even a single utility shows the strain: Dominion Energy's pipeline of contracted data center capacity climbed from roughly 40 gigawatts in early 2025 to 47.1 gigawatts by that October, pushing connection timelines out further.

The constraint is global, not American. Amsterdam, Dublin, and Singapore have all at various points restricted or paused new data center connections because the facilities were absorbing too large a share of local grid capacity, pushing growth toward Nordic, Spanish, and Italian markets where power and land remain available. Wherever AI capacity wants to land, the local grid, not the local real estate market, is the gating factor, which is why site selection has become as much an energy exercise as a property one.

The Equipment Bottleneck Behind the Grid Bottleneck

Even where power exists, the hardware to deliver it has become its own constraint. Lead times for large power transformers, the units that step grid voltage down for a data center, have stretched from roughly 24 to 30 months before 2020 to as long as four to five years, with prices up around 77% since 2019. Switchgear, circuit breakers, and other transmission and distribution gear face the same squeeze as data centers, utilities, and broader electrification projects compete for the same factory capacity. The shortage is severe enough that, by some 2026 estimates, more than half of planned US data centers could be delayed or canceled not for lack of money or land but for lack of the electrical equipment needed to energize them. For an underwriter, that means a project's timeline now depends on a procurement queue for transformers and switchgear that can rival the interconnection queue in length, and a developer who locked in equipment orders years early holds an advantage as real as a queue position.

When the constraint is power rather than land, how efficiently a facility turns its grid draw into useful computing becomes a competitive variable in its own right. The standard measure is Power Usage Effectiveness (PUE), the ratio of total facility energy to the energy that actually reaches the IT equipment. Every watt spent on cooling, power conversion, and distribution rather than on servers shows up as overhead above a perfect score of 1.0:

A facility running at a PUE of 1.5 burns half again as much grid capacity as it delivers to its chips, which in a power-constrained market is half again as many scarce megawatts consumed for the same compute. That is why the most efficient operators, pushing PUE toward the 1.1 to 1.2 range, can stand up more revenue-generating IT load behind a fixed interconnection allocation than a rival stuck at 1.4 or 1.5, turning efficiency into a way to ease the binding constraint rather than merely a cost line.

Because energization is so uncertain, time-to-power has become the variable that drives a project's value. A site with secured, near-term power is worth a large premium over an identical site stuck deep in an interconnection queue, and that premium can dwarf differences in land cost. Hyperscalers and colocation developers routinely expect power one to two years earlier than utilities believe they can deliver it, and that expectation gap is where deals are won and lost.

The path to energization is a sequence, and each step can stall:

That sequence is why a developer's interconnection-queue position is treated almost like a separate asset, and why it feeds directly into the build-cost and time-to-power math and into the valuation of any site. It is also why landlords with energized capacity command the kind of pricing power that has pushed wholesale lease rates up 20% or more a year.

That same logic has redrawn the map of where data centers get built. Site selection is now power-first: developers follow available megawatts rather than population or fiber, which is why the marquee AI campuses have landed in gas-rich or grid-spare locations like Abilene in West Texas, Richland Parish in Louisiana, New Albany in Ohio, and Memphis rather than in the legacy coastal hubs. A region that can offer firm power on a near-term timeline can attract billions of dollars of investment almost overnight, and one that cannot will watch that capital route around it regardless of how attractive its land or tax incentives look on paper.

The response to grid delay has been for developers to stop waiting. The phrase that now dominates the sector is Bring Your Own Power: securing on-site or near-site generation so a project does not depend entirely on a constrained grid connection. Tracking shows more than 40 gigawatts of announced behind-the-meter and co-located generation across the country.

Behind-the-Meter Generation

Power generated on the customer's side of the utility meter, consumed directly by the data center rather than drawn from the grid. It lets a project energize in months rather than the years a grid interconnection can take, at the cost of building and operating generation the developer would otherwise leave to the utility.

The technology mix reflects a near-term-versus-long-term split, and the options differ sharply in speed, cost, and carbon profile:

Natural gas, in the form of combined-cycle plants, reciprocating engines, and aeroderivative turbines, dominates near-term behind-the-meter power because it can be deployed quickly. For the longer term and to meet clean-energy commitments, the hyperscalers have turned to nuclear: Microsoft, Amazon, Alphabet, and Meta have collectively signed more than 10 gigawatts of nuclear power-purchase agreements with operators such as Constellation, Vistra, and Talen Energy, alongside a growing roster of small modular reactor developers. Some developers are also chasing stranded power, existing generation that has not been fully allocated, which is driving data center growth into new geographies. This is the point at which data center coverage overlaps heavily with the work of the energy investment banking group, since the deals increasingly involve generation, PPAs, and grid infrastructure rather than buildings alone.

The Gas-and-Fuel-Cell Bridge

The near-term reality is that most bring-your-own-power capacity is fossil-fired, because gas is what can be built in months rather than years. Demand for on-site generation has been so intense that GE Vernova's gas-turbine backlog reached 100 gigawatts in the first quarter of 2026, up 17 gigawatts from year-end, with management guiding the backlog past 110 gigawatts by the end of the year. Large-frame turbines are now effectively sold out through 2028, which has pushed developers toward reciprocating engines and, increasingly, fuel cells as a faster alternative. Bloom Energy has emerged as the marquee fuel-cell supplier, with its product backlog up roughly 2.5 times year-over-year on data center demand alone.

The fuel mix is therefore a near-term-versus-long-term split: gas and fuel cells to energize now, nuclear to decarbonize later, and a continuing scramble for any stranded or under-allocated generation in between.

