Major tech companies like Amazon and Microsoft are, quite literally, power-hungry. The staggering amount of electricity required to fuel artificial intelligence, and the data centers behind them, has driven these companies to such extreme measures that they are attempting to bring retired nuclear power plants back to life.
There’s no question that data centers use a lot of electricity. Before AI-related data centers came onto the scene, electricity demand in the US grew very little for a decade or more. Now, Virginia has more data centers than anywhere, and in 2023 those data centers used over 25% of electricity sold in the state.
This sudden spike in electricity demand—on track to far outpace the growth in supply—raises unique challenges. The data center industry believes they are in a race to build computer capacity for artificial intelligence (AI), and meeting the data centers’ electricity needs is key to their growth. Getting enough electricity is limiting the development of data centers. And so, desperate times call for desperate measures, like Microsoft co-locating data centers with nuclear plants, or Meta commissioning the construction of dirty fossil fuel plants to power data centers. To understand conversations about co-location, here’s some basics that you need to know.
Why do data centers need so much energy?
There are two reasons why data centers are quickly growing as part of our electricity demand. One is simply more things we do require access to shared data, such as shopping online, streaming videos, and business communications. But growth from artificial intelligence data processing is the big driver for more data centers and more energy demand at each data center. AI requests require more data analysis than typical internet searches or even cryptocurrency mining, which in turn requires more energy—take a look at my colleague’s recent blog post on the topic to understand just how much we’re talking about. The next step up in energy-hungry activity is generative AI.
Electricity demand is so key that a data center’s size is routinely described by how much electric energy it requires. A 50 megawatt data center has the energy demand of a small city. But new data centers are now proposed to be 500 megawatt, grouped on campuses with electric demand greater than nuclear plants.
Most data center energy use is consumed by data processing computer chips, but this work also produces a lot of heat; the processors can run as hot as 175°F, and the cooling equipment needed to keep them from overheating accounts for roughly 20% of the total electric energy consumed by the facility. Most data centers use conventional air conditioning for this purpose, but designs are available that keep the processing equipment submerged in mineral oil or similar liquid that is inherently more efficient than using air to absorb and move heat.
How do data centers get the electricity they use?
Today, thousands of data centers are attached to the electric power grid, like any other building or facility. The energy use is more intense than most buildings (often described as ‘energy dense,’ as in kilowatts per rack or cabinet), but until recently, most data centers were not so large as to overload the electric grid. That is no longer true. Commonly used AI processing chips were using 6-10 kilowatts per rack, which is a bit more than what my level 2 electric car charger in my garage can draw. Now, chips using 30–60 kilowatts are widely used. New data centers fill a large building with rows and rows of such cabinets.
All this growth in demand requires building the needed electrical supply for data centers as a central part of the growth of AI. Teaming up with an existing nuclear plant is just one of the dramatic approaches data companies pursue now.
Will data centers choose clean renewables or dirty fossil fuels?
Let’s look more closely at how the race to build AI is now a race to build more electricity supplies. Besides the five companies that own 39 US nuclear plants outside of monopoly utilities, the greatest opportunities for new data centers are in combination with new generation. The usual practice for typical growth of electricity use in the United States doesn’t require a new consumer to do more than call the local utility to bring a new pole and power line to the new business or residence.
Now, with data centers using as much energy as cities, that’s not going to work. The conventional approach to meeting the added demand from data centers requires more transmission lines, which can take 5-9 years to plan and build, as well as completing the construction of new power plants, which also require more transmission lines.
The options for today’s data center developers in a hurry are either to join in building fossil gas power plants or join in building clean energy power plants. The coordinated scheduling and planning of both supply and demand allows clarity of needs and solutions for these enormous investments at a time when confusion and delay are widespread in the electric industry over data centers.
Do data centers have a history of independent power generation?
