Behind-the-Meter Data Center Siting: An AI Cluster Playbook for Energy Companies

A behind-the-meter data center draws power directly from a generation asset (a gas peaker, wind farm, solar array, biogas digester, or nuclear plant) before that electricity crosses the utility meter and enters the public grid. Doing this lets a compute load skip the interconnection queues that the Lawrence Berkeley National Lab measured at roughly 2,600 GW of pending capacity and a five-year median wait as of 2024. For a first deployment, most energy companies don’t need a gigawatt campus. They need a 10-20 MW module: big enough to generate real revenue, small enough to fit behind switchgear a site already has.

This post picks up where the stranded-power business case ends and walks through how to screen a candidate site, size the first module, get through interconnection and permitting, pick a contract structure, and match infrastructure to a site that was never meant to host a data center.

What Is a Behind-the-Meter AI Cluster, and Why Does It Start at 10-20 MW?

Most energy executives hear “AI data center” and picture Stargate or Meta’s Hyperion campus in Louisiana, which is scaling toward 5 GW through 2030 on a reported $50 billion budget. That reference point kills good projects before they start. A behind-the-meter cluster at a wind farm, biogas site, or gas peaker solves a much narrower problem than those campuses ever will: what to do with power that’s already stranded, curtailed, or underused at a site that will never see gigawatt-scale transmission investment.

Behind-the-meter (BTM) means the compute load sits on the generation side of the meter, drawing electricity that hasn’t yet been sold onto the grid. Cleanview’s 2025 tracking identified 46 U.S. data center projects planning roughly 56 GW of behind-the-meter capacity, with about 90% of that pipeline (close to 50 GW) announced in 2025 alone. Most of it is gas. Texas leads, on the strength of Permian Basin supply, pipeline density, and a permitting environment built for speed.

That 56 GW figure spans everything from peaker-scale projects to campuses, but the mechanics don’t change with size. You need a generator with spare or curtailed capacity and a compute load parked next to it, connected by a wire short enough to avoid the queue. That’s the whole model.

The workload matters as much as the site. This playbook is built for inference-class AI clusters, not training superclusters, which remain a hyperscale-campus problem with different cooling, networking, and latency demands. Inference racks running at 40 kW per rack and above are where factory-built modular infrastructure does its best work. Run the arithmetic on a single NVIDIA GB200 NVL72 rack, which draws 120-140 kW, and a 10 MW data center module houses roughly 70-80 of them before overhead; a 20 MW module, 140-160. That’s a serious cluster. It’s also a size an energy company can host without becoming a power-plant developer first.

What Makes a Generation Site a Good Candidate for Compute?

Not every substation or well pad qualifies. Some sites have curtailment on paper but no realistic way to get a compute load connected without a multi-year transmission project, which defeats the entire point of going behind the meter.

The strongest candidates have real surplus or curtailed capacity and a short physical run to an interconnection point. Spare land helps. So does switchgear that already handles the voltage class a 10-20 MW load needs. Miss two of these and the project usually isn’t worth pursuing at this size.

We’ve made the economic argument for monetizing stranded power elsewhere — the curtailment losses, the flared-gas volumes, the margin gap between selling power at grid rates and hosting compute (see Turning Stranded Power Into Cash Flow). This post assumes you’re already convinced and goes straight to the site walk. The general fundamentals of evaluating any data center site, power and land and connectivity, still apply here, just weighted differently for a generation-adjacent site than for a traditional colocation build.

Generation firmness deserves a second look. Gas peakers and biogas digesters produce something close to dispatchable, predictable power, which is easy for a compute tenant to underwrite. Wind and solar sites can absolutely work. Texas regulators approved a 260 MW AI data center co-located with a similarly sized wind farm in July 2026, but the compute load has to tolerate variability, or the site needs storage to smooth it.

Why Is 10-20 MW the Right First Module, Not a Gigawatt Campus?

Gigawatt campuses are impressive, and mostly irrelevant to a mid-size utility or an independent power producer sitting on one gas peaker. Meta’s Hyperion is scaling to 5 GW. OpenAI’s Stargate venture with Oracle and SoftBank has roughly 7 GW planned. Neither is a utility-scale AI data center in the sense this playbook cares about: both require dedicated transmission-class substations at 138-230 kV, multi-year permitting cycles, and capital measured in tens of billions.

None of that is available, or necessary, at a 40 MW gas peaker or a 120 MW wind farm carrying 20 MW of chronic curtailment.

A 10-20 MW module fits behind the switchgear most mid-size generation sites already own; many can interconnect at an existing 34.5 kV feeder without a new substation. It’s small enough that a single generation asset, even a modest one, can supply it without threatening the site’s primary job, whether that’s selling power under an existing PPA or running an oilfield. And at this size an operator can prove the model (construction time, interconnection process, contract terms, uptime) before staging up.

