The marathon: Can behind-the-meter data center power solutions endure the decade in the US?

The first article in this series examined how the US data center sector is assembling a parallel energy system to close the gap between facility construction timelines of 12 to 24 months and grid interconnection timelines that stretch to seven years or more. The BTM pipeline now exceeds 130GW of proposed energy resources across the US, with gas accounting for more than 80% of the total. The market is responding decisively to the need for speed: It is selecting technologies that can deliver power within the data center construction window, accepting BTM power costs that are higher than grid-connected alternatives, and treating the result as a bridge to eventual grid service.[1]

But a bridge implies a destination. For a growing share of the pipeline, that destination is becoming less certain. Grid interconnection timelines are not compressing fast enough to match data center construction cycles, and equipment originally deployed for three to seven years of bridge duty could be asked to operate longer. In some cases, it may even become the primary power source for facilities with demand profiles that persist for a decade or more.

Longer durations change the nature of the power-source decision. A five-year BTM solution can be justified as the cost of speed. A ten- or fifteen-year solution starts to look more like an operating strategy. A twenty- or thirty-year asset becomes an infrastructure commitment, tied to a specific site, a specific fuel source, a specific regulatory regime, and a specific demand profile. In other words, duration changes the economics. It also changes the risk.[2]

This is the shift at the center of the “marathon.” The sector is moving away from a speed-to-power market, where the winning solution was often the one developers could secure, and toward a portfolio-selection market. In this new context, the winning solution is the one that preserves optionality, fits the location, and still makes sense if the path to grid power changes. That transition is best understood through three lenses: How duration impacts the cost of BTM power, how the market is segmenting in response, and how geography and regulation will determine which strategies endure.

[1] RaboResearch, The sprint: Data centers are building a parallel energy system in the US, May 2026

[2] FERC, 2025 State of the Markets Report, March 2025; Lawrence Berkeley National Lab, Queued Up, 2025

The key cost question is no longer whether BTM power is more expensive than grid supply on day one. It is whether the technologies selected for speed still make economic sense when the operating horizon extends from five years to 15 or 20.

At year five, the cost difference between BTM gas generation and grid-connected alternatives is narrower than it might appear. Gas reciprocating engines deliver power at roughly USD 107/MWh, aeroderivatives (gas turbines adapted from jet engine technology) at USD 106, and combined-cycle plants at USD 101/MWh, compared with a national average industrial grid rate of approximately USD 86/MWh (figure 1).[3] At that duration, the premium reflects the price of speed. The question the market has been slower to ask is how those costs evolve by year 15 or year 20, especially if the bridge period has not yet ended and the equipment is required to keep operating.

[3] RaboResearch SCDE Model, 2026; U.S. Energy Information Administration, average retail electricity price for industrial consumers, 2025 (USc 8.62/kWh). Grid rates vary significantly by market.

The sustained cost of delivered energy

To answer that question, this analysis introduces the sustained cost of delivered energy (SCDE), a metric that measures the all-in cost per megawatt-hour over progressively longer operating periods. Unlike standard levelized cost estimates, SCDE reflects the way BTM assets actually age. It escalates fuel and operating expenses for inflation and incorporates the periodic maintenance, overhauls, and fleet rebuilds required when gas equipment runs continuously at high capacity factors.[4]

The findings reshape the cost comparison. Gas reciprocating engines and aeroderivatives, the technologies deployed most widely in the “sprint,” look increasingly cost-effective through the early years as upfront capital costs are absorbed. That improvement starts to break down around year 10 (see figure 1). At high capacity factors, these units accumulate operating hours quickly, pulling them into recurring overhaul cycles and major rebuilds roughly once a decade. In our model, both technologies show the same pattern: Costs fall into the USD 80/MWh range by year 9, reset to around USD 100/MWh around year 10, improve again through year 15, and then step higher again around years 20 and 30.[5] The result is not a smooth long-run decline, but a sawtooth cost curve. That matters because these are the technologies most associated with bridge power. They perform well when the operating horizon is temporary, but their economics become less compelling when bridge duty turns into long-duration primary power.

Combined-cycle plants tell a different story (see figure 1). They start at a cost similar to gas engines at year five, at roughly USD 101/MWh versus USD 107/MWh, but their cost profile improves more steadily as the operating horizon extends. By year 15, combined-cycle SCDE falls to roughly USD 55/MWh and continues declining toward USD 50 by year 30 (see figure 1). The reason is straightforward: Combined-cycle plants are built for sustained baseload operation, with higher thermal efficiency and maintenance cycles better suited to long-duration use. The crossover with gas engines occurs around year 10. By year 20, the combined-cycle plant costs are roughly one-third lower.[6]

What gas-fired technologies can’t escape is fuel exposure. Even with stable natural gas prices, fuel represents a persistent cost floor that no amount of capital amortization can eliminate. Heavy-duty simple cycle turbines (the lower-cost gas option at year 15 at roughly USD 58/MWh) continue to decline modestly to USD 55/MWh by year 30 as capital costs spread further. However, this floor remains above where renewables settle. A sustained decline in natural gas prices would lower the floor for all gas technologies, but it would not remove the fundamental asymmetry that gas equipment carries a fuel cost in every year of operation, while solar and wind do not.

