| Work Detail |
Companies building data centers to train artificial intelligence (AI) models could power them with solar-powered microgrids in the southwestern U.S., researchers have found. The estimated energy demand for these data centers ranges from 15 GW to 150 GW by 2030.
Researchers have identified plots of land in the southwestern US with the technical potential to host 1,200 GW of off-grid solar plus gas-backed storage.
According to the researchers, off-grid microgrids with 44% solar power are economically viable for data centers dedicated to training new artificial intelligence models, while microgrids with up to 90% solar power may be economically viable for customers looking to limit their carbon emissions.
Six researchers from Paces, Scale Microgrids, and Stripe reported their findings in a report 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 AI data centers used for training. AI begins with training a new AI model; once the model is trained, it is used for “inference” in commercial applications.
While data centers used for inference require proximity to end users, training data centers are more geographically flexible and can be located in areas with high solar potential and cheap land, according to the authors.
According to the analysts cited by the authors, total AI power demand by 2030 is between 30 and 300 GW, with training data centers likely accounting for around half of that demand. This translates to 15-150 GW of AI training demand by 2030, which represents a small portion of the sites identified by the researchers, which add up to a total of 1,200 GW of potential off-grid capacity.
According to the study, the costs of off-grid systems that provide up to 90% of lifetime hourly energy demand with solar plus storage “are quite competitive” with the costs of off-grid systems powered by other means.
For a “levelized cost of energy (LCOE) significantly lower than restarting the Three Mile Island nuclear reactor, a 90% renewable microgrid can be achieved,” according to the study, which references Microsoft’s costs under its Three Mile Island power purchase agreement.
Or, for “about the same LCOE as an off-grid natural gas turbine, you can get a microgrid with 44% renewables.”
The authors raised the possibility that while “uptime requirements for AI training data centers are still unclear,” they “may offer opportunities to eliminate generators altogether.”
While all of the locations identified by the study were close enough to natural gas pipelines to allow for backup gas generation, the paper said that if backup diesel generation were used instead, “it can be built almost anywhere with good sun.”
The authors posed and answered the question, “If this is so great, why isn’t it being done?” One obstacle, they said, is inertia — “the fact that this hasn’t been done before.” Cost is also an issue, though the cost of a system with 44% renewables is “basically on par with gas alone and offers a valuable hedge against fuel price risk.”
The third hurdle is that “massive data centers dedicated solely to training are a recent phenomenon, and data center designers have historically been skeptical of off-grid solutions due to the perceived need to optimize uptime reliability.”
Staff from all three companies played different roles in preparing the white paper. Paces conducted a search for land in the southwestern U.S. that could host solar-powered microgrids with gas backup (in green in the image above). Scale Microgrids conducted a levelized cost of energy analysis. Stripe staff conceived the research project, assisted in the modeling, and wrote the white paper with the other co-authors.
According to the document, 95% of the plots identified are private. |