Staging matters because most sites don’t have 100 MW of stranded capacity sitting idle on day one. They have 10 MW, with another 10-15 MW likely once curtailment patterns firm up or a second turbine comes online.

Past Phase 3, the project starts to resemble the gigawatt campuses this playbook is explicitly not written for. Different capital stack. Different timeline. Different buyer.

How Does Behind-the-Meter Interconnection and Permitting Actually Work?

The entire appeal of BTM siting is skipping the queue. That’s not automatic, and the rules are moving fast enough that “how it worked last year” isn’t a safe assumption to build a deal on.

At the federal level, FERC directed PJM, the largest U.S. grid operator, in a December 2025 order to create new rules specifically for co-located data center load, formalizing a category that had previously been argued case by case. Regulators, in other words, are catching up to a model that’s already being built on the ground. Nobody should read that as permission to skip review.

State rules vary more, and Texas shows where this is heading. Senate Bill 6, signed in June 2025, gave ERCOT authority to disconnect co-located large loads during grid emergencies and required Public Utility Commission of Texas review for any new load of 75 MW or more seeking to co-locate with an existing generator. Loads above that threshold must be able to curtail their full draw within 30 minutes during an emergency. A 10-20 MW first module sits comfortably under that line in ERCOT. That’s one more reason starting small is often the lighter regulatory lift.

Below the review thresholds, expect a standard permitting checklist regardless of jurisdiction: air permits if pairing with new gas generation, water discharge permits if cooling requires anything beyond a closed loop or air, local zoning or conditional-use approval for the structure itself, and utility notification even when no new grid interconnection is being requested. None of it is unique to compute — it’s the same environmental and safety review any new load addition at a generation site would trigger.

The ground keeps shifting under bigger deals, which is instructive even at 10-20 MW scale. Talen Energy’s original arrangement to sell Amazon up to 1,920 MW behind the meter from its Susquehanna nuclear plant was restructured in June 2025 into a grid-connected, front-of-the-meter retail deal, after a FERC ruling created uncertainty around the original co-location structure. Build the first module to survive a reclassification. Don’t assume today’s BTM rules are permanent.

What Contract and Ownership Structures Can Energy Companies Offer?

Three structures cover most real deals, and the right one depends on how much operational risk the energy company wants to carry.

Colocation lease. The energy company acts as landlord: it leases site, power, and interconnection capacity to a compute operator who owns and runs its own IT equipment. Revenue is a base rent for capacity plus a metered power pass-through, structurally close to a real estate lease with a utility rider attached. The cost components most closely track what we break down in modular data center cost. Cleanview’s 2025 tracking put achievable annual compute revenue at $10-12 million per megawatt for BTM gas-fired sites, an order of magnitude above merchant power sales at a congested node. Of the three structures, this one is the simplest to originate and asks the least new expertise of the energy company.

Tolling agreement. Common in battery storage, and increasingly relevant to gas-paired compute: the compute operator, the “toller,” pays a fee for the right to run its own dispatch decisions through a plant the energy company still owns and physically operates. The energy company keeps O&M control and the underlying asset. The toller gets flexibility over when and how much power it draws. This hands over less operational control than a lease, which matters if the generation asset also serves other customers or existing grid obligations.

Joint venture. The energy company contributes site, power rights, and interconnection standing; a compute operator or infrastructure investor contributes capital and demand. Google’s December 2024 agreement with Intersect Power is worth studying here: co-locating data centers inside roughly $20 billion of renewable energy parks, with the first phase targeted for 2026. JVs carry the most complexity and the most upside. They’re also the structure most exposed to the kind of regulatory reclassification that hit Talen’s original Amazon arrangement.

Whichever structure you choose, treat contract term and regulatory assumption as separate risks. A 15-20 year colocation lease signed under today’s BTM rules needs explicit language for what happens if a state or ISO reclassifies co-located load mid-contract. Texas, PJM, and FERC all changed their rules in 2025 alone.

Why Does Factory-Built Modular Fit Distributed Generation Sites Better Than Custom Construction?

A custom-built data center wants a flat site with permanent access roads, plus a demand forecast that holds for the 18-36 months a traditional build typically takes. A wind farm’s substation yard, a biogas digester’s fence line, or an oilfield pad rarely offers any of that. The site was chosen for wind speed or gas geology, not construction logistics. It usually sits well outside anywhere a general contractor wants to mobilize a crew for three years.

Factory-built modular units are designed for exactly this mismatch. They arrive pre-integrated and tested, deploy in roughly 8-16 weeks after delivery, and land in standardized increments (commonly 0.5-2 MW per unit) that stack into a 10-20 MW cluster without requiring a fixed campus master plan. ModulEdge’s own units are rated from -35°C to +52°C, which on an exposed substation yard in North Dakota or a wellhead pad in West Texas is the difference between a workable deployment and a stalled one. The same logic holds for any distributed generation site where a permanent building was never on the table. See our mobile and portable modular siting playbook for the ruggedized end of this.

