Solar Microgrids – or “Hypergrids” – Are America’s Best Bet to Win the AI Race

Skyrocketing energy demand for AI data centers is an inescapable topic. Continued leadership in AI is critical to America’s economic growth and its national security, but its voracious appetite for energy is straining the capacity of our aging electricity grids – and potentially increasing electricity costs and emissions in the process, especially if this new demand is met primarily by reactivating nuclear plants or going all-in on natural gas.

Scale’s Duncan Campbell recently worked with collaborators from Stripe and Paces to investigate the potential of off-grid microgrids powered by solar, battery storage, and natural gas for backup to meet AI’s energy needs, and what they found has been making headlines: 90% solar-powered microgrids in the U.S. southwest offer a faster and more scalable solution than existing nuclear or natural gas-only options, with comparable costs and lower risk.

In short, the surest path for America to win the AI race will harness our world-class solar resources and technology leadership. 

Importantly, thanks to the combination of Scale’s modeling expertise and Paces’ siting expertise, we know that this opportunity isn’t just hypothetical – an incredible 1,200 gigawatts of these primarily solar-powered AI datacenters could be built in the U.S. southwest, enough to meet all projected US datacenter growth through 2030 several times over.

Read on for a summary of their key findings, dig into the team’s white paper for an in-depth look into this important research, and reach out to us if you’re interested in discussing how to make off-grid solar microgrids for AI data centers a reality.

Why Speed, Scale, and Certainty are Key – And Why Solar Wins

For individual companies and entire countries, the AI race has existential stakes: those who first develop the best models, powered by the biggest “hyperscale” data centers, could lock in long-term dominance over one of the most strategically important and profitable industries in history. The urgency of securing this first-mover advantage has led companies to leave aside their usual concerns about power costs or sustainability and instead prioritize the speed of accessing data center power supplies, in tandem with the scale of power that can be accessed – and, relatedly, the certainty that the power will be built at the expected speed and scale. 

After 20 years of essentially flat demand, the U.S. utility grid isn’t ready to accommodate the speed or scale of this load growth in the timeframe required, leading hyperscalers to either adopt desperate-sounding gambits like Microsoft restarting Three Mile Island (more on that below) or give up on utility power entirely and take matters into their own hands with off-grid solutions, most often with natural gas turbines. But, while gas turbines offer a familiar, tired-and-true approach, it’s not the fastest or the most scalable approach in 2025, and it faces supply chain, cost, and permitting risks that create significant uncertainty over its viability. 

Here’s what we found when we modeled different pathways for the development of a 500 MW data center:

• Solar+Storage is Faster to Build: Due primarily to faster procurement times, off-grid solar-plus-storage microgrids (with gas engines for backup) can be built sooner than an equivalent amount of off-grid gas turbine capacity. Typical deployment times are 2-4 years for solar microgrids, compared to 3-5 years for gas turbines – and innovative construction practices and design choices could accelerate solar’s timeline further. For example, AI training data centers with lower uptime requirements could begin operations on solar and storage alone, ahead of the installation of backup gas engines. 

• Huge Solar Resource and Ample Supply Chain: Paces found vast areas of the U.S. southwest that would be feasible sites for a 500 MW data center cluster, with criteria including proximity to airports and highways to facilitate construction and access to gas pipelines for backup power generators. With over 1,200 GW of suitable sites, this region could easily host enough solar-powered data centers to meet projected U.S. AI demand of 30-300 GW by 2030. Solar and storage supply chains are also already scaling up rapidly, and both have spare capacity well in excess of current demand.

• Turbines Face Supply Chain and Permitting Risks: Only a handful of companies around the world (GE, Siemens, Mitsubishi) manufacture utility-scale gas turbines, and booming demand is already beginning to strain their limited manufacturing capacity. Turbine lead times of 3+ years today are thus likely to increase. Moreover, gas-only projects may face air quality permitting issues in some areas, potentially further slowing down timelines and/or limiting their scale of deployment. 

• Limited Opportunities for Nuclear Restarts: As noted above, Microsoft made headlines with their plans to restart the Three Mile Island nuclear facility, closed in 2019, to power data center operations. This represents a much faster option than the decade-plus timeline for developing new nuclear power plants in the U.S., but it will still take four years (until 2028) to begin operations, and there are relatively few similar opportunities for recently-decommissioned plants that can be quickly and cost-effectively restarted.  

We expect that the massive demand pull of AI data centers will help drive the commercialization of emerging energy technologies such as enhanced geothermal, small modular nuclear reactors, and even nuclear fusion, and recent news announcements underscore this potential. However, timelines for deploying these pre-commercial technologies tend to start at 2028 at the most optimistic, and delays for first-of-a-kind (FOAK) facilities are highly likely, making them too speculative as a first choice for AI companies seeking to secure competitive advantages today. 

