Before the majority of individuals in the rest of the nation have ever seen the abstract, a specific type of scholarly article can change billions of dollars in market value once it reaches the appropriate inboxes. This spring, Stanford’s Electrical Engineering department has produced a number of new results aimed at significantly lowering the electricity usage of AI tasks. Investors in U.S. utility stocks have started to become uneasy about the research, which is based on a multi-year initiative centered around analog and in-memory computing paired with optical data transmission. There hasn’t been a disastrous selloff. It has been consistent. It has also been instructive.
Although the Stanford work is technically complex, it is surprisingly simple to describe its ramifications. Today, the majority of AI processing takes place on digital silicon, with data moving over electrical interconnects between memory and processors. A significant portion of the energy used by a data center is caused by this shuttling. The Stanford study focuses on two complimentary strategies that could significantly reduce that energy bill when combined. Much of the shuttling is eliminated by in-memory and analog computing, which conducts computations within the same physical components that contain the data. By using light instead of electrical signaling, optical transmission transfers data over greater distances at a cheaper energy cost. Both concepts have long existed in scholarly literature. Making them cooperate at sizes pertinent to commercial AI training is the goal of the Stanford contributions.
The timing of the study has caused discomfort for utility investors. The utility industry has been one of the most anomalous stock market performers in any category over the last two years, surging substantially on the theory that AI data center buildouts would result in a multi-decade boom in power demand. Talen Energy, NRG, Constellation Energy, Vistra, and a few other names were traded at multiples that would have been deemed unjustified in any typical utility setting. The thesis was simple. Large amounts of electricity were required by hyperscalers like Microsoft, Amazon, Google, and Meta. For years, utilities with generation capacity close to large data center clusters would get premium payments for such power under long-term contracts. Investors enthusiastically embraced the narrative.
The Stanford results are insufficient to refute the premise on their own. The gap between present data center efficiency and the theoretical limits suggested by the research will take years to close, and the energy demand from AI is continuously growing. More significantly, the results have added uncertainty to the growth estimates that were previously certain. The multibillion-dollar infrastructure expenditures utilities have been increasing to serve hyperscaler demand may end up being overbuilt if AI workloads can eventually run on a fraction of today’s electricity per productive computation. Real selling pressure has been generated by that one doubt, which was inserted into the analyst models of the biggest utility holdings.
Additionally, the more general “DeepSeek shock” has been exacerbating these worries. Earlier this year, the Chinese AI startup DeepSeek shocked international markets by proving that competitive AI models could be trained with a lot less processing power than the American hyperscalers had been assuming was required. Because it raised the possibility that the demand-side predictions supporting the entire AI capital expenditure boom were overstated, the announcement posed a somewhat greater immediate danger to U.S. utility valuations than the Stanford study. Another piece of evidence pointing in the same direction is the Stanford work, which arrived within the same general time frame. less energy used for each task. distinct chip designs. Growth was quieter than the bulls had predicted.
The aspect of stock pricing that is most readily apparent is the valuation reset that results from this reconsideration. During the buildout frenzy of 2024 and early 2025, utility names that had traded as “AI derivatives” have begun returning some of those gains. Profit-taking has been especially noticeable in businesses that have a lot of exposure to particular data center contracts as opposed to a variety of residential and commercial clientele. Institutional investors believe that the AI-utility trade’s easy phase is over and that businesses that can show consistent customer demand regardless of any one hyperscaler’s capex schedule will be rewarded in the next phase.
However, because the bulls have not disappeared, it is worth considering the opposite scenario. Because the addressable AI market is growing faster than per-workload efficiency can keep up, a number of prominent energy analysts have suggested that even with significant advances in AI efficiency, the overall electricity demand from data centers will continue to rise. The hyperscalers have not significantly reduced their intentions for capital expenditures. According to guidelines maintained by Microsoft, Google, Amazon, and Meta, data center buildouts are expected to continue at a nearly current pace through 2026. Even though the upside ceiling has been cut, the utility thesis still has runway if those estimates come true.
Duration is a practical question for long-term utility investors. Transmission upgrades, new generation capacity, transformer purchases, and other infrastructure investments being undertaken today are intended to last for thirty or forty years. Whether the demand for AI that warrants those expenditures will last long enough to recoup the expenses is the question. When combined with the DeepSeek results and the larger trend toward innovative chip architecture, Stanford’s study has added just enough doubt to that question to confound the math. Investors are now pricing in some of that complexity.