Power First, GPUs Second: Building the Right BTM Strategy

Data Centers & Energy By EnergyInnoCore Editor · Published Oct 16, 2026 · Updated Oct 18, 2026 · 12 min read

Grid interconnection delays have become a serious drag on AI data center deployment. Utility queues are pushing energization out by years, regional transmission systems are struggling to absorb gigawatt-scale requests, and utilities are more cautious about committing capacity. For developers chasing first-mover advantage, simply waiting for the grid is no longer a comfortable strategy.

The numbers explain the urgency. PJM's long-term forecast shows about 30 GW of peak load growth from data centers between 2024 and 2030, while new generation is arriving far more slowly. PJM has acknowledged that permitting, siting and supply-chain hurdles can make a new gas-fired plant take two or three times as long to develop as a hyperscale data center.

That mismatch has turned behind-the-meter (BTM) generation from an unusual alternative into a mainstream development strategy. Instead of waiting years for the utility to deliver the full power requirement, the developer builds some, or all, of the generation at or near the site. The idea sounds simple: if the grid cannot bring power fast enough, bring your own. In practice it is much more complicated. BTM shortens the road to operation only when the generation technology, permitting strategy, electrical architecture, fuel infrastructure and data center load are designed together. Otherwise the developer may simply trade one bottleneck for another. (For a basic introduction, see our BTM planning primer.)

When I evaluate BTM power for an AI data center, I put three questions ahead of almost everything else:

1. Time-to-power: how quickly can it actually deliver dependable megawatts?
2. Technical compatibility: can it follow the load the data center will put on it?
3. Fuel and total cost of ownership (TCO): will the fuel supply and long-term economics still make sense after the rush to energize is over?

1. Time-to-Power: The Schedule That Matters Most

Installed cost per kilowatt, efficiency and emissions all matter. But for many developers, time-to-power (TTP) has become the first screening criterion. A data center without power is not an operating asset: servers earn nothing, committed customers wait, financing costs keep accumulating, and expensive GPUs may be sitting somewhere other than where they are needed.

That changes how BTM should be evaluated. The question is not "How long does it take to install a generator?" It is "From site control today, how long until dependable megawatts are flowing to IT equipment?" That timeline includes much more than the generation equipment.

Air permitting can decide the winner before equipment is ordered

For combustion-based BTM systems, air permitting often sits directly on the critical path. Depending on size, configuration, technology, emissions profile and location, a project may face state and local permits under the Clean Air Act framework, including New Source Review and potentially Title V operating requirements. EPA notes that most data center air permits are administered by state and local agencies, which makes geography surprisingly important. Two technically identical 300 MW BTM projects in different states can have very different schedules because of differences in permitting procedures, local air quality, modeling requirements, public review, agency staffing and community response.

EPA has taken steps intended to clarify and speed up parts of the process. Guidance issued in 2025 clarified that certain construction activities unrelated to emissions units may begin before a Clean Air Act construction permit is issued, under specific conditions, and the agency has published data center resources as developers consider large amounts of on-site generation. In July 2026, EPA also stated that the Acid Rain Program does not apply to generating facilities that are not connected to the public grid, which may give truly islanded projects more flexibility.

None of this means permitting can be treated as an administrative task to handle after technology selection. It should be the opposite. The emissions profile and permitting pathway should be examined while the generation architecture is being chosen, because the wrong choice can erase months or years of supposed TTP advantage.

Then comes the equipment queue

There is another uncomfortable reality: everyone else has discovered BTM too. Gas turbines, reciprocating engines, transformers, switchgear, breakers and emissions-control equipment are all in high demand. GE Vernova's figures illustrate the scale. As of the second quarter of 2026, its gas power backlog and slot reservation agreements had reached about 116 GW, up from 100 GW a quarter earlier, and the company said it expected at least 125 GW under contract by year-end. Data centers are helping drive that backlog.