Data centers, and other vital facilities, commonly have some amount of on-site electricity generation as a back-up for grid outages. The conventional approach in the past was an inexpensive, but inefficient internal combustion generator that was not expected, or permitted, to run much. Today, as developers’ ambitions grow, they have expanded this in hopes of bypassing delays in transmission expansion or the costs of transmission, or both.
The progression has moved from bringing data centers to existing generators to building both new data centers and electric generation where there is an untapped energy resource. This can mean using wind, solar, or fossil gas generators built with the data center. Broadly, the idea is to have redundant generation, such as solar plus batteries plus the grid connection, to improve the reliability above what “going it alone” might provide. For now, there is no standard approach to meeting the physical and legal challenges to on-site power supply. Further complicating matters, the global capacity to build conventional energy infrastructure, including gas-fired generators is already stretched to meet new demand.
There are already examples of using renewable energy attached to data centers to meet corporate and public goals. The most prominent example is Google planning to bring your own new clean energy, “BYONCE” (not to be confused with Beyonce’s song “Energy”).
Can on-site power production reduce costs on consumers imposed by data center growth?
Customers adding on-site electricity production is familiar these days. Millions of Americans have solar on their roofs. An even earlier arrangement, industrial cogeneration of heat and electricity, provides greater efficiency and cost savings for large energy consumers. In those situations, electricity flows to the grid when the on-site supply is greater than the demand, and power flows from the grid to the customer when more is needed than is being produced. These clean and efficient energy additions lower energy costs for everyone, especially when overall supplies are tight. When renewable energy supply goes up, prices go down.
Looking ahead, the role of data centers will be important for the cost of energy in the US. Unless there are many adopters of BYONCE, consumers’ costs will be going up in response to generator shortages and transmission needed for data center growth. Health and environmental costs will also rise if fossil fuels are used directly for the data center, or to replace a nuclear energy supply now dedicated to data center use.
Which regulators can protect consumers from costs from data centers?
Sales of electricity from a power plant to a data center is a retail sale, which are regulated by the states’ public utility commissions as part of the supervision of utility companies. States should require data centers to pay for any additional generation, grid upgrade, and operating costs to protect consumers from potential cost increases. While developers will claim that on-site generation allows them to avoid these costs, that is not the case.
The Federal Energy Regulatory Commission has opened an inquiry into the rules for a data center at the site of a nuclear plant. A decision should clarify options and opportunities for improvements over the current wholesale requirements, which are limited and few. Federal law describes electricity sales as “affected with a public interest” and primarily regulated by states. Some states are starting to set rules at their own pace, leaving a patchwork.
Why do data centers want to locate adjacent to nuclear power?
Before data centers arrived, nuclear power plants supplied regional power grids as some of the largest electric generators built. Nuclear plants usually have a large land area around them as a safety and security buffer that excludes customers’ facilities. Now, a few nuclear plant owners have made deals for data centers to be built adjacent to existing nuclear plants. This is not workable at every US nuclear plant, as half of the 94 reactors are more than 40 years old and only 39 are owned by independent companies that could make new deals to sell their electricity directly to data centers. Also, a recent update on nuclear plant safety from UCS is a reminder that old plants may not be a good solution for the future.
The information technology industry has become interested in nuclear power recently because some tech companies have carbon-reduction goals that they fear cannot be fully achieved without nuclear power. Several tech companies have purchased large amounts of renewable energy, primarily electricity from wind and solar. Some tech leaders have bought into the grand plans of new nuclear power companies that offer unlikely promises of new designs and easy access to the electricity vital for future data centers.
However, new nuclear plants are not likely going to meet the demand from data centers easily, starting with delays to building conventional new nuclear plants, untested new designs, high costs, and the lack of solutions to nuclear power’s problems.
The spread of AI data centers is larger, more numerous, and faster-growing than anything that has come before to the electric utilities. The potential for this to cause supply shortages and costs for consumers is plain to see, and yet also falling through gaps in the existing rules. Dialog between the larger data companies, some utilities and some regulators won’t solve the problems right away. This is going to take some time and effort. Investors, elected officials, and everyday consumers need to keep asking questions.