Renewables face no fuel constraint. Solar PV reaches roughly USD 49/MWh at year 15 and USD 34/MWh by year 30 (see figure 1). Onshore wind follows a similar trajectory. However, renewables have lower capacity factors than thermal generation, and their output does not match the continuous, high-availability load profile data centers require. Meeting that demand with renewables alone would require significant overbuild, storage, or firm backup capacity. In hybrid configurations paired with gas, storage, or grid supply, however, their long-run cost advantage becomes more valuable over time. In that sense, renewables are less a stand-alone solution than a hedge against the fuel exposure embedded in long-duration gas reliance.

The strategic implication is that the market's near-term equipment choices and its long-term cost interests are not fully aligned. Engines and aeroderivatives are winning the “sprint” because they can be deployed quickly and retain backup or redeployment value once grid service arrives. Combined-cycle plants are better suited to the “marathon,” but their scale, permitting requirements, and equipment lead times make them less useful for immediate energization. Renewables have the most favorable long-run cost trajectory, but they require firm generation or grid supply to meet the reliability profile data centers need. The fastest solution is not necessarily the most durable, and the lowest-cost long-run solution is not necessarily available on the timeline the market requires. The best strategy is therefore not a single technology choice, but a portfolio decision shaped by duration, location, and exit path.

[4] RaboResearch SCDE Model, 2026. Methodology: capex amortization plus cumulative operating and maintenance, fuel, and periodic maintenance events, divided by cumulative MWh delivered. Assumes: CPI 2.5%, natural gas price USD 3.5 per million British thermal units (MMBtu), nuclear fuel (uranium) USD 6.5/MWh. Fixed/variable operating and maintenance, inflated to 2025 USD; Capex reflects BTM scale. Nuclear costs are based on first-of-a-kind (FOAK) estimates.

[5] Caterpillar publishes maintenance intervals for natural gas reciprocating internal combustion engines, and GE Vernova’s GER-3620 report outlines gas turbine maintenance schedules; together, these sources underpin assumptions in the SCDE model. Rebuild costs reflect new generation units plus balance-of-plant and engineering upgrades.

[6] RaboResearch SCDE Model, 2026. All reported cost figures are on an unsubsidized basis.

The BTM pipeline is often discussed as a single market trend. That framing is increasingly misleading. The commitments being made across the pipeline reflect at least three distinct strategies, each with a different investment thesis, a technology profile, and answer to the question of what happens when the bridge is no longer needed.

Bridge power

Bridge power represents the majority of the current pipeline. Providers such as Enchanted Rock, VoltaGrid, and Scale Microgrids are structuring energy-as-a-service contracts with built-in termination provisions, and the equipment being selected (primarily gas reciprocating engines and aeroderivative turbines) reflects that intent. Of the roughly 40GW of gas-based capacity where equipment class has been specified, aeroderivatives and small simple-cycle turbines account for 36%, and engines for 26%, while combined-cycle plants represent a similar 26% share.[7] The modular end of the fleet is being built for temporary use. The heavy end is being built for long-term operation.

Hybrid power

Hybrid power refers to configurations that supplement limited or conditional grid access rather than replacing it entirely. This model is gaining traction as the Federal Energy Regulatory Commission (FERC) and regional grid operators develop new frameworks for partial grid reliance. Hybrid strategies use on-site generation and battery storage to manage peak withdrawals, absorb curtailment risk, and bridge the gap between available grid capacity and full facility demand. They do not require the grid to supply 100% of the facility’s electricity demand, nor do they require the facility to generate 100% itself. The value proposition is flexibility, not independence.

Long-duration private power

Long-duration private power refers to embedded generation infrastructure designed to operate as the primary energy source for a decade or more, in locations where the grid may never be the preferred answer. This is the narrowest segment, but it represents some of the largest individual commitments in the pipeline. Crusoe Energy's Project Jade in Laramie County, Wyoming, is being developed with dedicated gas pipeline infrastructure delivering 419m cubic feet per day of firm capacity directly to a 1.8GW campus, scalable to 10GW. In Doña Ana County, New Mexico, the Green Chili lateral will deliver 400m cubic feet per day to a planned 4GW campus with on-site gas and renewable generation.[8] These are not merely data centers with power attached. They are integrated energy-and-compute developments where fuel access, land, generation, and load are planned as a single thesis.

This three-part segmentation matters because the strategies carry fundamentally different risk profiles. Bridge power is a timing play with manageable downside. Hybrid power is a flexibility play where value depends on whether conditional grid frameworks mature. Long-duration private power is a conviction bet, with the most exposed downside, because heavy-duty equipment is site-specific, capital-intensive, and difficult to redeploy if the thesis changes.