Specifying the theoretically most efficient turbine means little if it cannot arrive until long after the building is finished. Developers should therefore ask questions that traditional utility projects rarely prioritized:

• When is the manufacturing slot available, and how firm is it?
• Which long-lead components are included, and when will transformers and switchgear arrive?
• Is factory testing capacity available?
• Can the project be phased?
• Are there existing equipment packages or reserved slots that could be acquired?

Sometimes the fastest BTM solution is not the one with the lowest theoretical levelized cost; it is the one that can actually be delivered. That is partly why modular aeroderivative turbines and reciprocating engines are attracting attention. GE Vernova's agreement with Crusoe, for example, covers 29 LM2500XPRESS aeroderivative packages expected to provide nearly 1 GW for AI data centers. The units can run independently or together, which supports staged deployment. Being able to energize the first 100 MW while another 200 MW is still under construction can be enormously valuable, which is why TTP should be measured in commercially usable megawatts rather than final project completion.

2. Load Dynamics: A Data Center Is Not Just a Big Factory

A common trap in BTM development is assuming that a 200 MW generator can power a 200 MW AI data center. Not necessarily. Nameplate capacity answers only one question; how the generation system responds to the load matters just as much.

Traditional data centers have fairly stable load profiles, but AI training is changing that. Large GPU clusters move rapidly between computational states, and power demand rises and falls as workloads synchronize, pause, communicate and resume. Measurements and industry presentations to ERCOT have shown AI training loads with both second-scale and millisecond-scale variations, in some cases reaching large percentages of peak load. That behavior can cause voltage fluctuations, frequency deviations and oscillations, and can push on-site generators outside their normal operating envelope. (We explain the phenomenon in our article on AI load transients.)

ERCOT considers the issue significant enough that in 2025 it partnered with Texas A&M to develop dynamic models of how large electronic loads such as data centers respond to grid disturbances, and it has requested voltage ride-through information from large data center and crypto loads of 75 MW or more.

For BTM developers this matters enormously. Imagine a 500 MW campus running mainly on on-site generation. If the compute load suddenly drops by 80 MW, that energy imbalance has to go somewhere; if it returns moments later, the system must respond in the other direction. A conventional generator is a large rotating machine with inertia, governor response characteristics, minimum stable operating levels, ramp-rate limits, combustion constraints and protection settings. GPUs do not care about any of those things.

Architecture matters as much as generation

This is where system design becomes critical. A BTM plant serving AI infrastructure may need a combination of generation, battery energy storage (BESS), UPS systems, power electronics and microgrid controls, rather than relying on generators to chase every movement of the IT load. A battery can absorb or release energy almost instantly while slower generation follows the broader trend. In that role the battery is more than backup; it is a shock absorber between the AI load and the power plant.

There is also a strong case for modularity. Ten 30 MW units may behave very differently from one 300 MW block: they can run closer to efficient operating points, provide redundancy, accommodate maintenance and expand with the campus. But modularity has costs too, including more equipment, controls, auxiliaries, maintenance points, potentially more emissions sources and a more complex electrical system.

That is why a BTM design cannot be developed by the generation team alone. Power engineers need to understand the IT load, data center electrical engineers need to understand the generators, and the controls team needs to understand both. Developers should request realistic load profiles early: not only expected peak megawatts but ramp rates, step loads, redundancy philosophy, allowable voltage and frequency excursions, UPS architecture, power factor, GPU deployment schedule and expected workload behavior. This is especially important for facilities that will operate in island mode. When connected to a strong grid, the grid acts as an enormous stabilizer. In a fully islanded configuration that cushion disappears, the BTM plant becomes the grid, and that is a very different engineering problem.

3. Fuel Infrastructure and Long-Term TCO

Once TTP and technical compatibility are addressed, the third question is economic, and it needs to go beyond simple capital-cost comparisons. BTM is often promoted with a straightforward calculation: equipment costs X per kilowatt, gas costs Y per MMBtu, so at a given heat rate electricity costs Z per megawatt-hour. It makes a tidy spreadsheet, but a real data center does not run inside a spreadsheet.

