超大规模AI扩张背后的数据中心电力挑战

Data Center for AI Computing in Rural Georgia, illuminated at dusk

With utility timelines stretching into the next decade, many large AI data center developers are turning to natural gas as the fastest route to new dispatchable power.Getty Images

The power grid is straining under the ambitions of hyperscale AI data center developers. Microsoft, Meta, Amazon, and Google have increased their data center investment roughly tenfold over the past eight years, according to Industrial Info Resources (IIR). Up from a combined $64 billion in 2018, their announced spending is expected to exceed $700 billion this year.

“Twelve straight months of new project announcements with spending running consistently elevated indicates a structural, multiyear theme of digital infrastructure buildout that is set to continue for the foreseeable future,” said Shane Mullins, an analyst with IIR at the Data Center World Power event in September 2026.

Community Opposition Grows, but Project Momentum Holds

The rapid expansion of data center development, fueled in large part by AI demand, has whipped up a growing wave of community resistance. According to IIR, more than 500 data center moratoria are on record across 42 states. New Jersey and Michigan lead with 41 and 36, respectively. New York has enacted a statewide ban on new data centers larger than 50 MW, at least temporarily, while Texas has paused new grid interconnections as it works through about 200 GW of requests.

“$130 billion in projects were rejected in Q1 of 2026 alone, more than the total rejected for the whole of 2025,” Mullins said. “The number of opposition groups has more than doubled over the past year.”

Even so, resistance has slowed only a fraction of the broader AI infrastructure buildout. Capital continues to pour into the sector. IIR data show at least $400 billion per month in newly announced data center projects globally since April 2026, with May exceeding $650 billion. That has pushed cumulative capex beyond $7 trillion.

In the US and Canada alone, that pipeline amounts to roughly $4.75 trillion and more than 4,250 planned or in-progress data center projects. Texas accounts for nearly $1.25 trillion of that total. Altogether, more than 400 GW of new data center load is in planning, and almost 100 US sites are being designed for at least 1 GW.

Where Will the Power Come From?

The scale of the challenge is already visible. The US now has 53 GW of operational data center load, up from 23 GW in 2023. Another 42 GW has been approved or is under construction. That could push data center electricity demand from 35 GW in 2024 to 125 GW by 2030.

Grid planners have struggled to keep up. In 2022, US planners projected that the country would need 24 GW of additional electricity demand within a decade. By 2025, that forecast has climbed nearly sevenfold, to 166 GW.

As a result, hyperscalers and other developers are increasingly being pushed to build, own, or co-develop power generation. Gas turbines remain the preferred option for projects that need to scale beyond 1 GW. But turbine manufacturers are facing capacity constraints of their own. GE Vernova, Mitsubishi Power, and Siemens Energy are sold out of larger turbines through 2029, with some orders now stretching into 2031. All three are expanding production, while other suppliers are offering fuel cells, reciprocating gas engines, and jet engines as faster alternatives.

“If you need more than 300 MW and the utility queue says 2034, building or contracting for your own gas plant is the fastest and most certain path to power,” Mullins said.

Natural Gas Reemerges in the AI Power Mix

AI data center growth has reshaped the North American power market. In 2022, wind and especially solar were ascendant, and some analysts predicted the end of gas turbine sales in North America. Solar project announcements continue to rise, but so does interest in natural gas generation and battery energy storage systems (BESS). IIR estimates that $228 billion in new natural gas power plant projects are currently in play.

Hyperscalers and AI data center developers are, in practice, largely power-source agnostic. Most organizations want clean energy, but the need for reliable power now is taking precedence. That has contributed to renewed investment in coal plants scheduled for retirement, as operators seek to keep them running for a few more years. Older natural gas plants are also being upgraded to increase output.

“Natural gas is filling the AI data center demand gap, with 347 grassroots, expansion, and conversion projects,” said IIR analyst Britt Burt. “Gas is the fastest, most reliable path to new dispatchable capacity.”

According to IIR, the North American power industry could attract more than $1.5 trillion in investment across nearly 7,000 planned projects. But the story is not only about gas. Nuclear energy now accounts for 111 projects worth $311.4 billion, including 61 small modular reactors (SMRs) as well as several nuclear plant restarts.

Meanwhile, renewables continue to expand. One of the fastest-growing combinations is solar paired with BESS, which extends solar’s usefulness beyond sunset and improves dispatchability. IIR counts 701 BESS projects worth almost $200 billion. Wind development has slowed, and offshore wind may continue to fade in the US, but onshore wind remains active.

Speed to Power Is Reshaping Energy Strategy

The same set of constraints is shaping every energy source: interconnection delays, labor and equipment shortages, grid congestion, limited transmission capacity, and increasing community opposition. In that environment, data center developers are pursuing whatever path offers the fastest, most reliable route to power at the required scale.

“AI and data center buildout, reshored manufacturing, and an aging generation fleet are driving electricity demand growth not seen in a generation,” Burt said. “This is one of the broadest capital cycles the US and Canadian power sectors have seen in decades.”

About the Author

Contributor

Drew Robb has been a full-time professional writer and editor for more than twenty years. He currently works freelance for a number of IT publications.