AI data centers lock in nuclear, doubts linger

What changes after a regulatory breakthrough is not just sentiment but the finance architecture around demand. In PJM, AI load is already being translated into executable obligations: the March 4 Voluntary Taxpayer Protection Pledge by seven hyperscalers created enforceable, legally binding financial structures under U.S. law, requiring new rate structures and long-term capacity payments whether power is used or not, reducing the risk that generation projects stall, while PJM anticipates a 5.2% capacity shortfall by 2027-2028, implying about $15 billion of new plant investment and giving utilities and investors a concrete target for project planning.

That demand pull is being forced onto a near-term clock. The Emergency Electricity Auction had to be completed before September 2026 with the intent that each data-center operator sign long-term electricity purchase contracts, turning AI power needs into bankable offtake, while Jensen Huang said at GTC 2026 that “the limiting factor regarding silicon is the ability to supply power to the chips,” underscoring why buyers need dependable supply; in the interview, the value proposition for SMRs and nuclear is tied directly to the operational needs of AI and data-center customers, with 24/7 reliability helping move the market from skepticism toward execution, because for those customers nuclear is “really the only power source” proven to provide 24/7 power for long periods and can run 18 to 24 months without refueling.

The hardest argument against SMRs is not about physics but project economics: critics say nuclear still has not shown it can be built cheaply or on schedule, and they see little proof that smaller designs change that record. On the Inside Data Centre Podcast, one speaker praised the modular concept as a possible way to “manufacture and bring down the cost,” yet still concluded, “I just haven't seen evidence of being cheap and… meeting its… deadlines,” echoing a broader view that recent nuclear builds have been defined by overruns rather than repeatable commercial execution.

That skepticism hardens when financing and market behavior are examined, because even supporters acknowledge SMRs become “very capital intensive” as projects move toward construction, while the hoped-for factory model still must clear licensing, safety, security, and waste requirements. In Floating SMRs, Market Skepticism, and the Path to Scale, the speaker said, “I think the market in 2026 doesn't believe it anymore,” pointing to stock declines including “NuScale… $13 per share… Last year it was at 50, now it’s down to 8,” “Oklo IPO… $15 went up to 136 last year, now down to 35,” and “X Energy IPO… at $30 per share, now at 15,” then adding that “the management of this company, especially at OKLO and Nano… They sold $150 million of their own shares.”

Behavior is already shifting from interest to pre-commitment. GE Vernova said, “In the first quarter of 2026 alone, GE Vernova signed 21 gigawatts of new gas equipment agreements. Only 2 gigawatts of that were firm orders. The other 19 gigawatts were reservations,” while “The company’s reserved slots rose from 43 to 56 gigawatts in a single quarter,” and it “told investors it expects to reach at least 110 gigawatts of combined orders and reservations by the end of 2026” — evidence that buyers are paying to secure future capacity before final configurations are set.

Utilities and hyperscalers are also rewriting contracts around AI load rather than waiting for supply to appear. Georgia Power said it “took a curtailment right in exchange” and “OpenAI committed up to 1,000 megawatts of flexible demand response,” which the utility says is “close to a third of the load” and helps it “serve peak demand” while projecting “about $950 million a year of incremental large-load revenue” and “$2.8 billion of customer savings between 2029 and 2031”; meanwhile, TradingView reported Constellation signed a 20-year nuclear power deal with Amazon, and one investor analysis argued, “Private Capital underwriting Nscale project debt or pre-listing equity should treat the $71 billion Phase 1 figure as unverified reporting and rebuild the per-megawatt stack from the chip line up,” showing capital is already underwriting early nuclear-linked projects.

What makes the TVA decision consequential is not simply that a reactor cleared another licensing step, but that it was publicly framed as a first-of-its-kind federal approval for a commercial SMR. PR Newswire underscored that point in the headline, “TVA Wins NRC Construction Permit for GE Vernova Hitachi BWRX‑300 SMR at Clinch River—first U.S. utility approval,” positioning the permit as a regulatory precedent that moves advanced nuclear beyond planning and into an execution stage under NRC authority at a named project site.

That interpretation was reinforced by Contrary Research, which described the NRC action as the first U.S. commercial SMR construction permit and therefore a broader signal about federal readiness to let advanced nuclear projects proceed. As it put it, “The Nuclear Regulatory Commission approved a construction permit for a 300 MW GE Vernova Hitachi BWRX-300 reactor in Oak Ridge, TN, which will be the first commercial small modular reactor in the US,” while noting the review “concluded four months ahead of schedule” and was “the NRC’s second new-reactor permit of 2026, after TerraPower’s Natrium unit in Wyoming.”