Cheap generation needs dependable fuel

A 500 MW or 1 GW on-site plant requires a large, continuous fuel supply. If natural gas is the primary fuel, the first question is not whether a pipeline crosses somewhere near the site. It is whether firm transportation capacity exists to supply the project when it is needed most. Pipeline diameter and pressure, compression, lateral construction, interconnection and metering, upstream capacity, competing customers, seasonal constraints, and firm versus interruptible service all affect whether supply is truly dependable.

The gas connection can become a miniature utility project of its own. A developer may save three years on the electrical interconnection only to find that a new lateral, compressor station or upstream expansion has a long permitting and construction schedule, so one queue has simply replaced another. Fuel security must also match the facility's reliability target. High-availability applications may need dual-fuel capability, on-site storage, multiple pipeline connections, redundant pressure regulation or alternative backup generation. Each improves resilience, and each costs money.

Total cost of ownership, not just capex

A realistic 15- or 20-year TCO model should include fuel escalation, transportation and pipeline reservation charges, operations and maintenance, major overhauls, emissions-control consumables, water, auxiliary power, degradation, replacement reserves, insurance, staffing, financing, potential carbon costs and eventual integration with the grid. Heat rate is especially important for a baseload plant: a small efficiency difference looks insignificant during development, but across hundreds of megawatts running about 8,000 hours a year for 15 years it becomes very significant.

Yet the most efficient option may conflict with TTP. A large combined-cycle plant offers attractive efficiency but usually takes longer and is more complex to build than modular simple-cycle turbines, engines, fuel cells or hybrid systems. That leaves the developer with a real trade-off: how much long-term efficiency is worth sacrificing to gain two years of revenue? There is no universal answer. For some projects, bringing 200 MW online 24 months earlier is worth far more than decades of fuel savings. For a campus expected to run continuously for 20 years, waiting a little longer for better efficiency may create more lifecycle value. The right analysis combines TTP and TCO instead of optimizing either alone.

BTM may become a bridge, not an island

Today's BTM solution does not have to remain tomorrow's. Some developers will start mainly or entirely on on-site generation and connect to the grid when transmission becomes available. Others will keep on-site generation permanently alongside utility service, renewable PPAs, storage and, eventually, nuclear. Grid operators are moving in the same direction: in January 2026, PJM's board endorsed pathways for large customers to bring their own new generation or accept curtailment in exchange for earlier connection, and FERC has pushed regional operators to reform their large-load rules.

That suggests the smartest BTM architecture may be a bridge rather than a permanently isolated system: bring the first 200 MW online quickly, add another 200 MW as equipment arrives, connect to the grid when capacity is available, add storage where it creates value, and keep efficient on-site generation for resilience, peak management or economic dispatch. The objective is optionality.

The Real Question Is Not "Which Technology?"

Discussions about BTM power quickly turn into technology comparisons: gas turbine or reciprocating engine, fuel cell or turbine, simple cycle or combined cycle, battery or no battery. Those are important questions, but they usually come too early. I would start somewhere else:

1. Fix the date of the first commercially usable megawatt, then work backward through permitting, equipment availability, construction, fuel infrastructure, commissioning and site approvals.
2. Understand what the AI load will actually do to the generating system, not just "300 MW" on a slide but the real electrical behavior of several hundred megawatts of GPUs on an islanded or semi-islanded system.
3. Calculate what it costs to keep the system running for 15 to 20 years, including the pipeline, fuel transportation, maintenance, overhaul reserves, storage, emissions controls and degradation, and the economic value of getting online earlier.

Only then should the developer choose which technology belongs behind the meter. The best BTM solution is not necessarily the cheapest, cleanest or most efficient generator. It is the one that delivers dependable power when the data center needs it, copes with the fast-changing electrical behavior of AI computing, and keeps making economic sense long after the race to energize is over. For today's AI infrastructure developers, that combination can be the difference between having a power strategy and actually having power.

EnergyInnoCore Editor

Independent analysis of power, grid and policy for AI data centers. Every article lists its sources. About the editor · Editorial policy

This article is for general information only and is not engineering, legal, financial or investment advice. See the disclaimer.