北美 AI 数据中心迈入 200 GW 节点，电网与高密度部署如何重构

High-Voltage Power Lines at Sunset with Digital Energy Flow

Utilities are now fielding gigawatt-scale data center requests, but power assets move on decades-long timelines, NRG Business EVP Scott Hart noted.Getty Images

Global data center power demand is projected to exceed 200 gigawatts (GW) by 2030. AI workloads are driving typical rack densities from roughly 10-30 kW into 80-150 kW, with advanced deployments approaching 1 MW per rack. Individual campuses are scaling into the hundreds of megawatts and beyond. The pace of growth is unprecedented.

At the Data Center World Power conference in Dallas, held Sept. 21-23, leaders from utilities, power producers, hyperscalers, data center developers, operators, and supply chain partners discussed how to scale capacity while safeguarding grid reliability, protecting ratepayers, and accelerating economic development.

Bill Kleyman, executive chair of the Data Center World Power Advisory Board and CEO and co-founder of Apolo.us, opened an early session by laying out the realities of deploying Nvidia DGX-class servers. He noted that a standard rack can accommodate six such systems, each built around eight Nvidia H100 accelerators, with individual accelerator cards costing up to $40,000.

“Sixty-nine percent of respondents to AFCOM’s State of the Data Center 2026 survey expect that rack density is going to increase like nothing we've seen in the past,” Kleyman said. “In parallel, we are seeing 36% growth in liquid cooling adoption in data centers, and that's continuing to expand.”

Utilities Confront the Gigawatt-Scale Campus

Scott Hart, executive vice president of NRG Business, a major provider of retail electricity, natural gas, and demand response in North America, described how customer requirements have escalated. “Traditionally, data centers would be a grid-connected, 20- to 40-MW facility,” he said. “Then we saw the cloud campuses in the 100 MW range, but now conversations have pivoted to a gigawatt and beyond.”

With projects of that scale, it can be challenging for a utility to commit to constructing a power plant. Hart explained that there are two clocks running during data center development: a technology clock that operates on timelines of months or a few years, and a power infrastructure clock that operates on decades. That creates friction when customers want capacity on near-term timelines, but utilities must plan and finance assets over 15-30 years. “A utility needs a minimum 15-year time horizon and commitment with a secured [offtake agreement] that can pay for that entire period,” Hart said.

Some sites will remain grid-only. Increasingly, however, projects are pursuing hybrid arrangements that combine grid supply with on-site generation and storage, Hart noted. That can involve mixing technologies such as geothermal, gas turbine plants, and renewables.

Hart added that incorporating demand response – programs in which customers curtail load or shift to on-site resources during grid stress in exchange for incentives – is becoming more attractive. “Most of the data center projects we work on are excited about being part of a demand response program,” he said.

Site Selection: Power and Community

Alise Porto, vice president of sustainability and strategic initiatives at Switch, said the site selection calculus has changed. A few years ago, proximity to major metro areas and robust fiber connectivity often dominated decisions. Today, available and reliable power is paramount.

“Everything now revolves around power,” Porto said. “And more recently, the scope has expanded to both power and community – it is those two together that now coordinate where you go, how big you go, and what you do.”

“Everything now revolves around power.” – Alise Porto, vice president of sustainability and strategic initiatives, Switch

Utilities, initially surprised by multi-hundred-megawatt and gigawatt-scale requests, are increasingly engaging earlier and collaborating on flexible solutions, Porto noted. That can include agreeing to reduce power draw during peak conditions or temporarily switching to batteries or backup generators to support local grid reliability when needed.

To maintain optionality as chip roadmaps evolve, Switch is pursuing designs that accommodate both current and future cooling methods and densities. “Depending on the type of chips our customers have, they can deploy air cooling now and utilize liquid cooling later,” Porto said. She added that Switch’s “AI factory” design targets up to 2 MW per rack, with modularity to allow cabinet-level power draw to scale as denser server architectures enter the market.

Developer and Operator Economics

Dado Slezak, executive vice president of energy capital and strategy at QTS, emphasized the capital intensity of AI data centers, particularly when on-site power is part of the plan. “For every dollar you've got of generation, you've got multiple dollars of data center infrastructure,” he said. “With that scale of investment, there must be a decades-long financial return secured contractually.”

Slezak also cautioned against rigid long-term prescriptions. “There is no point when you can look three or four years in advance and know what you should build,” he said. “A lot can change in that time, and decisions keep shifting, each with a downstream impact.” He urged deep collaboration among partners to manage evolving technology targets and interconnection realities.

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.