Power Demand and the Investment Implications Across the AI Value Chain
As artificial intelligence moves from experimentation to industrial-scale deployment, the electricity required to power next-generation data centers is, in our view, creating a compelling investment opportunity set across utilities, generation, infrastructure, and technology.
A New Demand Cycle for Electricity
For roughly two decades, U.S. electricity demand was largely flat. Efficiency gains in appliances, the offshoring of manufacturing, and the replacement of aging coal plants with cheaper natural gas helped create an environment in which utilities had little need to build new power plants. In some markets, excess supply even forced certain generation assets, including nuclear plants, to shut down. Artificial intelligence is changing that equation quickly.1
We believe the emergence of AI data centers has introduced a durable source of incremental load growth. Utilities are now seeing credible evidence of sustained demand through long-term contracts, large-scale site commitments, and substantial infrastructure investments by hyperscalers.
As a result, AI power demand appears durable to support a multiyear investment cycle for regulated utilities, creating opportunities to invest in transmission, substations, grid modernization, and off-grid technologies. Because utilities earn regulated returns on capital invested in their systems, a higher-growth demand environment could support stronger earnings growth. Equally important, large data-center customers may help fund upgrades that were already needed across aging grids, potentially spreading fixed costs over a larger customer base and easing affordability pressures over time.
Why AI Requires so Much Power
The driver is not simply the growth in data centers; it is the intensity of compute inside them. Traditional enterprise data centers often operated at roughly 5 to 10 kilowatts per rack, while cloud data centers typically rose to 15 to 30 kilowatts. Current-generation AI systems built around advanced GPUs can require more than 100 kilowatts per rack, and future designs are expected to push density even higher.
At the same time, the size of AI clusters is expanding. A large data center once meant a facility requiring 50 to 200 megawatts.2 Today, hyperscalers are discussing campuses measured in gigawatts. One gigawatt of demand is roughly equivalent to the power needed for about one million homes. In our view, that scale helps explain why AI has moved power from a background operating cost to a central strategic bottleneck.
Efficiency improvements in chips and systems will matter, but they may not fully eliminate the demand challenge in the near term. Each new generation of GPU may deliver more compute per watt, yet overall demand may continue to rise as AI requires substantial computing capacity not only to train increasingly advanced models, but also to support inference, i.e., the everyday use of trained models to generate outputs at scale. In practical terms, this appears to suggest that efficiency gains are being offset by accelerating usage, a dynamic consistent with the idea that cheaper or more efficient compute can stimulate more consumption rather than less.
Supply Constraints Are Real, but Solvable
The challenge is not whether enough power can eventually be built, but whether it can be delivered fast enough. Large nuclear projects typically take many years to build and have historically faced cost overruns and delays. Gas plants can be built faster, but turbine backlogs remain a constraint as new capacity may not arrive until the end of the decade. Transmission interconnections, permitting requirements, and local approvals add further complexity.
As a result, the solution set may continue to broaden. Existing coal and gas plants may run more frequently, and some coal retirements may be delayed until cleaner and more reliable resources are available. Renewables and batteries can be deployed more quickly, although they do not always provide round-the-clock power. Modular and behind-the-meter solutions such as reciprocating engines, aeroderivative turbines, fuel cells, and other on-site generation are also gaining traction because they may provide deliver faster access to power.
This “all of the above” approach is important for investors. AI power demand does not create a single winner. It creates opportunities across regulated utilities, independent power producers, grid equipment, cooling, construction, and specialized power solutions. In our view, the clearest beneficiaries may be those positioned around the most acute bottlenecks.
The Global Power Investment Opportunity Set
Where Growth and Value Perspectives Meet
For growth investors, AI-driven power demand represents a major rate-of-change event. Historically, many growth portfolios focused on consumer technology, software, semiconductors, and cloud infrastructure within the AI ecosystem. But as AI model development pushed GPU roadmaps forward, it became clear that each generation of compute would require more physical infrastructure and more electricity. That realization expanded the opportunity set into areas traditionally viewed as value oriented.
