Off-grid solar-plus-storage could power data centers used for training AI models

Firms building datacenters to train artificial intelligence models could power the centers with high-solar microgrids in the southwest U.S., researchers found. The power demand for such datacenters is estimated at 15 GW to 150 GW by 2030.

December 20, 2024 William Driscoll

Feasible land for 90% solar off-grid microgrids in the southwest U.S.

Image: Paces

Researchers have identified land parcels in the southwest U.S. with the technical potential to host 1200 GW of off-grid solar-plus-storage with gas backup.

Off-grid microgrids with 44% solar are economically viable for datacenters focused on training new artificial intelligence models, researchers found, while microgrids with up to 90% solar may be economically viable for customers that seek to limit their carbon emissions.

Five researchers, from the firms Paces, Scale Microgrids and Stripe, reported their findings in a white paper titled “Fast, scalable, clean, and cheap enough: How off-grid solar microgrids can power the AI race.”

The paper focuses on using off-grid systems to power artificial intelligence datacenters used for training. Artificial intelligence begins with training a new AI model; once the model is trained, it is used for “inference” in commercial applications.

While datacenters used for inference require proximity to end-users, training datacenters are more geographically flexible and may be located in areas with high solar potential and inexpensive land, the authors said.

Total AI power demand by 2030 ranges from 30-300 GW, according to analysts the authors cited, and training datacenters likely represent about half of that demand, they said. That translates to 15-150 GW of AI training demand by 2030, representing a small portion of the sites the researchers identified totaling 1200 GW of potential off-grid capacity.

For off-grid systems that provide up to 90% of lifetime hourly energy demand with solar-plus-storage, the costs “are quite competitive” with costs of off-grid systems powered by other means, the study said.

For a “significantly lower levelized cost of energy (LCOE) than the Three Mile Island nuclear reactor restart, you can get a microgrid that is 90% renewable,” the study found, referencing Microsoft’s costs under its contract to buy power from Three Mile Island.

Or, for “nearly the same LCOE as an off-grid natural gas turbine, you can get a microgrid that’s 44% renewable.”

The authors raised the possibility that while “the uptime requirements for AI training datacenters are still not entirely clear,” they “may offer opportunities to eliminate generators entirely.”

While the locations identified by the study were all close enough to natural gas pipelines to permit gas generation backup, the paper said that if diesel generation backup were used instead, “you can build almost anywhere with good sun.”

The authors posed and answered the question “If this is so great, why isn’t it happening?” One hurdle, they said, is inertia: “the fact that this hasn’t been done before.” Cost is also an issue, although the cost of a 44% renewable system is “basically at parity with gas only and offers a valuable hedge on fuel price risk.”

The third hurdle is that “massive datacenters dedicated to training only are a recent phenomenon, and datacenter designers have historically been skeptical of off-grid solutions due to the perceived need to optimize for uptime reliability.”

Staff of the three firms served different roles in preparing the white paper. Paces conducted a search for all land in the southwest U.S. that could accommodate solar microgrids with gas backup (shown in green on the featured image above). Scale Microgrids performed a levelized cost of energy analysis. Staff of Stripe conceived of the research project, assisted with modeling, and jointly drafted the white paper with the other co-authors.

The paper says that 95% of the identified land parcels are privately held.

This content is protected by copyright and may not be reused. If you want to cooperate with us and would like to reuse some of our content, please contact: editors@pv-magazine.com.