Over the past 12–18 months, Tech giants and U.S. utilities have ramped up investments in alternative power sources—particularly nuclear energy and on-site natural gas fuel cells—to address the growing power needs of data centers. Recent deal announcements (Meta with Vistra/Oklo/TerraPower and AEP with Bloom Energy) underscore an accelerating trend, reinforcing Powering AI and Nuclear Power as among the most attractive themes for 2026.
Meta announced Friday a series of landmark agreements to secure up to 6.6 GW of nuclear power by 2035 for its data centers. These include long-term deals with electric utility Vistra covering over 2.6 GW from existing nuclear plants in Ohio and Pennsylvania, as well as backing new SMR (Small Modular Reactors) projects with SMR specialists Oklo and privately-held TerraPower. Specifically, Oklo will develop a 1.2 GW power campus in Ohio and TerraPower will deploy two reactors to generate up to 690 MW of power as early as 2032, and provide Meta with rights for energy from as many as six other reactors by 2035. In both cases, Meta will prepay for power and provide funding, significantly de-risking the two nuclear start-ups and further validating their technology.
Notably, these agreements far exceed Meta’s previous nuclear supply deal with Constellation (1.1 GW) and the prior “record” among tech giants (1.9 GW for Amazon/Talen), marking a clear step change in power needs and procurement.
Likewise, there was a change of scale in the deal between U.S. utility AEP and fuel-cell manufacturer Bloom Energy. AEP, which had purchased in November 2024 100 MW of Bloom’s fuel cells, exercised its option to acquire additional 900 MW for a total commitment of $2.65 billion. This deal, along with the recent ones with Oracle and infrastructure investor Brookfield, clearly positions Bloom’s fuel cells as a favored on-site energy supply solution for utilities and data centers seeking to bypass grid constraints.
This acceleration likely reflects upcoming “titan” data centers (e.g. Meta’s Prometheus and Hyperion) over the next few years, as well as the rapidly rising power consumption of AI chips. Industry sources expect Nvidia’s next Rubin generation to consume roughly 2,300 W, compared with 1,000–1,400 W for Blackwell, an increase close to 100%! Here again, the change of pace is impressive, as the power consumption increase of AI chips over 2020-25 was closer to 30% per year.
In conclusion, we believe that AI’s electric power challenge remains in its early stages. While data centers continue to pursue efficiency gains, the growing number and scale of these AI factories, combined with increasingly power-intensive chip architectures, should continue to drive sustained demand for reliable and scalable power solutions.
