Behind-the-meter or colocated gas-powered electricity promises to deliver firm power at $44 per MWh, a really good deal…under the right conditions. Download the model and tinker with those conditions yourself.
[This post is still in draft form.]
The energy needs of AI datacenters have turned the U.S. energy sector upside down in the last year or two. Companies at the center of the datacenter construction boom – many of which had, and officially still have, some of the most ambitious corporate climate goals in the world – have turned everywhere for new power, and that has included gas. Specifically, giant co-located and behind-the-meter deals. For context, let’s answer explain who is making these moves. If you’re curious about why they are making these moves, spoiler alert: the promise of electricity at $44/MWh. But we’ll come to that.
If you’re just here for the model, scroll down for the links.
Who’s doing this? Basically, all of the hyperscalers.
| Lead company (and major partner, where known) | Deal scale and site |
| Meta and Entergy Louisiana | 5 GW, northeast Louisiana |
| Microsoft and Chevron | 2.67 GW, west Texas |
| Microsoft | 1.35 GW, West Virginia |
| Amazon | 7.65 GW, west Texas |
| Alphabet (Google) and Black Hills Energy | 564 MW of co-located gas, additional 2.1 GW of various resources, in and near Wyoming |
I’ve no doubt left a few off the list, but don’t fixate on the details here: This list is incomplete and contains oversimplifications; for example, several Google projects and possibly others will include solar resources and grid connections for additional capacity. The point of the list is the massive scale of this ‘speed-to-power’ moment.
Behind-the-meter gas: a (very) simple model of LCOE
As part of my clean energy teaching, I teach students how to calculate Levelized Cost of Energy (LCOE), as well as how to think about the strengths and limitations of the metric. Acknowledging that it isn’t perfect, I nonetheless believe it is useful. I’ve defended LCOE’s relevance elsewhere so the focus here is on the model to generate that metric.
The model included here does not examine any one of these sites. Rather, it is a generic model that could be for any site.
The technical and financial parameters of the model:
- capacity factor (%)
- heat rate (MMBTU/MWh)
- capital expenditure ($/kW)
- operational expenditure ($/kW per year)
- cost of natural gas ($/MMBTU)
- equity rate (expected rate of return for equity investors) and debt rate (financing rate)
I have largely relied on Lazard’s LCOE+ (19th Edition, 2026) for these values; you’ll find the midpoint in the spreadsheet as a starting point, though I’ve made a few changes as a starting point. I’ll describe those assumptions and then the results.
Notes on the default assumptions
The shape of this model is for another purpose: understanding utility-scale generation that goes on the grid and competes with other resources there. The behind-the-meter or colocated (and de facto behind-the-meter) scenario by hyperscalers warrants a few changes – and these will help you understand why my results diverge from Lazard’s summary stats for grid-connected resources:
- Capacity factor: I’ve used 95% as a default. This above the top of Lazard’s massive range (30-90%), but that range is oriented toward what’s on the grid, and even the newest combined-cycle gas in the U.S. average about 65% because it’s a mix of baseload and dispatchable. I’ve chosen this high number to take the hyperscalers at their word (they’re prioritizing ‘firm power’) and to ‘steel-man’ the gas case.
- Price of natural gas: I’ve used $3.50 as the default. I really encourage tinkering with the model here since only an ahistorical fool would assume a low and stable price of natural gas for the 10- and 20-year terms that most companies have signed. (I’m assuming the price of gas will be passed through to the hyperscalers, though such deal details have not yet slipped into public view.) If you would like to turn this dial, use EIA’s natural gas historical data (data here and 25-year price snapshot here).
- Debt rate (i.e., cost of bank finance): I’ve used 7% a default. The hyperscalers, as the most profitable companies on the planet, enjoy access to cheap debt capital.
- Debt share: I’ve used 100%, to keep things easy. This parameter has less meaning here because of the scenario, but using 100% makes the extreme case in favor of gas.
- Equity rate of return: I’ve left in 12% (a default from Lazard), but when debt share is 100%, equity funding drops out of the calculation by definition.
- Depreciation: The model uses the Modified Accelerated Cost Recovery System, the fastest allowable depreciation schedule for this asset class (IRS Publication 946).
Again, if you disagree with these starting points, then change them! Scroll down for the model.
Results: in the best of circumstances, gas provides low-cost firm power
With the default values, my model generates a LCOE of about $44 per MWh.
Wait, what? you might say. That seems like a good deal. Indeed it is. That’s with expensive turbines and reasonably priced gas. The viability is surely
You can turn the dials yourself, but just to help you get your bearings, consider the following scenarios. In each bullet point, the parameter value change is the only change to the model, i.e., these are individual changes, rather than several at once (though you can try those yourself):
- Assume $6 gas (instead of $3.46) throughout the project life –> $60/MWh.
- Assume a capacity factor of just 80% (instead of 95%) throughout the project life –> $47.50/MWh.
- Turbine prices spike to $3,000/kW (instead of $2,100) –> $51/MWh
- Project life is 15 years (instead of 30) –> $51/MWh
- More typical financing, including debt rate of 8% (vs. 6%) and 30% equity financing (vs. zero) –> $52/MWh
Keep in mind that these values are an oversimplification. LCOE describes what every MWh would have to earn for the project to make investors whole. Projects are messier than that in real life, with MWh in good years earning a lot and vice versa. Still, these calculations and sensitivities give a sense of the investments that hyperscalers are signing up for. As they say: all models are wrong, but some are useful.
A few shortcomings and omissions worth mentioning:
- For ease of use, I’ve made the model rigid: you can enter a different assumption, but then the assumption holds for the entire life of the project. A richer version of the model would allow for certain assumptions to shift gradually. Two candidates:
- Capacity factor: Why not have it fade gradually over time? It seems likely that the projects will gradually lose out to renewables and batteries, so that 95% might slowly turn into 65% over the course 10-15 years.
- Price of gas: One might imagine modeling gas in different ways, perhaps with upward or downward trends, or with a series of spikes – after all, we’ve had two energy price shocks in less than five years.
- Think of this calculation as capturing energy, but not power quality. Datacenters have infamously challenging power quality needs – not just a lot of power and energy, but huge load oscillations that cause big problems when they’re connected to the grid (e.g., Shoukse et al (2025), Power Stabilization for AI Training Datacenters). Despite the well-known ramp speed of natural gas, it isn’t enough for AI loads; a full model would include battery storage and a whole lot of electronics I’ve left out.
- Relatedly, there’s only gas generation here. What about fuel cells or onsite solar or onsite wind or oversized storage for real flexibility, in addition to gas? That’s all beyond this exercise, but smart people have thought about this (e.g., Haggi and Morankar (2026), Resilient Design and Optimal Operation of Battery Energy Storage Systems for Behind-the-Meter Data Center Microgrids).
- The price of natural gas is not stable, and if everyone moves toward gas-powered generation, we could end up in a supply-constrained world, or even just a few periods of tight supply and high prices. To predict or assume $3.46 over the long haul is optimistic.
- There’s no price on carbon. That’s right, no assumed cost of the greenhouse gas emissions that all of the companies in question have pledged to eliminate and have mostly already done so from their electricity portfolios. Enjoy tinkering with that cell.
Downloads
You’ll find the read-only Google Sheet here: BTM gas: a simple model (https://docs.google.com/spreadsheets/d/1Jbo0I5zbQXcqn21li-viQpubmyMNCguSgY3NqUfuROA/edit?usp=sharing)
Under the File option in the menu bar, you can (a) select Make a copy to import a version to your Google Drive or (b) select Download to export the document in a variety of different file formats.
