Data Centers for the Energy Sector: Turning Stranded Power Into Cash Flow

Energy companies sit on two underused assets: power that can't reach demand, and time. Renewable curtailment is measured in tens of terawatt-hours a year. CAISO curtailed 3.4 TWh of wind and solar in 2024, ERCOT over 8 TWh (U.S. EIA, 2025; Modo Energy, 2025), while global gas flaring hit 151 billion cubic meters, the most since 2007 (World Bank, 2025). Meanwhile the grid interconnection queue has grown to roughly 2,600 GW with median waits past five years (Lawrence Berkeley National Lab, 2024). Siting a modular data center behind the meter turns curtailed or flared energy into contracted compute revenue, raises asset utilization, and skips the queue entirely.

This post covers the scale of wasted energy, the proven behind-the-meter compute model, the margin math of selling power versus running inference, why federated inference favors many small sites, and how modular data centers fit an energy site.

How much energy is actually being wasted?

Enough to run a serious amount of compute. The problem was never generation. It was that the electrons couldn't get to demand.

In 2024, CAISO curtailed 3.4 TWh of utility-scale wind and solar, a 29% jump over 2023, with solar making up 93% of it (U.S. EIA, 2025). ERCOT curtailed over 8 TWh; at constrained locations, roughly 22% of all renewable output was thrown away, and the worst-hit generators lost up to 60% of production (Modo Energy, 2025). In Great Britain, more than a tenth of all wind generation was curtailed in 2024 at a cost of about £1.23 billion, including £380 million paid to wind farms to switch off (Utility Week, 2024). Germany couldn't deliver 3.5% of its renewable electricity because of grid bottlenecks (Clean Energy Wire, 2025).

Then there's gas. Global flaring reached 151 bcm in 2024, releasing hundreds of millions of tonnes of CO2 for zero economic value (World Bank, 2025).

Compute is a way to move the demand to the energy instead of the other way around.

Is behind-the-meter compute a proven model or a pitch?

Proven. Across gas, nuclear, and repurposed mining sites, the behind-the-meter playbook is already running at scale.

Crusoe Energy built the template with "Digital Flare Mitigation", capturing gas that would otherwise be flared, generating power on-site, and running modular container data centers at the wellhead. It reports mitigating 2.7 million tonnes of greenhouse gas and keeping about 22 billion cubic feet of gas from being flared (Crusoe / Contrary Research, 2025). On the nuclear side, Talen Energy contracted up to 1,920 MW of behind-the-meter power from its Susquehanna plant to Amazon through 2042, and Constellation is restarting Three Mile Island Unit 1 under a 20-year deal with Microsoft (Utility Dive, 2025).

And the miners pivoted. After the 2024 Bitcoin halving squeezed margins, operators repurposed their power contracts and sites for AI. Core Scientific converted capacity toward roughly 400 MW of HPC infrastructure and signed about $3.5 billion of hosting with CoreWeave over 12-year terms (Core Scientific SEC filings, 2025). The remote-site version of this is its own discipline; we covered the ruggedized end in mining data centers for remote sites.

Why not just sell the spare power to the grid?

Because where power is curtailed, its grid value is near zero, and sometimes negative. British operators are literally paid to switch off (Utility Week, 2024). Redirect that same megawatt to compute and the economics change category.

Data center power demand is the buyer. The IEA projects global data center electricity consumption roughly doubling from about 415 TWh in 2024 to 945 TWh by 2030, growing four times faster than total electricity demand, with AI-focused consumption tripling (IEA, 2025). In the U.S., data centers used about 4.4% of national electricity in 2023 and could reach 6.7–12% by 2028 (Lawrence Berkeley National Lab, 2024).

The margin comparison is the whole argument:

Treat the high end as illustrative, it varies enormously with local power prices and contract terms. But even the conservative floor, a bare-metal lease at a few million per MW per year, is an order of magnitude above the value of curtailed power. Electricity stops being a pass-through cost and becomes a profit center. That's the cash-flow case. The economics of building your own generation rather than waiting for the utility are laid out in on-site power: why data centers are done waiting for the grid, and the renewable-pairing version in the solar-powered data center.

What is federated inference, and why does it favor many small sites?

Federated inference spreads AI serving across many sites, keeping data local and moving compute to where the data and the power already sit. It's the opposite of hauling everything to one hyperscale campus.

This matters for energy siting because inference behaves differently from training. Training clusters can be isolated and centralized. Inference lives in the live user path, so latency is a functional requirement, not an optimization. Distributed sites can hit sub-10ms where a distant central region adds 30–150ms round-trips (Vertical Data, 2025; Red Hat, 2025). That means inference tolerates, even prefers, being spread across many mid-sized sites of half a megawatt to a few megawatts each.

