The Grid Can't Keep Up With the Model: Siting Power for the AI Era
Written by Bruno P. Allaire
AI data center demand has collided with a grid built for the last century. The winners won't be whoever announces the biggest campus, they'll be whoever can deliver power on data center timelines instead of utility timelines.
Share this perspective
The Demand Shock
Every so often an industry asks the electric grid for something it has never been asked for before. The AI buildout is doing it annually. Individual campuses are now requesting capacity on the order of a small city's peak load, hyperscalers are describing their pipelines in gigawatts rather than megawatts, and the requests keep arriving while the previous ones are still being studied. A planning apparatus built to accommodate load growth of one or two percent a year, deliberated over rate cases and forecast decades, is suddenly staring at demand curves that bend like a hockey stick.
The numbers get quoted a lot, so I'll add only the framing that matters for what follows. Large-load interconnection requests have exploded, and a meaningful share of the announced capacity is speculative, the same campus shopped to three utilities as a siting option. The demand side of the equation is real regardless: EPRI projects data centers consuming between 4.6 and 9.1 percent of all US electricity by 20301. But even discounted aggressively, the underlying load is real, it is concentrated in a handful of regions, and it is arriving on a schedule the grid cannot meet with its normal machinery. The result is the first genuinely binding physical constraint the technology industry has hit in a generation. Compute scales. Power does not.
That constraint is now the organizing fact of data center siting. Anyone who has worked a board meeting on a new facility knows the agenda has flipped: it used to open with fiber, water, and incentives. Now it opens with one question. When can we get power, and how firm is the answer?
Why the Interconnection Queue Is the Bottleneck
The instinctive assumption is that the constraint is generation, that we simply are not building enough of it. That is partially true and mostly beside the point. The binding constraint in most load-growth regions is the interconnection queue: the process by which a new load or a new generator gets studied, and connected, to the transmission system. In key markets that process now runs four to seven years, and the national queue peaked at nearly 2,600 gigawatts of active generation and storage requests at the end of 2023, roughly twice the installed capacity of the entire US fleet2. A data center that breaks ground this year on a standard grid interconnection may be waiting past the end of the decade for full service.
The queue is slow for structural reasons, not bureaucratic ones. Each large request triggers system impact studies across a transmission network that was engineered for a different era, and each study has to model what every other pending request would do to the system. The more requests pile in, the slower each study gets, which pushes more developers to file earlier and more speculatively, which piles more requests in. It is a self-reinforcing backlog, and no amount of urgency at the utility end dissolves it, because the physical upgrades a large interconnection requires, new lines, new substations, new voltage support, take years to site and build even after the studies clear. The thinning happens mostly on the way in: of the projects that requested interconnection between 2000 and 2019, only about 19 percent, representing roughly 13 percent of capacity, had reached commercial operation by the end of 20243.
The practical consequence is that "when can I get power?" has become the first question in every serious siting conversation, and the honest answer is usually a date that kills the project's economics. A model release cycle is measured in months. The grid's answer is measured in years. Any developer who reconciles those two clocks without the utility has found the central trick of this era.
The Behind-the-Meter Playbook
That is why the most interesting work in this space is happening behind the meter. If the queue cannot deliver service on the compute schedule, the developer brings the generation to the load: on-site gas turbines and reciprocating engines, fuel cells, islanded microgrids, and hybrid arrangements that bridge a campus from energization until the grid connection matures. What was once a niche resilience strategy has become the default opening move in serious AI-era siting.
I have spent enough of my career on the generation and midstream side of this equation to say this plainly: behind-the-meter power is not a hack, it is an underwriting discipline. The questions that determine whether a bridged project works are the same ones I have asked of every energy asset I have ever built. Is there firm fuel supply at a contractable price, or are you exposed to basis blowouts at exactly the wrong moment? Is the generation equipment procurable inside the window, in a market where turbine lead times have stretched past the interconnection queue itself? Can the air permit be secured in the state and county where the campus sits, on a timeline that matches the energization date? A developer who can answer those three questions holds something rarer than a gigawatt of announcements.
The playbook also has a fiscal logic that surprises people the first time they run it. Behind-the-meter generation avoids a share of the transmission and distribution charges that come with grid service, it can be scaled in increments that match actual load ramp rather than a single large utility commitment, and it converts the campus's largest cost, firm power, from a variable under someone else's control into an asset on its own balance sheet. The bridge frequently pencils as permanent infrastructure. It is why we formed Defiant Power Solutions: modular natural-gas generation, parallelable from a single 450-kilowatt unit to more than a hundred megawatts, deployable in as little as seventy-two hours, so firm power can arrive on the compute's schedule rather than the queue's.
