Is it possible that the digital infrastructure and energy/power industries are overestimating the need for data centers, and for development of new energy facilities to be used to power them?

The growth of generative artificial intelligence (Gen AI) is expected to account for 3.5% of global electricity consumption by 2030, according to Gartner, a consultancy. According to a Goldman Sachs forecast, by 2030, data center-driven power demand is expected to reach 272% of the 2023 level (397 TWh compared to 142 TWh), and, of the additional 255 TWh of demand, 89 TWh will be attributable to additional AI demands. These forecasts are predicated on the anticipated non-stop explosive surge in development and adoption of Gen AI technologies.

The bullish forecasts, however, have their skeptics. According to the skeptics, two types of constraints should make Gen AI enthusiasts at least take a second look.

The first type relates to the capabilities of Gen AI. According to Daron Acemoglu, last year’s economics Nobel laureate, the productivity and GDP gains from the use of Gen AI may be less than otherwise forecasted, because Gen AI may not prove to be as efficient at performing more complex tasks that do not presuppose a defined desired outcome, compared to the easier tasks for which productivity improvements have been observed so far.

Acemoglu’s concerns are seconded by Jim Covello, head of Global Equity Research at Goldman Sachs. According to Covello, Gen AI is currently poised to function as a high-cost replacement of already relatively low-cost labor, which is the opposite of prior transformative innovations. Accordingly, without significant reduction in cost or output capabilities, Gen AI’s peak may come sooner than expected.

Constraints of the second type are limitations and bottlenecks on Gen AI development, such as concentrated supply of GPU processors. Taiwan Semiconductor Manufacturing Corp., which dominates global GPU production, projects AI chip shortages throughout 2025 and probably 2026, and operates in the shadow of the geopolitical risks constantly surrounding Taiwan.  Separately, power supply constraints may in the short- to medium-term restrict the development of Gen AI capabilities by limiting the implementation of sufficient data center resources and other digital infrastructure needed to optimize the evolution of Gen AI.

Related to the economic constraints are the legal uncertainties that, if resolved adversely, may end up limiting the use of Gen AI or increase its cost, thereby limiting the potential economic advantages related to Gen AI. As of today, regulators generally take a preventative risk-based approach depending on the risk and scale of AI systems and models employed, generally setting the guardrails for the Gen AI development process and outright prohibiting only a minority of AI systems (such as certain biometric- and human-behavior related AI systems prohibited by the EU AI Act or California prohibitions on certain types of deepfakes, some of which are currently stayed by a U.S. district court). Yet, by some appraisals (although disputed as an overestimation, the costs for developers to comply with the EU AI Act may reach EUR 31 billion and result in a 20% reduction of AI investments over a five-year period, making Gen AI significantly less cost-effective and potentially slowing down the pace of Gen AI development and implementation.

Still, it is not a purely theoretical risk that the regulatory environment may take a less favorable turn, resulting in limitations on Gen AI development or use. As of today, there are at least 20 infringement cases being heard across the U.S. against Gen AI development companies. If determined adversely, the case law may curtail the developers’ ability to train their models. Further, one should not discount the risk of a single high-impact event that is yet to occur as a prompt for more aggressive regulation, similarly to how the Enron scandal led to the passage of Sarbanes-Oxley Act in 2002.

Where do these uncertainties leave digital infrastructure and energy providers? The good news is that neither is an industry captive to Gen AI. Electric power is set to remain a universal commodity redeployable in more traditional fields such as transportation and manufacturing, in which power generation and transmission capacities are currently acting as major challenges of their own. Similarly, it is conceivable that, should the scale of Gen AI industry be reduced, the freed-up capacity will, in itself, provide a boost of alternative use.

If the Gen AI optimists are right, we may expect complete transformation of our lives by one of the most impactful technologies ever conceived. If they are wrong, however, the infrastructure impetus it has unleashed may still be put to constructive uses beyond its initial purpose.

—Neil Torpey is a partner in Corporate Department at Baker Botts, specializing in mergers and acquisitions, private equity, venture capital, and capital markets. Torpey has executed transactions across the U.S., Asia, Latin America, and Europe. Nikolai Gryzunov is an associate with Baker Botts, specializing in corporate transactions including mergers and acquisitions, corporate reorganizations, joint ventures, debt financings, and securities offerings. He began his career in Moscow, advising domestic and international clients on corporate, finance, and regulatory matters.