Energy and the AI buildout: An investor’s perspective

Key takeaways

• The AI infrastructure buildout is revealing capacity constraints in many key inputs, with power being one of the most strained.
• We see the energy system adjusting across multiple time horizons, solving for immediate gaps while expanding the structural base.
• The upshot: Power constraints are unlikely to impede the AI buildout.
• For investors, a more dynamic environment means greater rewards may be achieved by shifting from a broad AI lens to more precise stock selection in the theme.

A once-in-a-generation industrial buildout is underway. Infrastructure to power artificial intelligence (AI) requires semiconductors, equipment, labor, data centers ― and massive amounts of power.

We estimate approximately 148 gigawatts (GW) of additional power capacity will be needed by the end of the decade to satisfy data center demand. This is multiples above the roughly 42GW in power capacity consumed by data centers in 2025 and represents a growing share of total demand, as shown in the chart below.

While power is a feared bottleneck in the AI buildout, our analysis finds that the system is adapting across both short- and long-term horizons, easing concerns of an energy-deficiency-led AI slowdown.

A collision of two disparate worlds

The increasing use of power to build AI Infrastructure has forced a collision between two radically different operating cultures.

The AI ecosystem, including labs, hyperscalers and semiconductor companies, moves quickly, with competitive and financial incentives for rapid iteration, aggressive capital deployment and first-mover advantage.

The power complex, including utilities, grid operators, regional markets and equipment manufacturers, has seen little structural growth for nearly two decades and is optimized for reliability, regulatory compliance, capital preservation and risk minimization.

When unprecedented demand collides with slow-moving, risk-averse institutions, strain is inevitable. The response is creativity.

The AI ecosystem is shifting to new geographies, underwriting long-term contracts, and pushing supply chains to expand. Developers are also reducing grid dependence by building generation alongside compute, a trend known as Behind the Meter or Bring Your Own Power (BYOP). Projects launch with anchor capacity and then scale in phases, accelerating capacity beyond what the grid can deliver.

Power constraints as an AI spoiler?

Against this backdrop, will power constraints derail the AI buildout? Our comprehensive review of ongoing and future power supply needs for the AI data center buildout leads us to answer “no.”

Near-term constraints are being addressed through pragmatic solutions. Our analysis finds the power supply systems fueling AI are more flexible than commonly assumed. Gas turbine manufacturing is already scaling and can accelerate further, solar is growing rapidly, and emerging technologies such as fuel cells are contributing. Batteries play a growing and meaningful role by raising the yield of power-generating sources. Through 2026–2027, available capacity, ongoing buildouts and interim solutions appear sufficient. After 2027, the system is more stretched, but uncertainty on both sides is higher.

In the out years of 2029-2030, we see incentives and innovation as likely combining to meet modest shortfalls. The system is adjusting across multiple time horizons, solving for immediate gaps while expanding the structural base.

2030 and beyond: Two long-term solutions

As we move into the 2030s, two developments have the potential to materially change how AI infrastructure is powered and scaled. One is nuclear energy, and hyperscalers are already contracting for output tied to extending, refurbishing or restarting existing plants.

The second development is data centers in space, a concept moving from theory to exploration. SpaceX has proposed orbital data centers powered by continuous solar exposure, leveraging natural cooling and eliminating terrestrial grid constraints. Google, Amazon and others are evaluating similar concepts. With reusable launch systems reducing launch costs and increasing payload capacity, deploying large-scale orbital compute over time appears technically plausible, introducing an entirely new supply frontier.

A self-regulating system

To encapsulate our broad view: Power constraints are unlikely to derail AI data center deployment on their own. Over the next two years, the gap is bridged by a mix of in-flight capacity, gas-led additions, storage, BYOP and site re-optimization. Beyond 2027, the challenge becomes less about, “Is there enough power?” and more about how quickly the system can translate capital into energized capacity.

Importantly, we believe these frictions create a self-regulating mechanism: They cap the pace of expansion, prioritize the highest-return projects, and reduce the risk of indiscriminate overbuilding.

For investors, this means the opportunity is shifting from broad “AI demand beta” to selective stock exposure. In an environment with dynamic bottlenecks and timelines, dispersion should rise, and we believe an active, nimble approach is best positioned to capture the winners.

The above offers a summary of our thinking on AI power needs and solutions. For comprehensive analysis and commentary, please refer to the full report.