Can the U.S. power the AI boom?

What happened

The AI race is slamming into the physical limits of the U.S. power system.

Hyperscalers are building data centers at a pace that did not exist two years ago. In Q3 2025, U.S. data center leasing exceeded 7 GW, with another 10 GW under active negotiation. This is more than double last year’s run rate.

The same five giants are behind the surge: OpenAI , Anthropic , Amazon, Alphabet and Microsoft. Each is announcing a rollout of multi-gigawatt campuses as casually as launching a new product line. OpenAI’s U.S. roadmap alone is approaching 10 GW. Amazon’s Project Rainier combined with Anthropic’s long-term buildout adds roughly 15 GW on top.

To put the scale in context: 1 GW can power around one million U.S. homes.
The capacity announced in 2025 could eventually require as much electricity as an entire developed country.

We are past the question of whether AI will consume more power. The only real question is whether the U.S. can build enough generation, transmission and firm capacity in time to prevent the AI boom from becoming an energy bottleneck.

Impact on our Investment Case

How big is the U.S. power shock?

The past year flipped the entire electricity debate on its head. Suddenly everyone in finance, tech and utilities is talking in gigawatts and terawatt hours. The problem is that most people mix the two concepts, so here is the simplest way to reset the basics before looking at the scale of what is coming:

• A terawatt-hour measures energy, the total amount consumed over time.
• A gigawatt measures power, the instantaneous output of a plant.

If a plant runs at 1 GW for a full year, it produces about 8.8 TWh.
The catch is that few power generating technologies run all year. Solar averages about 25%, wind around 35%, while gas and nuclear usually run above 85%. This gap between nameplate capacity and firm capacity is the heart of the story, because data centers need power that is available every single hour.

The challenge is that nobody truly knows how large the new wave of AI-driven demand will be. Forecasts for U.S. data center electricity use in 2030 vary more than for any major energy category. Some estimates cluster around 350 to 400 TWh, others point to 800 TWh or more, and the most aggressive cases exceed 1,000 TWh. The spread reflects deep uncertainty: faster training cycles, heavier multimodal inference, synthetic video generation, and the rise of agentic workloads. It is similar to trying to predict future global mobile data traffic in 2008.

A reasonable mid-range view puts U.S. data center demand near 500 to 700 TWh by 2030. That means roughly 320 to 520 TWh of extra load compared to today.

This sits inside a broader shift. After almost two decades of flat consumption, the U.S. is entering a real electricity demand boom. The 5-year CAGR for expected U.S. electricity load (utilities' forward estimate of how much power their customers will require) has surged from 2.6% in 2023 to 5.7% today. Utilities now expect total U.S. electricity use to rise by 800 to 1,000 TWh by 2030. Data centers alone account for nearly half of that increase. Nothing in modern U.S. energy history compares to one single load category driving this much new consumption.

Translating energy into capacity gives the real scale of what must be built. Meeting 800 to 1,000 TWh of incremental demand requires roughly 90 to 115 GW of new firm capacity by 2030.

This is not a small step up. It is a shock to a system that has grown far too slowly for far too long.

Can generation scale in time?

To know whether the U.S. can add enough power by 2030, the first step is to look at what the system has actually been adding.

Over the past six years, net additions bounced from almost zero to a bit above 40 GW per year. Coal retirements alone removed 4 to 15GW annually. Nuclear stayed flat until life extensions began in 2023. Gas barely grew, rarely exceeding 6 GW of net additions and even turning negative in 2024. Virtually all real growth came from solar, wind and, more recently, storage.

The challenge is that meeting this demand requires a broader mix of capacity. Data centers need firm power, which renewables support but cannot provide alone, so the real question is what can realistically be added in the next five years given supply chains, construction timelines and developer execution.

A fair but optimistic buildout looks like this:

• Coal: roughly net zero per year. This assumes retirements fall sharply, with life extensions keeping units online and a few restarts offsetting inevitable closures. It is a bullish stance versus the historical 10GW per year decline, but somehow aligned with the policy shift.
• Natural gas: around 10 GW per year. This is higher than the past decade but aligned with current orders and the rise of on-site generation at AI campuses. Developers are installing gas turbines and modular fuel cells to bypass interconnection queues (the long waiting lines to connect new projects to the grid).
• Nuclear: zero net additions. Life extensions help energy availability but do not deliver new nameplate capacity before 2030. Small modular reactor (SMR) announcements represent 5 to 15 GW of potential commitments, but none of that will become operational in this decade.
• Wind: about 10 GW per year, consistent with historical installation rates.
• Solar: around 40 GW per year. This is more than twice the five-year average and above any historical record, but the module manufacturing build-out makes it plausible.
• Battery storage: roughly 20 GW per year. Useful for shifting energy, but it does not create firm power and therefore is not counted as firm capacity.

Convert this mix into firm output using standard capacity factors, and you get about 22 to 23 GW of new firm power per year. Over five years, that is roughly 110 to 115 GW. In other words, just enough to cover mid-range U.S. demand growth by 2030.