Staged, modular capacity matches how stranded power actually shows up in the field. MARA doubled its flared-gas hosting capacity to 50 MW across Texas and North Dakota sites in partnership with NGON in late 2025, and it did so by adding standardized modular capacity at multiple wellhead locations as gas volumes justified it rather than building one large facility. Vespene Energy took the same approach to landfill gas, pairing on-site generation equipment with data processing load at a municipal landfill in Marathon County, Wisconsin. Neither looks anything like a hyperscale campus, and both look exactly like the kind of site this playbook is written for. Renewable-paired sites raise a related set of questions we cover separately in the solar-powered data center.

A fixed building bets the entire capex on one forecast. Modular capacity lets an energy company add the next 10 MW when the second turbine or the next well comes online. It lets them stop cleanly if it doesn’t.

How Should an Energy Company Decide Where to Start?

Screen for a site with real surplus, a short wire to existing switchgear, and a generation type your compute partner can live with. Firm gas and nuclear are the easiest sell. Wind and solar take more structuring, but they work.

Size the first module at 10-20 MW, not more. Pick a contract structure that matches how much operational control you’re willing to give up: lease for simplicity, tolling for shared risk, joint venture for the sites worth building a real partnership around. Then build in modular capacity rather than a permanent structure, because the next phase, and the regulatory rules governing it, may not look like this one. The full delivery model, from spec to commissioning, is in our definitive guide to modular data centers.

The energy companies moving now are trading a stranded asset for a contracted one. The ones still waiting are still deciding whether their surplus power is a liability at all.

Frequently Asked Questions

What is a behind-the-meter data center? A behind-the-meter data center draws power directly from a generation asset (a gas plant, wind farm, solar array, biogas site, or nuclear facility) before that electricity crosses the utility meter and enters the public grid. Because the load never touches the transmission network in the same way a new grid-connected project does, it can often avoid the multi-year interconnection queue, which the Lawrence Berkeley National Lab measured at roughly 2,600 GW of pending capacity and a five-year median wait as of 2024.

Why is 10-20 MW considered the right first size for an energy company’s AI cluster? A 10-20 MW module typically fits behind switchgear a generation site already has, avoiding the dedicated transmission-class substation that gigawatt campuses require. It houses roughly 70-160 GB200-class AI server racks depending on density, enough to generate meaningful colocation or hosting revenue, while small enough to prove out construction, interconnection, and contract terms before committing to a larger buildout.

How much surplus or curtailed power does a site need to host a 10 MW compute cluster? Most operators target 15-25 MW of consistently available surplus or curtailed capacity to host a 10 MW draw, leaving margin for ramp-up, redundancy, and a future expansion to 20 MW. The exact number depends on how firm the generation source is: gas and nuclear sites typically need less margin than wind or solar sites with variable output.

Do I need FERC or state utility approval to co-locate compute behind my meter? It depends on load size and jurisdiction. FERC’s December 2025 order directs PJM to create new rules specifically for co-located data center load, while Texas already has a statute: Senate Bill 6 (June 2025) requires Public Utility Commission of Texas review for any new load of 75 MW or more co-locating with an existing generator, plus a 30-minute emergency curtailment capability. A 10-20 MW first module generally sits below that review threshold.

What’s the difference between a colocation lease and a tolling agreement for energy-to-compute deals? A colocation lease makes the energy company a landlord: it charges a compute operator base rent for capacity plus a metered power pass-through, and the operator owns and runs its own IT equipment. A tolling agreement keeps the power plant under the energy company’s ownership and operation, with the compute operator paying a fee for the right to draw power on its own schedule, which hands over less operational control than a lease.

Can curtailed solar or wind power a behind-the-meter AI cluster, or does it need to be gas or nuclear? Renewables can work, but variability changes the deal terms. Texas regulators approved a 260 MW AI data center co-located with a similarly sized wind farm in July 2026 under a net metering arrangement, though the load must be able to curtail fully within 30 minutes during grid emergencies. Firm generation types (gas peakers, biogas, nuclear) require less compute-side flexibility and are generally the simpler first pairing.

Why not skip straight to gigawatt scale if hyperscalers are already building at that size? Gigawatt campuses like Meta’s Hyperion, scaling toward 5 GW through 2030, or OpenAI’s Stargate, with roughly 7 GW planned, require dedicated transmission-class substations, multi-year permitting, and tens of billions in capital that most utilities, IPPs, and independent generation owners neither have nor need for a first deployment. A 10-20 MW module proves the operational model on a single generation asset before any larger commitment is considered.

How fast can a 10-20 MW modular data center actually be deployed at a generation site? Roughly 8-16 weeks after delivery for factory-built modular units, compared with 18-36 months for a traditional custom-built facility. The speed comes from arriving pre-integrated and factory-tested, in standardized increments that don’t require a fixed campus master plan.