Energy could be the ultimate frontier for AI

The Jevons Paradox frames the dynamic well: as AI becomes cheaper, adoption soars but there’s a later part to this statement that goes “until supplies become a problem. Chips may no longer be the bottleneck, but electricity almost certainly will be.

As AI scales into trillions of models and mega-data centres, the constraint shifts from chips to power sockets.

The US grid has seen minimal growth over the past two decades. However, electricity consumption is now expected to increase at an average annual rate of 1.7% from 2020 through to the end of the short-term forecast in 2026, after growing by only a CAGR of 0.1% in the past 2 decades, driven by structural demand factors outlined below.

Demand Drivers

a)     Data centres

Today, training ChatGPT-3 reportedly consumed enough electricity to power around 120 US homes for one year, while GPT-4 likely consumed 50 times more — equivalent to approximately 6,000 household-years in a single run.

Mega tech companies are pouring money into building data centres and purchasing the most advanced semiconductor chips. Companies like Meta Platforms and Alphabet even raised their CAPEX forecast for the year 2025 even as most companies scaled back expenditure amid uncertain macroenvironment. Cumulatively, mega tech firms such as Meta Platforms, Alphabet, Microsoft and Amazon.com are expecting to spend more than USD 330 billion in CAPEX this year.

Individual GPUs are becoming more power-hungry as well. Nvidia’s Blackwell (GB200) chip, although significantly more power efficient, is designed for nearly seven times the power draw of the A100 chips used to train GPT-3.8. By 2030, US data centres could house millions of such advanced GPUs, while requiring significant energy for cooling. As such, electricity demand from data centres is set to more than double within five years, potentially consuming as much power by 2030 as present-day Japan.

Read more: Data centre infrastructure outlook: AI is crunching on power demand!

Figure 1: Data centres expected to contribute around 12% of US total electricity consumption

b)     Trump’s MAGA: Manufacturing Reshoring

Part of President Donald Trump’s “Make America Great Again” (MAGA) agenda is also to encourage reshoring of manufacturing activities to the US. 

To advance his “Made in America” agenda, Trump has leveraged tariff relief as a bargaining tool to attract foreign investment into the US, striking trade deals with countries such as Japan, South Korea, and the EU, while also warning of a 100% tariff on semiconductors unless companies commit to investing in the US.

All these will amplify the need for electricity. Producing semiconductors requires significant energy, single sites can draw 100–300 MW. For instance, Taiwan Semiconductor Manufacturing Company (TSMC) uses 8% of Taiwan’s electricity to run its chip fabrication facilities in the country, which could jump to 24% by 2030. As the United States reshores chip production, starting with high-value chips, energy demand could rapidly increase.

Just the first phase of TSMC’s fab facility in Phoenix, Arizona, requires 200 megawatts (MW) of peak connected load, enough to power roughly 30,000 homes. By 2030, that power need could grow nearly six times. Seventy-five semiconductor facilities are planned or are currently under construction in the United States, and by 2030 the country could produce 20% of the world's most advanced chips, further amplifying power needs coming from manufacturing.

Figure 2: Manufacturing activities are on the rise

Figure 3: Renewable energy are contributing more to total energy consumption

c)     Broader electrification

After nearly 14 years of stagnation, electricity demand is now firmly rising. It increased by 3%, marking the fifth-highest rise this century, partly as a rebound from a milder summer in 2023, when electricity demand fell by 1.3%.

The broader electrification of the US economy is driving a substantial and sustained rise in electricity demand, marking a sharp shift from the flat consumption patterns of past decades. Several structural forces are converging to reshape the demand landscape: EV adoption, charging networks, and building electrification are adding structural baseload demand.

At the same time, policies promoting electrification of buildings—through electric heating, cooling, and appliances—are gradually displacing direct fossil fuel use. Together, these trends are fuelling one of the strongest electricity demand growth cycles in recent US history, with profound implications for utilities, grid infrastructure, and energy policy.

Figure 4: US electricity consumption expected to increase

Figure 5: Electrification are driving more electricity demand

Supply Constraint

a)     Grid Stagnation

Although the backlog of projects waiting to connect to the grid is shrinking in some regions, interconnection wait times are still lengthy, averaging about four years from queue entry to proposed operation. This indicates that the pace of grid upgrades is failing to keep up with the rapid growth of AI-driven demand.

These loads are concentrated in regions such as PJM and ERCOT and operate continuously throughout the year, making them impossible to manage through temporary demand reduction during peak periods. Traditional methods of grid expansion—such as cyclical capacity additions and power flow adjustments—are no longer sufficient to match the speed of load growth.

b)     Retiring baseload & intermittency issue

Compounding this issue, a large number of coal and natural gas units are set to retire in the coming years, while most new capacity additions are variable renewable sources like wind and solar. Stable baseload capacity additions are projected at just 22 GW, far below the 104 GW expected to retire, pointing to a sharp deterioration in overall grid stability.

