BlockBeats reported on July 21 that independent semiconductor and AI research firm SemiAnalysis said fast-rising AI compute demand is changing the power supply model for data centers.
According to the firm, reciprocating engines, which were previously used mainly as backup power during grid failures, are increasingly being repositioned as round-the-clock baseload power sources.
Signed behind-the-meter projects are growing
SemiAnalysis said reciprocating engine manufacturers have signed contracts this year to supply about 1GW of behind-the-meter power projects. It expects annual new supply to top 4GW in both 2027 and 2028.
U.S. grid headroom may run out between 2027 and 2028
After modeling U.S. grid carrying capacity, SemiAnalysis said existing power headroom is likely to be exhausted between 2027 and 2028. It added that planned utility-scale generation capacity through 2030 is still not enough to cover the additional load expected from data centers.
About 140GW of potential projects still lack power contracts
Using its data center model, SemiAnalysis estimated that around 140GW of potential data center projects have not yet signed power supply agreements. Those projects may end up using behind-the-meter power at scale to bypass grid expansion bottlenecks.
Reciprocating engines seen taking the largest share
Among several behind-the-meter technologies, including reciprocating engines, aeroderivative gas turbines, and fuel cells, SemiAnalysis expects reciprocating engines to account for the largest share.
It cited lower cost, faster deployment, modular expansion, and stronger financing capacity. The firm also said equipment makers such as Caterpillar, INNIO, and Cummins are expanding production capacity to support larger-scale deployment in the next few years.
In SemiAnalysis' view, as AI data centers continue to expand, on-site self-generation is shifting from a traditional backup facility to core energy infrastructure, with reciprocating engines emerging as a potential way to help close the power gap in the compute era.