Co-Location Beats the Queue
Which leads to the stronger form of the idea: stop treating the generation as a bridge and treat it as the design. Co-locating generation directly with the load, on the same site or across the fence line, collapses the siting problem. No queue position for the load itself, no dependence on a distant upgrade schedule, and a power quality story that a hyperscaler's engineers can verify at the equipment rather than take on a utility's word.
The structural advantages compound. Co-location removes transmission loss and the congestion exposure that comes with it. It pairs naturally with heat recovery, since a data center is, among other things, a very expensive cold-weather machine, and a turbine across the fence is a heat source that monetizes fuel twice. And it gives the campus an energy profile that can be tuned: gas-fired firm capacity today, with room on the site for whatever the fuel mix and the politics support in the next decade.
I have a bias here worth admitting, because I have lived the alternative. Building midstream and gas-fired infrastructure through a commodity trough teaches you exactly how long energy promises take to become electrons, and how much of the announced is really just the studied. The co-located structure works because it strips out every party whose timeline you cannot control. You are left with fuel supply, equipment, and permits, and those are the only three clocks an energy operator ever trusted anyway.
What Breaks, and When
None of this is frictionless, and the frictions are worth naming honestly, because they determine which regions and which sponsors actually deliver. Gas pipeline capacity is the first constraint to bind. Several of the most attractive load-growth markets are also the most pipeline-constrained, and a co-located campus without firm transport is a stranded investment waiting for a cold week. Turbine and engine lead times are the second: order books have stretched across the industry, and a supplier's promised delivery date is, as every energy executive learns eventually, an opening position. Emissions permitting is the third, and the most political: load-growth states are writing rules in real time, and a permit strategy that worked last year may not survive the next legislative session.
There is also a systemic risk that deserves more attention than it gets. Bridged power is quietly becoming permanent. Campuses designed around temporary on-site generation have a way of never surrendering it, which means a meaningful share of the AI buildout's emissions will come from infrastructure that was never planned, modeled, or permitted as part of the long-term resource picture. Regulators will eventually catch up, and the sponsors who ran their bridges to a documented standard, with emissions controls and a defined exit, will be treated very differently from the ones who improvised.
The failures, when they come, will follow a recognizable pattern: a campus announced at a scale the fuel supply couldn't support, energized on equipment that arrived late, in a jurisdiction where the permit assumption was outdated by eighteen months. The announcements will not tell you which projects those are. The interconnection filings and the fuel contracts will.
The Siting Window
So who wins this era? Not, I would argue, whoever announces the largest campus. The winners are the operators who control the three scarce inputs simultaneously: generation sites with credible fuel access, procurement positions that secure equipment before the queue clears, and the permitting experience to move through state and county processes without a default setting of litigation. Each of those is a quiet asset. None of them photographs well. All of them compound.
The advantage compounds because every delivered project deepens the moat. A sponsor who has energized one co-located campus has the supply chain, the permitting record, and the operating data to price the next one accurately, while competitors are still assembling their first. The gap between the operators and the announcers widens with each cycle, quietly, while the headlines chase renderings.
For the energy executive watching this from the outside, the window is the thing to understand. The grid will eventually catch up: new transmission, new firm generation, and interconnection reform will re-normalize the relationship between compute and the utility. But that normalization is a decade out in the markets that matter, and the capital deployed in the meantime will lock in the winners of this era. The grid can't keep up with the model. For a while yet, the model will run on whatever the energy industry can put behind its fence line, and the rewards will go to those who can put it there on schedule.
Notes & Sources
- 1.EPRI projects that data centers could consume between 4.6% and 9.1% of all US electricity by 2030, up from roughly 4% today. Electric Power Research Institute, "Powering Intelligence: Analyzing Artificial Intelligence and Data Center Energy Consumption," 2024
- 2.US interconnection queues peaked at nearly 2,600 GW of active generation and storage capacity at the end of 2023, roughly twice the installed capacity of the entire US generation fleet, before easing to roughly 2,060 GW by the end of 2025. Lawrence Berkeley National Laboratory, "Queued Up," 2026 edition
- 3.Of the generation and storage projects that requested interconnection between 2000 and 2019, only about 19% of projects, representing roughly 13% of capacity, had reached commercial operation by the end of 2024. Lawrence Berkeley National Laboratory, "Queued Up: 2025 Edition"