Is GenAI villain and hero in data center power drama?

Power demand shifts and accelerates

In the early days of GenAI, power demands tilted toward the compute-intensive task of training models. Energy consumption now stems from other sources.

"It used to be these large language models … took a lot of training and that training was highly energy consumptive," said Autumn Stanish, director analyst at Gartner. "That has now largely shifted to inferencing."

With inferencing, a GenAI model draws upon its training phase to interpret new data when users submit queries. This action generates output, such as text, images and code. Inferencing puts a strain not only on compute, but also on networking, memory and cooling systems. "It's not just the servers with the GPUs and various accelerators," Stanish noted. Liquid cooling is close to becoming an absolute requirement for high-performance computing, she said, adding that hyperscalers have purchased so much memory capacity that they've created a shortage.

Dell'Oro Group has also cited the rising demand for infrastructure. The market researcher earlier this month projected that the worldwide data center physical infrastructure market will exceed $80 billion by 2030, with an annual growth rate in the mid-teens. Within that market, direct liquid cooling is expected to eclipse $8 billion by 2030 as it becomes "a foundational technology for AI factories," Dell'Oro noted.

Gartner, meanwhile, forecast that infrastructure will drive worldwide AI spending in 2026, which the research firm expects to surpass $2.5 trillion. That's a 44% year-over-year increase, for which AI infrastructure is expected to add $401 billion in spending compared with 2025 levels. Gartner didn't include cooling and power infrastructure cost in its AI forecast, but John-David Lovelock, a vice president analyst at the research and advisory firm, said cooling and power infrastructure needs "are slowing the construction of new data centers and adding significantly to the final cost."

Ways to ease power constraints

Dell'Oro pointed to "power scarcity" as a constraint on near-term data center expansion, citing on-site power generation as a workaround for large AI campuses. On-site generation is emerging as a popular "bridge solution" that can power large data centers while they wait to be interconnected into the grid through a substation, said Alex Cordovil, research director at Dell'Oro Group. The process of bringing a site online at gigawatt scale involves "very long timelines," he explained. "We expect [on-site power generation] to be a very important strategy. We're looking at it becoming almost a mandate for new deployment."

Cordovil cited Microsoft's Fairwater AI data center in Mount Pleasant, Wis., xAI's Colossus data center in Memphis, Tenn., and Crusoe's Abilene data center in Abilene, Texas, as examples of data center projects that involve on-site power generation. It's not an approach for every data center, however. On-site power generation only becomes cost-effective for large data center campuses around the 100-megawatt threshold and above, Cordovil said. That typically includes hyperscalers and colocation facilities.

Data centers are turning to natural gas turbines and reciprocating engines as their primary on-site generation options, Cordovil added. That's an appropriate choice in states like Texas and other regions with abundant natural gas and a well-developed pipeline network. But other generation technologies, such as fuel cells, could gain importance as they become more economical. Fuel cells are more expensive than gas turbines, Cordovil said, but the gap is closing.

Gartner views microgrids as an approach large data centers might take to address power constraints. It describes microgrids as independent power networks that run on their own or in conjunction with the main power grid to meet a data center's electricity needs. Most businesses are weighing whether to create a microgrid based on nuclear power, geothermal power or some other energy source, which, Gartner's Stanish said, could lead to widespread microgrid deployment in the 2027-2028 timeframe.

Social considerations could delay that schedule, however. Data center builders might have to address questions such as whether they have an obligation to invest in the surrounding community's power infrastructure as well as their own private grids, Stanish said.

GenAI to the rescue?

GenAI as an energy manager, as well as a top contributor to power demands, could play a greater role among large data centers.

Today, GenAI is already embedded in data center infrastructure management tools, Stanish noted. GenAI and AI-driven automation in general are moving through a series of maturation steps. AI has been used for years to monitor data center environments, and more sophisticated uses such as recommendations and pop-up alerting are becoming established, Stanish added. These features give data center managers a heads up based on preset triggers and offer power-management guidance.

The next step would involve an AI agent, or another automation mechanism, managing energy use autonomously. Stanish said this phase will be important for power-savings measures, such as spatial and temporal shifting -- for example, moving an AI workload to a data center location with a surplus of green energy or scheduling a workload to run at the most energy-efficient time.

The top 5% or so of businesses, such as those with experience in technologies like digital twins, will be the first to deploy this style of AI, perhaps by late 2027 or early 2028, Stanish conjectured. Other organizations will follow in the early 2030s.

Along those lines, Quinn said Caylent is talking with an energy producer about using AI to "surface where flexible AI workloads can actually support grid stability instead of undermining it."

Cordovil also believes  AI can play a role in managing energy use. He expects data centers to use AI in maintenance and facilities management. Another plus is AI's ability to provide real-time information for energy decision-making. It could provide insight into when to turn data center chillers on or off, Cordovil said. And AI could help data centers determine when to sell excess capacity to the grid -- on hot days, for instance -- generating revenue through price arbitrage.

"We expect these technologies to be used more and more within the data center to make them more efficient," Cordovil added, "and make the overall grid more efficient and more reliable."

John Moore is a freelance writer who has covered business and technology topics for 40 years. He focuses on enterprise IT strategy, AI adoption, data management and partner ecosystems.