[7] S&P Global Energy, Colocating Data Centers with Energy Resources Help Plug the Power Supply-Demand Gap in North America, March 2026

[8] Rockies Express Pipeline LLC, Switchgrass Lateral Filing, FERC

The locations where long-duration private power is emerging share a common profile. They tend to have limited existing generation capacity and a relatively underdeveloped transmission network (so-called low generation and transmission depth), along with available land, accessible natural gas infrastructure, and state-level policy that enables non-utility generation. Northern Nevada, Wyoming, and parts of New Mexico and West Texas fit this description. Without enabling policies for on-site generation, data center capacity expansion would likely be 40% to 60% lower in Texas (ERCOT), the Northwest, and the Rockies, where grid constraints would otherwise be binding.

The pattern is that sustained BTM reliance becomes more defensible where the grid was relatively constrained to begin with. In legacy hubs – long-standing, established data center markets, such as Northern Virginia and Silicon Valley – the grid infrastructure is more developed, but permitting and regulatory environments are more restrictive. As a result, the long-duration private power model is far less likely to take hold. The geography of data center growth is being reshaped by power availability, not just connectivity, proximity to users, or tax incentives. That represents a structural shift, favoring markets that most participants were not watching five years ago.

This does not mean legacy hubs lose their importance. Northern Virginia, Silicon Valley, Phoenix, Dallas, and other established markets will remain critical for latency-sensitive workloads, enterprise connectivity, and existing cloud ecosystems. However, power availability is changing the marginal decision. The next increment of AI capacity is more likely to follow energy depth rather than historical data center clustering alone. That shift does not erase the old map of data center development, but it adds a second one: defined by fuel access, transmission scarcity, land availability, and state-level willingness to let large electricity users (so-called large loads) generate their own power or manage their own supply instead of relying entirely on the grid.

The BTM buildout exists because the grid cannot deliver power fast enough. That condition has not changed. FERC's Order No. 2023 reduced speculative queue volume by roughly 10% nationally, with sharp drops of 59% in the Pennsylvania-New Jersey-Maryland Interconnection (PJM) region, 53% in California (CAISO), and 66% in New York (NYISO). However, these reforms have not translated into faster processing. The average time from interconnection request to commercial operation has reached approximately five years nationwide, twice as long as a decade earlier.[9] That timeline has not changed meaningfully in recent years, even as data center construction cycles have compressed to 12 to 24 months.

Through most of 2025, the regulatory response beyond queue reform was reactive rather than systematic. FERC's December 2025 co-location order addressed PJM in isolation. The proposed DATA Act would create a federal framework for private power systems but remained unresolved. That changed on June 18, 2026, when FERC issued tailored show-cause orders to all six regional grid operators, directing each to justify or reform the tariff rules governing how large energy users connect to the transmission system within 60 days. The orders target five categories of reform, including co-location arrangements and behind-the-meter generation, cost allocation to prevent shifting upgrade costs to existing ratepayers, and new transmission services for flexible large loads. The action represents a shift from responding to individual proceedings to requiring all six markets to develop large-load integration frameworks simultaneously.[10]

The more consequential sorting is happening at the state level. States competing for data center investment have moved to enable self-generation: Ohio, Utah, West Virginia, Oklahoma, and New Hampshire all enacted enabling legislation in 2025. By contrast, states that are absorbing the grid impact have moved in the opposite direction: New York, California, and New Jersey have proposed or enacted ratepayer protections, cost-sharing requirements, and dedicated rate classes for large loads.[11]

For developers planning a 10- or 15-year BTM commitment, the regulatory environment in the host state will matter as much as the economics of the equipment being deployed. The market that enabled the “sprint” phase was permissive almost everywhere. The market that sustains the “marathon” phase will be selective.

[9] Lawrence Berkeley National Laboratory; LBNL

[10] FERC, Order on Show Cause Proceeding, Docket Nos. EL25-49, AD24-11, EL25-20, December 18, 2025; Decentralized Access to Technology Alternatives Act (DATA Act), proposed January 2026

[11] S&P Global Energy, US Data Center Load Regulations Emerge Amid Efforts to Mitigate Electricity Rate Concerns, December 2025

The most common question about BTM power for data centers is whether it serves as a bridge or a permanent shift. The more useful answer is that it is both, and the market is sorting itself accordingly.

Which power strategies still work if the grid arrives late, arrives partially, or arrives on terms that change the original economics? The answer will not be the same in every market. In some places, engines and aeroderivatives will do their job and step back into backup duty, as intended. In others, hybrid configurations will become the practical middle ground between full grid reliance and full energy independence. In a narrower set of geographies, private power will become embedded infrastructure that persists for decades.

The strongest strategies will be those that preserve uptime without locking capital into a single grid outcome, retain value after interconnection, and avoid matching a 30-year generation asset to a demand profile that may not last as long. The “sprint” showed that data centers could move faster than the grid. The “marathon” will reveal which of those moves became infrastructure, and which were only temporary workarounds.