The world's AI infrastructure faces a critical bottleneck: most existing data centers lack the electrical capacity to run the latest generation of graphics processing units (GPUs), forcing companies to choose between expensive new construction or settling for older, slower systems. More than 80% of current AI capacity sits in facilities built for far less demanding workloads, yet fewer than 8% of enterprise data centers will meet the power and cooling requirements when next-generation systems arrive.

What's Driving the Data Center Power Crisis?

The problem stems from a dramatic acceleration in GPU power consumption. Older systems like the DGX H100 and H200, each with eight GPUs, consumed only 10.2 kilowatts of power, making them compatible with most existing facilities. But the latest Blackwell DGX GB200 NVL72 draws up to 120 kilowatts per rack, a more than tenfold increase. Even more demanding systems are on the horizon: the Vera Rubin NVL72 operates at 120 to 130 kilowatts, while the Rubin Ultra NVL576 skyrockets to as much as 600 kilowatts per rack, enough to power roughly 400 homes. The Feynman system expected to follow may require as much as 1 megawatt per rack.

Meanwhile, most current data centers can only support 30 kilowatts or less per air-cooled rack. This mismatch has created what industry analysts call the "AI datacenter power crisis," a growing gap between what next-generation AI systems demand and what existing facilities can actually deliver.

How Much Will New Data Centers Cost to Build?

Building new facilities to accommodate these power-hungry systems carries staggering costs. Foxconn estimates that a 1-gigawatt data center built to power NVIDIA VR200 NVL72 systems would cost approximately $47 billion to construct, with an annual electric bill of roughly $1.3 billion. NVIDIA CEO Jensen Huang has confirmed that building a data center from breaking ground to operational AI supercomputer takes about three years in the United States.

The timeline and expense are already stalling projects. One report estimates that the inability to secure adequate power will delay or cancel 30% to 50% of AI data centers planned for deployment in 2026. Of the 16 gigawatts of capacity slated to come online in 2026, only 5 gigawatts are actively under construction. With typical build times of 12 to 18 months, the remaining sites are unlikely to become operational within their target dates.

Why Are Power Utilities Struggling to Keep Up?

The infrastructure challenge extends beyond data center construction itself. Data centers that demand enormous amounts of power can be built far faster than power utilities can expand their capacity. Delivering increased power to data center locations often requires new transmission lines from existing generation facilities, a process that can take years for permits alone. In some cases, entirely new power plants must be built, a process that can take decades.

This gap is particularly acute in rural areas, where most new data centers are being built. While 87% of existing data centers are located in urban areas, 67% of those currently under construction are in rural areas, and 39% are in areas with no data centers today. Many rural sites lack grid connections capable of meeting the power demands of modern AI infrastructure.

"Datacenters that demand huge amounts of power can be built far faster than power utilities can expand their capacity. Delivering increased power to datacenter locations often requires new transmission lines from existing generation facilities, which can take years for permits, but can even necessitate building new power plants, which can take decades," according to Gartner research cited in the source material.

Gartner, AI Infrastructure Guide for Power-Constrained Data Centers

What Are the Broader Implications for AI Deployment?

The power crisis carries significant consequences for artificial intelligence (AI) development globally. Gartner estimates that meeting the incremental power needs of AI data centers in 2027 will require 500 terawatt-hours per year, a 2.6 times increase from 2023 power requirements and nearly as much as Germany's entire power consumption in 2022. This explosive demand compounds existing infrastructure challenges.

The crisis also threatens "sovereign AI" initiatives, national AI deployments built to benefit particular countries where data is stored, processed, and managed within national borders. While AI can offer clear national benefits, many nations are constrained by existing data center capacity and the availability of power to operate new facilities. Aging power grid infrastructure compounds the problem; Capgemini reports that outdated grids are a top concern for reliable data center power delivery, cited by 74% of electricity executives globally.

How Can Organizations Deploy AI Within Existing Data Centers?

• Adopt Energy-Efficient Architectures: Organizations can deploy AI systems specifically designed to operate within existing power constraints, such as platforms that run air-cooled at 10 to 20 kilowatts per rack, allowing them to drop into existing data centers without requiring new construction or liquid cooling infrastructure.
• Implement Design and Operational Best Practices: According to Gartner, organizations must adopt design and operational practices to deploy AI infrastructure effectively within existing spaces, ensuring performance, scalability, and energy efficiency without exceeding current facility limitations.
• Combine Multiple Approaches: Rather than relying solely on new construction, companies can run energy-efficient systems alongside existing GPU infrastructure, maximizing utilization of current facilities while gradually transitioning to next-generation systems as power infrastructure improves.

The AI data center power crisis represents one of the most pressing infrastructure challenges facing the technology industry. While new construction will eventually expand capacity, the immediate solution lies in deploying systems that work within existing constraints, buying time for power utilities to catch up with demand. Organizations that can operate advanced AI workloads on current infrastructure will gain a significant competitive advantage as the industry navigates this critical transition period.