Which is exactly the shape of stranded power. You don't have one 2 GW surplus in one place. You have hundreds of substations, wind farms, and gas plants each with some spare capacity. Federated inference maps onto that geography. The distributed-infrastructure economics are the same ones driving edge colocation.

How does behind-the-meter siting skip the interconnection queue?

The queue is the moat, and behind-the-meter walks around it.

About 2,600 GW of generation and storage was waiting in U.S. interconnection queues at the end of 2023, over 95% of it zero-carbon (Lawrence Berkeley National Lab, 2024). The median time from interconnection request to commercial operation has doubled, from under two years for early-2000s projects to over four years for recent ones, and in PJM, the largest U.S. grid, the timeline now exceeds eight years (LBNL, 2024; Introl, 2025). A grid-connected project you start today may energize near 2033.

Compute co-located behind an existing generator's meter doesn't have to clear that transmission queue. It can energize in months. Regulators are catching up to the model: FERC's December 2025 order directed PJM to reform its behind-the-meter generation rules and create new categories for co-located data-center load (FERC, 2025). For an energy owner with existing generation and grid rights, that's the difference between monetizing now and waiting most of a decade.

Why modular data centers, specifically, for energy sites?

Because energy sites are rarely tidy, rarely urban, and rarely willing to wait 30 months for a building.

Factory-built modular units deploy in 8–16 weeks after delivery versus 18–36 months for traditional builds, arrive pre-integrated and factory-tested, and come as ruggedized containerized increments of roughly 0.5–2 MW (Data Center POST, 2025; AND Cable, 2025). ModulEdge units run from −35°C to +52°C, which matters on an exposed oilfield pad or a substation yard. That lets an operator follow the power, dropping standardized capacity onto a wind farm, a gas plant, or a substation and scaling in steps as availability dictates, rather than betting a fixed shell on a single forecast. The siting logic for putting capacity where the power is, not where the fiber used to be, is the core of our mobile and portable modular siting playbook.

The grid won't get faster. Curtailment won't fix itself. The compute demand isn't slowing. If you own power that can't reach a market, the question isn't whether to monetize it as compute, it's how fast you can get a module on site. The full delivery model is in our definitive guide to modular data centers.

Frequently asked questions

How much renewable energy is wasted through curtailment? In 2024, CAISO curtailed 3.4 TWh of wind and solar (up 29% year-on-year, 93% of it solar), and ERCOT curtailed over 8 TWh (U.S. EIA, 2025; Modo Energy, 2025). In Great Britain, more than a tenth of all wind generation was curtailed at a cost of about £1.23 billion (Utility Week, 2024). That is energy an on-site data center could consume instead of it being switched off.

Is powering compute with flared gas a real, proven model? Yes. Crusoe Energy pioneered generating electricity from otherwise-flared gas to run modular container data centers at the wellhead, reporting 2.7 million tonnes of greenhouse gas mitigated and about 22 billion cubic feet of gas kept from flaring (Crusoe / Contrary Research, 2025). Globally, 151 bcm of gas was flared in 2024, the most since 2007, so the addressable resource is enormous (World Bank, 2025).

Why not just sell spare power to the grid instead of building a data center? Where power is curtailed, its grid value is near zero or negative. British operators are paid to switch off (Utility Week, 2024). Redirected to compute, the same megawatt can support colocation revenue of roughly $4.4 million per year and more for AI hosting, with electricity becoming a margin profit center rather than a pass-through cost. The exact spread for any given site depends on local prices and contract terms.

What is federated inference and why does it favor many small sites? Federated inference spreads AI serving across many interconnected sites, keeping data local and moving compute to where data and power already are (Red Hat, 2025). Because inference sits in the live user path, latency is a functional requirement, distributed sites can hit sub-10ms versus 30–150ms to a distant central cloud (Vertical Data, 2025), so spreading modest modular sites across energy assets often beats one giant campus.

‍How does behind-the-meter siting avoid the interconnection queue? U.S. interconnection queues held about 2,600 GW at the end of 2023, and median wait to commercial operation has risen past five years nationally and over eight years in PJM (LBNL, 2024; Introl, 2025). Compute co-located behind an existing generator's meter doesn't have to clear the transmission queue, so it can energize in months, and FERC's December 2025 order is actively reforming those rules (FERC, 2025).

Why are modular data centers the right fit for energy sites? Factory-built modular units deploy in 8–16 weeks versus 18–36 months for traditional builds, arrive pre-integrated and factory-tested, and come in ruggedized increments of roughly 0.5–2 MW (Data Center POST, 2025). That lets an energy operator follow the power, dropping capacity onto oilfield pads, substations, or wind and solar farms and scaling in standardized steps as availability dictates.