The catch is that everything must fire at the high end of its potential: record solar deployments every year, high gas turbine throughput, no slowdown in wind, and coal retirements held in check. There is almost no margin for permitting delays, supply chain hiccups or policy shifts.

For context, the U.S. has never added more than 40 GW of net capacity in a single year. Moving toward 50 to 60 GW of net additions every year through 2030 would be a structural acceleration across manufacturing, permitting and project execution. It is possible, but it requires the entire system to run at full speed, without missing a beat.

Can the grid deliver it?

Even if the U.S. manages to build enough generation, the electricity still needs to reach the data centers. That is where constraints become structural. Connecting a large load requires grid studies, substation upgrades, breakers, transformers and high-voltage switchgear. These components are now stretched to multi-year lead times.

Lead times for the most critical equipment now stretch across two to three years for grid transformers, around a year for high-voltage switchgear, and several months for wire and cable. These are not minor bottlenecks. They dictate the actual pace at which any new project can connect, whether it is a wind farm, a gas turbine or a hyperscale data center.

Transmission adds another layer of constraint. Last year, the U.S. added only 888 miles of new high-capacity transmission. The Department of Energy estimates the country needs about 5,000 miles per year to keep up with electrification and rising industrial load. Yet building a major transmission line typically takes 8 to 10 years, while hyperscalers build campuses in 18 to 24 months. The timelines simply do not match.

This is why developers are shifting toward a “build next to your load” approach. By installing private substations and, in some cases, on-site gas turbines or modular fuel cell blocks, hyperscalers can bypass the interconnection queue, which is the most congested step in the process. However, this is not a complete workaround. These installations still depend on medium and high voltage transformers, switchgear and protection systems, all of which have long lead times and multi-year backlogs.

Even with private infrastructure, not every location can be served in this manner. Some campuses still depend on utility substations or must tap existing transmission corridors that are already saturated.

It is difficult to precisely quantify how much grid bottlenecks will slow data center development. As a reference point, the latest IEA World Energy Outlook estimates that around 20% of expected U.S. data center additions through 2030 are at risk of delay due to grid congestion, interconnection backlogs and equipment shortages. In other words, even if generation ramps successfully, the grid remains a serious constraint.

How we’re positioned

The data center buildout is global, but the center of gravity is the United States. Roughly half of all new hyperscale projects in the pipeline are located on U.S. soil, and this is where grid congestion, transformer shortages and on-site generation are becoming the defining bottlenecks. Positioning the portfolio around the U.S. infrastructure stack gives us the highest visibility, the strongest regulatory support and the clearest pricing power.

On the generation side, we focus on the segments that benefit directly from the scale-up in utility-scale clean power. First Solar and Nextracker Inc remain the two U.S. leaders best-aligned with Made in America incentives, domestic manufacturing credits and a growing wave of long-term solar procurement. On the firm power side, GE VERNOVA INCO. anchors our exposure to gas turbines and long lead equipment. The return of gas as a bridging fuel for data centers continues to feed a healthy order pipeline.

Our grid exposure is even more deliberate. We maintain significant positions in U.S. EPC contractors and in Korean grid equipment manufacturers that supply the transformers, switchgear, protection systems and digital controls required to connect new load. These companies sit in the most structurally constrained part of the value chain, supported by multi-year backlogs, limited global manufacturing capacity and sustained pricing power.

In short, the portfolio is built to capture both sides of the AI-driven power cycle: the acceleration in clean and gas-based generation, and the chronic grid infrastructure shortage that now shapes the economics of the entire sector.

Our Takeaway

Since our May 2024 note, every major forecast has been revised higher. U.S. load expectations have doubled. Data centers are on track to drive roughly half of all new U.S. electricity demand this decade. This pulls forward investment in solar, storage, gas turbines and, above all, the grid.

Our analysis uses a mid-range scenario, but it is important to note that outcomes could land lower or considerably higher depending on model sizes, training cycles and the speed of AI adoption.

This is not simply an AI hardware cycle. It is the early stage of a restructuring of the U.S. power system. Data centers are starting to behave like small utilities. Electricity is turning into the limiting resource. The companies that generate it, move it and stabilize it will define the next ten years of energy infrastructure spending.

Our Sustainable Future strategy is positioned for this shift. Grid infrastructure represents about 36% of the strategy, compared with roughly 13% for peer cleantech funds. Another 17% is allocated to U.S. power generation segments with the fastest growth profiles, led by solar and battery storage.

All holdings are profitable companies with strong technology, pricing power and operational track records. We intentionally avoid early-stage nuclear developers, pre-revenue SMR concepts, fuel cell manufacturers and other speculative names that depend on aggressive data center growth to justify their valuations.

We apply the same discipline in our AI & Robotics strategy. There, we focus on the growth of inference, vertical applications and the data layer. Sustainable Future complements this by owning the power infrastructure that makes that expansion possible: generation, storage and the grid.

Whether data center capex accelerates or normalizes, the structural pressures on the U.S. power system remain: grid congestion, transformer scarcity, electrification needs and long interconnection timelines. These issues predate hyperscalers and will persist regardless of AI cycles, keeping demand strong for the companies that provide the critical hardware.