According to Table 1, supply margins are expected to become increasingly tight in the coming years (note: a positive number indicates surplus supply). If AI growth exceeds current projections, supply constraints may intensify.

Grid operators are raising warnings. PJM Interconnection, the largest US power grid operator, has cautioned of potential shortages as AI companies request gigawatt-scale allocations, outpacing the grid’s planning capacity. PJM has warned that delays in new transmission infrastructure and generator connections could lead to power shortages unless grid upgrades are accelerated.

In addition, Google has recently agreed to reduce power consumption at its US AI data centres during peak demand periods. The initiative is designed to relieve stress on the already strained US power grid. The agreements include load-shifting strategies that help preserve grid reliability. While this may release some stress of the US grid during peak periods, this may cause AI training or inference activities to slowdown during these periods.

Overall, as electricity supply currently runs at just enough capacity, but with little room for forecast error, there is still risk of electricity supply being less than demand, which could hinder AI development.

Table 2: Supply margins are expected to be very tight

Investment Opportunities

While supply constraints could impede AI development, structural growth drivers are making some of the utilities and energy companies more attractive, especially those who can supply continuous electricity with less disruption.

Read more: Data centre infrastructure outlook: AI is crunching on power demand!

In the near term, some immediate solutions could help buy time for longer-term measures to mature. One such option is Combined Cycle Gas Turbines (CCGTs), which generate electricity using both a gas turbine and a steam turbine, achieving efficiencies of around 60% compared with about 35% for traditional gas plants. Since coal generation peaked in 2007, U.S. electricity generation has continued to rise, largely because clean sources – predominantly solar and wind – have absorbed a significant share of the growth in demand. Unlike wind and solar, however, CCGTs provide stable, dispatchable electricity – a quality that data centres and AI workloads require.

Moreover, CCGTs may face shorter interconnection lead times than many renewable projects, as grid operators often prioritise them in regions where stability is a concern. Demand for CCGTs has been reflected in costs, which are reported to have risen by around 300% compared with a decade ago. For example, GE Vernova has disclosed an order backlog equivalent to 3.3 times its revenue, with the earliest delivery dates for newly ordered CCGT units now pushed back to 2028–2029.

Another solution is nuclear energy. Nuclear plants run 24/7 with very high-capacity factors (~90%), unlike renewables which are intermittent. Plants can operate for 60+ years with extensions, providing stable supply over decades. Companies are pouring into nuclear power as well, with Meta having signed a 20-year agreement to buy roughly 1.1 gigawatts of energy from Constellation’s Clinton Clean Energy Center in Illinois, which is the entire output from the site’s one nuclear reactor.

However, the limitation for nuclear energy is that traditional nuclear plants can take 10–15 years from planning to operation, at a significant cost. This is too slow for near-term electricity shortages caused by AI-driven demand.

This lead to the Small Modular Reactors (SMRs). Like CCGTs, SMRs provide firm, dispatchable electricity that can run 24/7, which AI data centres and industrial loads require. SMRs can be built in smaller units and added incrementally, making them easier to site near demand centres compared with large traditional nuclear plants. In recent years, demand for SMR has been on the rise as electricity supply has become a major problem. The U.S. government wants to triple nuclear power by 2050 to shore up an electric grid that is under growing pressure from surging power demand. But large nuclear projects, in the U.S. at least, are notoriously plagued by multi-billion dollar budgets (10-15 billion vs SMR 2-4 billion), cost overruns, delayed construction timelines and, sometimes, cancellations. Companies like GE Vernova see demand for as many as 57 small reactors in total across its target markets in the U.S., Canada, the United Kingdom and Europe by 2035, and the first of four reactors is already under construction near Toronto. Meanwhile, big tech firms like Google recently pledged to fund the development of three new nuclear sites, after  teaming up with Kairos Power, a developer of small modular reactors (SMRs) last year. Amazon invested more than USD 500 million to develop SMRs in October and bought a data centre campus powered by the Susquehanna nuclear plant in March 2024.

Figure 6: GE Vernova’s expected demand for SMR

Source: GE Vernova, Data as of August 2025.

Conclusion: Power constraints present a hurdle to AI; but an opportunity for electricity companies

The rapid expansion of data centres, manufacturing reshoring, electric vehicle (EV) adoption, and broader electrification are driving substantial growth in U.S. electricity demand. While this rising demand risks making energy the primary bottleneck for AI development, it has also contributed to sizeable order backlogs for utilities and power equipment manufacturers. Combined with policy support — including measures to streamline licensing and accelerate project deployment — this presents structural opportunities for the sector.

However, we believe much of this optimism is already reflected in current share prices, with companies such as GE Vernova and Siemens Energy trading at historical highs and PE multiples above their long-term averages. As such, we maintain a Neutral stance on the sector. Investors seeking exposure to this structural growth theme may consider waiting for a pullback, until valuations become more reasonable. Alternatively, investors who seek passive exposure to the utilities sector can consider the iShares Global Utilities ETF (NYSE: JXI) that invest in major global utilities companies, which include power generators and utilities companies.