The first tranche of new US data centers coming on line this decade to run electricity-intensive artificial intelligence (AI) software will probably rely mostly on power generated by natural gas, while the nuclear renaissance hoped for by Big Tech comes later in the 2030s.
Microsoft, Amazon, Facebook-parent Meta and Google-parent Alphabet want clean, reliable power as quickly as possible so they can be early movers in the development of AI, which is rapidly advancing and finding new user bases around the world. While these companies do not relish the optics of powering AI development with fossil fuels, gas-fired power is widely expected to fulfill most of the gap between current supply and future demand through at least 2030.
Unlike wind and solar, gas can be relied upon for steady, baseload power, a necessary ingredient for always-on data centers. And crucially, unlike nuclear, gas-related infrastructure can be built out quickly. The most recent additions to the US nuclear fleet, Vogtle units 3 and 4 in Georgia, took 15 years to build and cost $30bn, double the expected time and cost. A few decommissioned nuclear reactors can be restarted, as Microsoft is paying to do with a unit of Three Mile Island in Pennsylvania. But this low-hanging fruit will be quickly exhausted.
Questions around the meter
While there is broad agreement that gas will power the AI data center boom through at least 2030, questions remain about what this rapid gas-fired power build-out will look like.
Data center operators can secure power in two ways: wade through the long, arduous interconnection process through which new customers connect to the grid, or bypass the grid altogether and secure their own personal electricity supply through so-called "behind-the-meter" agreements. Many in the gas industry are betting tech companies' need for speed will force them to opt for the latter.
"The data centers are not going to wait," Alan Armstrong, chief executive of Williams, the largest US gas pipeline company, told Argus in an interview. "They are going to go to states that allow you to go behind the meter."
In this scenario, construction of an AI data center in a state like Louisiana, for instance, might accompany construction of a new intrastate pipeline connecting the state's prolific Haynesville gas field with a new gas-fired power plant. Intrastate pipelines bypass the federal oversight triggered by interstate pipeline construction, and new gas power plants only take 2-3 years to build, East Daley Analytics analyst Zachary Krause told Argus. Most of the incremental power needed to run AI data centers this decade will be generated by new gas plants, Krause said.
Even ExxonMobil in December said it was in talks to provide "fully islanded" gas-fired power to AI data centers. It claimed it could even capture 90pc of the CO2 emissions from power generation, appeasing tech companies' climate ambitions.
ExxonMobil's non-grid gas generation fleet is "independent of utility timelines, so they can be installed at a pace that other alternatives — including US nuclear — just can't match," ExxonMobil chief financial officer Kathy Mikells said.
But connecting to the grid may offer better reliability and economics than behind-the-meter gas power. If an off-grid gas generator trips off line, for instance, an always-on data center without back-up generation depending on that facility would be in trouble. Grid connection also allows generators to sell excess power into the grid. For those reasons, most new data centers this decade will rely on the grid as their primary power source, Adam Robinson, research associate at consultancy Enverus, told Argus.
Small modular future
But if the 2020s become the decade of gas-powered AI, the 2030s may be when nuclear-powered AI gets its due.
The long-awaited nuclear renaissance may come not from conventional reactors, but from next-generation small modular reactors (SMRs), which can theoretically be built much faster and cheaper. No US SMRs yet exist, but given the number of SMR start-ups with expected start dates before 2030, and money pouring into the sector from the likes of Google and Microsoft, at least one of these next-generation reactors should be operating by 2030, Adam Stein, director of nuclear energy innovation at research center Breakthrough Institute, told Argus. SMRs' smaller price tag relative to conventional 1 GW nuclear reactors may also accelerate their adoption, Stein said.
"Not every utility needs a GW-scale plant of any kind, but they might need a 300 or 600MW plant," he said. "So the total addressable market is larger for SMRs."