The marquee nuclear deals

A handful of landmark transactions have become the reference points for the entire clean-baseload push. In late 2024, Constellation Energy agreed to restart Unit 1 of the shuttered Three Mile Island plant, rebranded the Crane Clean Energy Center, under a 20-year power-purchase agreement that dedicates its full 835 megawatts to Microsoft's data centers. Constellation is investing roughly $1.6 billion to bring the reactor back online by 2028, the first time a retired US nuclear plant has been revived for a single commercial buyer. Amazon contracted with Talen Energy for up to 1,920 megawatts from the Susquehanna plant through 2042, ramping to full volume by the early 2030s, after buying the data center campus adjacent to the plant outright. Meta signed its own 20-year agreement with Constellation for the 1.1 gigawatt Clinton plant in Illinois. Beyond restarts, the same buyers are funding small modular reactors that will not deliver until late this decade at the earliest, a wager on baseload that does not yet exist at commercial scale. The common thread is that hyperscalers, not utilities, are now the marginal buyers setting the price of clean power, and they are increasingly willing to fund the generation directly. For the real estate, the implication is stark: a platform's portfolio of energy contracts and queue positions is becoming the asset, and the buildings the wrapper around it.

From Plant Restarts to Reactors That Do Not Exist Yet

The restarts are only the near-term half of the nuclear push. The longer-dated bet is on small modular reactors (SMRs), factory-built units that promise faster, cheaper deployment than conventional gigawatt-scale plants but that will not deliver power until late this decade at the earliest. Google signed the first US corporate SMR fleet agreement with Kairos Power, covering up to 500 megawatts across six to seven reactors, with the first unit targeted for around 2030. Amazon invested roughly $700 million in X-energy and committed to bringing more than 5 gigawatts of its reactors online across the United States by 2039, on top of the Susquehanna campus. Meta has been the most aggressive of all, assembling up to 6.6 gigawatts of nuclear commitments spanning plant restarts and next-generation reactors, including TerraPower's Natrium, Oklo's Aurora, Vistra, and Constellation, and issuing an open request for proposals for 1 to 4 gigawatts of additional new nuclear. Taken together with the restarts, the hyperscalers have contracted well over 15 gigawatts of nuclear capacity in roughly a year, a market that did not exist before AI demand created it.

The headline nuclear commitments map cleanly by buyer:

The significance for a banker is less about any single reactor than about who is now funding generation. Utilities historically built power plants and sold the output; today a handful of technology companies are underwriting reactor restarts, pre-ordering SMR fleets, and signing twenty-year offtake agreements that make the projects financeable in the first place. That reversal, with the load taking on the generation risk, is the clearest sign that power, not real estate, now sits at the center of the data center investment case.

The fastest way to get power is to plug a data center straight into an existing power plant, bypassing the congested grid entirely. That arrangement, known as co-location, became the sector's biggest regulatory battleground. In November 2024, FERC rejected an expanded interconnection agreement that would have let Amazon draw more power from the Susquehanna nuclear plant beside which Talen Energy had sold it a data center campus, on the grounds that co-located load raised unresolved questions about who pays for grid reliability and how costs are shared. The rejection froze a financing model much of the sector had been counting on.

The fight resolved over the following year. Talen and Amazon restructured their arrangement in June 2025 into a 17-year, roughly $18 billion power-purchase agreement for up to 1,920 megawatts from Susquehanna, reworked as a front-of-the-meter deal that routes through the grid rather than around it, with the transition expected to complete in spring 2026. Then, on December 18, 2025, FERC issued a unanimous order directing PJM to write clear rules for co-locating large loads at power plants, creating new transmission-service options and reforming behind-the-meter generation terms. The order effectively reopened the co-location path the 2024 rejection had closed, but on regulated terms.

The regulatory fight resolved in a clear sequence:

States are moving in parallel, and the most consequential is Texas. Its Senate Bill 6, enacted in 2025, gives ERCOT authority to order large loads with on-site backup generation to curtail or switch to that backup during grid emergencies, requires new large loads after the end of 2025 to install equipment for remote disconnection during firm load-shedding events, and makes facilities of 75 megawatts or more share interconnection costs and pay upfront study fees. The practical effect is to turn data centers into a curtailable, flexible load rather than a guaranteed draw, and to push developers toward the on-site generation and battery storage that let them ride through curtailment. SB6 is widely treated as a bellwether for how other states will regulate data center demand as it collides with grid reliability.

Power as the binding constraint changes the entire diligence checklist. The first questions on any data center mandate are now about electricity: How much power is contracted, when does it energize, who pays for the interconnection, and is there a behind-the-meter backstop if the grid slips. A platform's value increasingly rests on its portfolio of secured power and queue positions, not just its buildings, and a development pipeline without committed power is a pipeline of options, not assets. The financing follows the same logic, with collateral commitments, availability payments, and long-term PPAs now standard features of how these projects are capitalized, a structure that informs the valuation of any data center.

Power has displaced land as the binding constraint, and the consequences run through everything downstream: time-to-power sets a site's value, an interconnection-queue position can be worth more than the dirt beneath it, and a development pipeline without committed electricity is a book of options rather than assets. The deeper shift is what the constraint has done to the developers themselves. Securing gas today and nuclear tomorrow has turned data center owners into de facto energy companies, negotiating twenty-year PPAs and funding reactor restarts, and the platforms that pull ahead are the ones that show up to a site with power already in hand rather than a place in the queue.