Independent power producers with existing fleets, particularly nuclear and gas assets in competitive markets, were early beneficiaries because they could offer immediate access to scarce power. Long-term contracts with hyperscalers helped validate the demand story and supported higher valuations. More recently, however, concerns around affordability, regulatory treatment, and grid reliability have caused uncertainty. Against this backdrop, policymakers and grid operators are working through rules to ensure data centers either bring incremental power generation or accept some form of curtailment during system stress.
For regulated utilities, the opportunity is different but potentially more durable. Data center load can support capital investment and earnings growth, provided rate structures are designed so large customers pay for the incremental infrastructure they require. In several regions, utilities have developed special tariffs intended to protect residential customers while still enabling hyperscaler development. If executed well, that structure can turn data center demand into a source of system-wide investment rather than a burden on existing customers.
Politics, Affordability and Community Acceptance
Power demand is now as much a political challenge as it is an infrastructure issue. In competitive markets, rising electricity prices have raised concerns that data centers are increasing bills for households and small businesses. While higher prices may be necessary to incentivize new generation, elected officials are sensitive to affordability, particularly in states where residential bills have already moved higher.3
Community acceptance is another constraint. Data centers can face local opposition related to land use, water consumption, noise, and visual impact. Hyperscalers increasingly appear to recognize that paying for grid upgrades is only the starting point. To win support, they may also need to demonstrate direct community benefits, including local investment, tax revenue, workforce development, and support for schools or public services. The more tangible those benefits become, the more likely communities are to view data centers as economic assets rather than outside intrusions.
For investors, these political dynamics may slow deployment but not necessarily undermine the cycle. In fact, regulatory friction could moderate the pace of buildout, potentially extending the investment period and helping to reduce the risk of a short, overheated surge. The critical question is whether policy frameworks can balance reliability, affordability, and the strategic importance of AI infrastructure.
Beyond Power: The Next Bottlenecks
The AI infrastructure buildout is also creating opportunities beyond generation. Electrical infrastructure is one of the most visible. Substations, transformers, switchgear, transmission equipment, and related components are facing long lead times as utilities and data center developers compete for supply. Cooling is another major area of change. As rack density rises, traditional air cooling becomes less effective, pushing the industry toward liquid cooling and other advanced thermal-management solutions.
Labor may be the least discussed bottleneck. Building data centers, substations, transmission lines, and power plants requires skilled electricians, engineers, and construction labor at a time when many of those trades are already constrained. Physical infrastructure, not just chips or software, will determine how quickly AI capacity can come online.
Investment Implications
In our view, AI power demand is reshaping how investors should think about the technology value chain. Semiconductors and cloud platforms remain central to the AI story, but the next phase may increasingly depend on electricity, grid capacity, cooling, and physical deployment. We believe that could create a broader and more diversified opportunity set than the first phase of the AI trade.
In our view, the most attractive opportunities are found where demand is durable, supply is constrained, and companies have pricing power or regulated visibility. Regulated utilities could benefit from higher capital investment and load growth. Independent power producers can benefit from scarce existing generation and long-term contracts. Equipment suppliers and cooling providers can benefit from bottlenecks in the buildout. Behind-the-meter and modular power providers may gain share as hyperscalers seek speed and flexibility.
At the same time, risks are meaningful. Demand forecasts could prove too aggressive, regulation could limit returns, community opposition could delay projects, and technology efficiency could alter the path of load growth. But in our view the broad direction appears clear: AI is no longer only a digital infrastructure story. It is an energy infrastructure story as well.
Sources
1 Source: IEA, Energy supply for AI
2 Note: A watt is a unit of power. One megawatt equals one million watts; one gigawatt equals one billion watts, or 1,000 megawatts.
3 Source: Yale University, The Politics of AI: Will Bipartisanship Last or Is Polarization Inevitable?
