After a year of concerted hand-wringing about the growing energy needs of data centers, a report that dropped just before the holidays proposed a solution that had been strangely absent from the discussion.
AI companies have seemingly grasped for every imaginable source of clean energy to quench their thirst for power, including pricey, left-field ideas like restarting shuttered nuclear plants. Some are foregoing climate concerns altogether and ordering up off-grid natural gas turbines. In a pithily named new analysis — “Fast, scalable, clean, and cheap enough” — the report’s authors make a compelling case for an alternative: off-grid solar microgrids.
An off-grid solar microgrid is a system with solar panels, batteries, and small gas generators that can work together to power a data center directly without connecting to the wider electricity system. It can have infinite possible configurations, such as greater or smaller numbers of solar panels, and more or less gas-generated capacity. The report models the full range of possibilities to illustrate the trade-offs in terms of emission reductions and cost.
An eclectic group of experts got together to do the research, including staffers from the payment company Stripe, a developer called Scale Microgrids, and Paces, which builds software to help renewable energy developers identify viable sites for projects. They found that an off-grid microgrid that supplied 44% of a data center’s demand from solar panels and used a natural gas generator the rest of the time would cost roughly $93 per megawatt-hour compared to about $86 for large, off-grid natural gas turbines — and it would emit nearly one million tons of CO2 less than the gas turbines. A cleaner system that produced 90% of its power from solar and batteries would cost closer to $109 per megawatt-hour, the authors found. While that’s more expensive than gas turbines, it’s significantly cheaper than repowering Three Mile Island, the fabled nuclear plant that Microsoft is bringing back online for an estimated $130 per megawatt-hour.
One challenge with solar microgrids is that they require a lot of land for solar panels. But a geospatial analysis showed that there’s more than enough available land in the U.S. southwest — primarily in West Texas — to cover estimated energy demand growth from data centers through 2030. This shouldn’t be taken as a recommendation, per se. The paper doesn’t interrogate the need for data centers or the trade-offs of building renewable power for AI training facilities versus to serve manufacturing or households. The report is just an exercise in asking whether, if these data centers are going to be developed, could they at least add as few emissions as possible? Not all hyperscalers care about climate, and those that do might still prioritize speed and scale over their net-zero commitments. But the authors argue that it’s possible to build these systems more quickly than it would be to install big gas turbines, which currently have at least three-year lead times to procure and fall under more complicated permitting regimes.
Before the New Year, I spoke with two of the authors — Zeke Hausfather from Stripe and Duncan Campbell from Scale Microgrids — about the report. Stripe doesn’t build data centers and has no plans to, but Hausfather works for a unit within the company called Stripe Climate, which has a “remit to work on impactful things,” he told me. He and his colleagues got interested in the climate dilemma of data centers, and enlisted Scale Microgrids and Paces to help investigate. Our conversation has been lightly edited for clarity.
Why weren’t off-grid solar microgrids really being considered before?
Zeke Hausfather: As AI has grown dramatically, there’s been much more demand for data centers specifically focused on training. Those data centers have a lot more relaxed requirements. Instead of serving millions of customer requests in real time, they’re running these incredibly energy intensive training models. Those don’t need to necessarily be located near where people live, and that unlocks a lot more potential for solar, because you need about 50 times more land to build a data center with off-grid solar and storage than you would to build a data center that had a grid connection.
The other change is that we’re simply running out of good grid connections. And so a lot of the conversation among data center developers has been focused on, is there a way to do this with off-grid natural gas? We think that it makes a lot more sense, particularly given the relaxed constraints of where you can build these, to go with solar and storage, gas back-up, and substantially reduce the emissions impact.
Duncan Campbell: It was funny, when Nan [Ransohoff, head of climate at Stripe] and Zeke first reached out to me, I feel like they convinced me that microgrids were a good idea, which was the first time this ever happened in my life. They were like, what do you think about off-grid solar and storage? Oh, the energy density is way off, you need a ton of land. They’re like, yeah, but you know, for training, you could put it out in the desert, it’s fine, and hyperscalers are doing crazy things right now to access this power. We just went through all these things, and by the end of the call, I was like, yeah, we should do this study. I wasn’t thinking about it this way until me, the microgrids guy, spoke to the payments company.
So it’s just kind of against conventional logic?
Campbell: Going off-grid at all is wild for a data center operator to consider, given the historical impulse was, let’s have 3x more backup generators than we need. Even the off-grid gas turbine proposals out there feel a little nuts. Then, to say solar, 1,000 acres of land, a million batteries — it’s just so unconventional, it’s almost heretical. But when you soberly assess the performance criteria and how the landscape has shifted, particularly access to the grid being problematic right now, but also different requirements for AI training and a very high willingness to pay — as we demonstrate in our reference case with the Three Mile Island restart — it makes sense.
Hausfather: We should be clear, when we talk about reliability, a data center with what we model, which is solar, batteries, and 125% capacity backup gas generators, is still probably going to achieve upwards of 99% reliability. It’s just not gonna be the 99.999% that’s traditionally been needed for serving customers with data centers. You can relax some of the requirements around that.
Can you explain how you went about investigating what it would mean for data centers to use off-grid solar microgrids?
Campbell: First we just built a pretty simple power flow model that says, if you’re in a given location, the solar panel is going to make this much power every hour of the year. And if you have a certain amount of demand and a certain amount of battery, the battery is going to charge and discharge these times to make the demand and supply match. And then when it can’t, your generators will kick on. So that model is just for a given solar-battery-generator combo in a given location. Then what we did is made a huge scenario suite in 50-megawatt increments. Now you can see, for any level of renewable-ness you want, here’s what the [levelized cost of energy] is.
Hausfather: As you approach 100%, the costs start increasing exponentially, which isn’t a new finding, but you’re essentially having to overbuild more and more solar and batteries in order to deal with those few hours of the year where you have extended periods of cloudiness. Which is why it makes a lot more sense, financially, to have a system with some gas generator use — unless you happen to be in a situation where you can actually only run your data center 90% of the time. I think that’s probably a little too heretical for anyone today, but we did include that as one of the cases.
Did you consider water use? Because when you zoom in on the Southwest, that seems like it could be a constraint.
Hausfather: We talked about water use a little bit, but it wasn’t a primary consideration. One of the reasons is that how data centers are designed has a big effect on net water use. There are a lot of designs now that are pretty low — close to zero — water use, because you’re cycling water through the system rather than using evaporative cooling as the primary approach.
What do you want the takeaway from this report to be? Should all data centers be doing this? To what extent do you think this can replace other options out there?
Hausfather: There is a land rush right now for building data centers quickly. While there’s a lot of exciting investment happening in clean, firm generation like the enhanced geothermal that Fervo is doing, none of those are going to be available at very large scales until after 2030. So if you’re building data centers right now and you don’t want to cause a ton of emissions and threaten your company’s net-zero targets or the social license for AI more broadly, this makes a lot of sense as an option. The cost premium above building a gas system is not that big.
Campbell: For me, it’s two things. I see one purpose of this white paper being to reset rules of thumb. There’s this vestigial knowledge we have that this is impossible, and no, this is totally possible. And it seems actually pretty reasonable.
The second part that I think is really radical is the gigantic scale implied by this solution. Every other solution being proposed is kind of like finding a needle in a haystack — if we find this old steel mill, we could use that interconnection to build a data center, or, you know, maybe we can get Exxon to make carbon capture work finally. If a hyperscaler just wanted to build 10 gigawatts of data centers, and wanted one plan to do it, I think this is the most compelling option. The scalability implied by this solution is a huge factor that should be considered.
Courtesy of U.S. Forest Service