By Andrew Wetzlel, CFA®, Managing Director of Sustainable Investing
We are living through a major transformation in how computing works, with the shift from serial processing to parallel processing and from CPUs (central processing units) to GPUs (graphics processing units), both types of microprocessors. GPU chips, which can run trillions of operations per second, have accelerated the development of artificial intelligence (AI), and in particular generative AI (“GenAI”), a type of AI that can create many types of content including text, images, video, and even computer code. The models that make GenAI possible are created and trained at data centers, many owned and operated by the largest technology companies on the planet. The growth in AI-capable data centers, which require significantly more energy than traditional cloud computing, is driving power demand and raising questions about the sustainability of this new era of computing.
Research from Alex de Vries shows that using AI to generate search results at Google could use between seven-to-ten times the energy consumption of a traditional search. Further, Empirical Research Partners’ work shows a strong correlation between the number of parameters in an AI model and the energy required to do the training. While newer GPUs are more energy efficient, the shift toward this type of computing along with growth in model size and complexity are contributing to electricity demand growth in the US.
Data from the Energy Information Administration (EIA) show that US electricity demand has been roughly flat at around 4,000 billion kilowatt-hours (kWh) or 4,000 terawatt-hours for over 15 years as efficiency has offset growth in the economy. With the recent trend toward electrification, the EIA estimates that demand will grow 1.5% annually, with nearly a third of this from data center growth. Private forecasters, such as Goldman Sachs, forecast demand growth of 2.4% through 2030, while others see growth above 3% annually. Should the Goldman forecast prove accurate, data center power demand in 2030 would represent 8% of total US power demand, well above today’s 3%. Note that the average American home uses roughly 10,800 kWh (10.8 megawatt hours) of electricity per year.
It’s important to note that electric generating capacity has been steadily increasing over the last 15 years, which may seem at odds with the historically flat demand. The reason for this is an increasing mix of more intermittent forms of power generation, primarily wind and solar, which have lower capacity factors (run less often) relative to nuclear, coal, and natural gas-fired generation. A 1-megawatt (MW) solar array can generate about 2,100 megawatt hours (MWh) of output in one year based on a national average capacity factor. A 1MW gas plant can generate roughly 5,000 MWh of output based on an average capacity factor. Nuclear and coal tend to have higher capacity factors than gas plants. The trend toward renewables and away from coal is well established, though the time it takes to connect this new capacity is increasingly a roadblock, creating a challenging setup for the largest tech companies seeking to win the GenAI arms race.
A recent Department of Energy report noted that the typical duration between connection request and operation date for new power projects is about 5 years, with 1,480 gigawatts (GW – 1 GW equals 1,000 MW) of renewable capacity and 1,030 GW of storage capacity currently seeking access. This is a growing problem for the large cloud providers (also known as hyperscalers) and other AI developers. Recent research from Morgan Stanley quantifies the value of “time-to-power,” the percent power price premium willingly paid to bring an AI data center on line 2 years faster. The research points to premiums of 60% to 100% based on GPU useful lives of 10 to 6 years respectively, relative to an indicative all-in utility rate of $100/MWh.
In the near-term, the growth in GenAI-related power demand will clearly outstrip renewable energy capacity additions to the grid, setting up fertile ground for innovative deals to secure power. Amazon recently agreed to buy a data center campus from Talen Energy, operator of an adjacent nuclear plant, while Microsoft announced a partnership with nuclear operator Constellation Energy for regional hourly matching of clean energy. In May, Intel deepened a partnership with Bloom Energy for fuel cell power at a data center expansion project. While all positive steps, the nuclear deals capture clean power that could be used for other drivers of demand growth and the scale is simply not sufficient to meet net-zero targets, at least in the near term.
One clear example of this is Microsoft, a company that has set aggressive sustainability targets including being carbon negative, water positive, and zero waste by 2030. While Microsoft has made significant progress, they’ve hit a bump in the road. The 2024 Environmental Sustainability Report posted recently highlighted a 31% increase in scope 3 (value chain) CO2 emissions, driven by construction of data centers. This same metric was flat in the 2022 report, while scope 1 & 2 (direct and indirect) emissions continue to decline. The 29% increase in Microsoft’s total CO2 emissions (scopes 1-3) will likely continue as data center construction picks up pace. To address this, over the last two years, MSFT has signed 6.3 GW of renewable energy power purchase agreements (PPAs) and added an additional 2.4 GW through June of this year. Most recently, Microsoft announced a framework agreement with Brookfield, one of the largest renewable energy developers, to deliver 10.5 GW of renewable energy in the US and Europe between 2026 and 2030.
Over time, data center construction and AI model training will turn to running this infrastructure to drive GenAI-related revenue from commercial applications. As renewable energy and storage assets, backed by PPAs, are developed, along with other innovative power solutions, Microsoft and GenAI leaders are likely to hit sustainability goals. Eventually, GenAI may well accelerate efforts to tackle extremely complex challenges such as climate change and degradation of natural capital, efforts Microsoft and others are already taking on.
As we evaluate the broad landscape of companies supporting the development of AI infrastructure, it’s clear that the combination of growing power demand and corporate climate commitments will drive accelerating renewable energy development. The inherent intermittency of renewables combined with grid connection and transmission/distribution challenges will drive local development and a focus on energy storage. Companies able to execute on this development at scale are likely to prosper along with the better-known technology companies, creating potential investment opportunities for sustainable investors who have long-term horizons.