Comparing Cost (and Sustainability) of Different Options

Given the absolute importance of securing hyperscale energy supplies ASAP, the cost of power is relatively much less important for AI data centers than for traditional commercial and industrial customers. That said, costs are never irrelevant, and most solutions being given serious consideration have been in the neighborhood of $100 per MWh. For reference, the average price for grid-connected industrial users in the U.S. is a bit over $80 per MWh, and rates are going up in most places due to the growing costs of maintaining our aging utility grids. 

To estimate costs, the Scale team worked with viable sites identified by Paces and ran 20-year powerflow models for thousands of microgrid configurations supporting data center load 24/7, and ran these through Lazard’s industry-standard Levelized Cost of Energy (LCOE) model. 

What we found was surprising, even to us – even including the cost of full backup, solar is quite competitive with natural gas-only solutions at about 50% solar, and remains within an acceptable range (and significantly lower than Three Mile Island-type restarts) at 90% solar. 

• Natural gas is still cheapest, but with downside risk: We estimate a cost for off-grid natural gas turbine-based solutions as $86 per MWh, which is higher than average industrial rates but is likely the cheapest off-grid option today. However, it’s important to note that natural gas generation prices are largely determined by fuel costs, which can go through periods of volatility – most recently following Russia’s invasion of Ukraine in 2022. As the U.S. ramps up exports of LNG and becomes increasingly integrated in the higher-priced international market for gas, it is projected to lead to further upward pressure on domestic prices as well. 

• Solar is cost-competitive, with room to improve: Our model found that microgrids integrating solar, battery storage, and natural gas backup have very competitive costs with significant upside. For standard designs, this ranges from $93 per MWh for a 44% solar configuration to $109 per MWh to meet 90% of energy needs with solar. These already-competitive costs can be improved further by adopting cost optimizations such as fixed-tilt systems and DC-coupled storage, which could drop costs to $87 per MWh for 44% solar – essentially the same price as gas-only – to $97 for 90% solar. 

• Nuclear remains expensive, even in a restart scenario: Nuclear power has seen its share of the electricity mix fall over the past decade due to its inability to compete with gas and renewables on cost, and the onrush of AI energy demand doesn’t change that fact. Even in a case like Three Mile Island with a recently-decommissioned plant, the costs of restarting the plant plus delivering power over the utility grid results in an estimated LCOE of $130 per MWh. And, according to Lazard, the cost of building a new nuclear plant in the U.S. has an estimated LCOE of $190 per MWh.

Finally, while these estimates don’t include any adders for the cost of carbon or other environmental costs, the fact is that essentially all of the major data center operators, including Google, Microsoft, Amazon, and Meta, have set decarbonization goals (mostly by 2030) and in some cases use internal prices on carbon of up to $100 per ton to align their decisionmaking – or even contract for carbon removals at $100 per ton. Even if those goals have a lower profile today amidst the urgency of the AI race, and perhaps in response to shifting political winds in Washington, the dramatically lower carbon profile of solar microgrids has real implications: 

• Enormous Value for Sustainability Goals: Using off-grid solar microgrids to meet U.S. data center demand instead of 100% gas solutions could reduce carbon emissions by between 0.4 billion and 4.1 billion gigatons of emission from 2026 to 2030. Even at just $50 per ton, that’s up to $200 billion in value – enough to more than offset the marginal difference in power costs while also dramatically reducing regulatory risk.

A New Model of Energy to Power the New Tech Economy

The idea of off-grid, solar-powered superclusters of AI datacenters in the U.S. southwest may sound radical – but given AI’s unprecedented transformative potential, its incalculable economic and national security implications, and its vast energy needs, it’s not surprising that a radical-sounding solution could also be the surest path forward. And in fact, solar microgrids are hardly radical: as Scale can testify, it’s a proven solution for a growing range of commercial and industrial companies, and it leverages the two fastest-growing energy technologies in the world in solar and battery storage. 

Thus, in the words of hockey legend Wayne Gretzsky, powering AI with solar microgrids – maybe more aptly termed megagrids, or hypergrids – located in an area with some of the very best solar resources in the world is simply “skating to where the puck is going.” Since publishing this white paper, we’ve received lots of inbound responses from stakeholders across both the tech and energy industry affirming that we aren’t the only ones considering this solar-forward approach, and its compelling combination of speed, scale, certainty, and cost-competitiveness means that it’s only a matter of when – not if – we see it deployed in the real world. 

If your company is interested in discussing this opportunity further, please reach out to us at info@scalemicrogrids.com.