Artificial intelligence data centers are hitting a power problem that has little to do with computer chips, according to a new report from Thornburg Investment Management.
Key Takeaways
- Average AI server rack density has nearly quadrupled since 2021, straining building power systems.
- Transformer lead times now stretch up to five years amid a broader equipment crunch.
- TAOZ and TFGZ, two active Thornburg ETFs, hold power infrastructure names like Vertiv and Coherent.
Nvidia Corporation’s (NVDA) latest AI hardware draws far more electricity per rack than entire data centers required a decade ago. The transformers, switchgear and cooling systems inside the building have not kept pace, the report found.
At Nvidia’s GTC 2026 conference, chief executive Jensen Huang described AI infrastructure as a five-layer cake. Energy, he said, forms its foundation. “Energy is the first principle of AI infrastructure and the binding constraint on how much intelligence the system can produce,” Huang said.
Thornburg equity research analyst Baadal Chaudhary calls that imbalance “Watts and Wafers.” Chips have scaled at a pace that keeps surprising investors, he wrote. The physical systems that deliver electricity to run them move on timelines measured in years, not quarters.
Transformers take two to five years to procure and switchgear can take up to three years, according to the report. The grid interconnection queue in Northern Virginia, a hub for data center construction, now runs seven years.
See more: Matthew Tuttle on Investing in AI Infrastructure
The broader AI power debate has focused on the electrical grid. This report, however, argues the sharper constraint sits inside the building. Average server rack density across the industry climbed to 27 kilowatts in 2026. That’s up from seven kilowatts in 2021, the report found. AI hardware is overwhelming electrical systems built for a different era.
Electricity Demands Surge Inside the Rack
Traditional server racks, the metal frames holding a data center’s servers and networking gear, once drew 5 to 15 kilowatts. Nvidia’s GB200 platform, built for AI computing, runs at 100 to 137 kilowatts, the report found.
The upcoming Vera Rubin platform is projected to reach 200 to 300 kilowatts per rack, according to the report. Rubin Ultra is expected to exceed 600 kilowatts. Air cooling stops working above 40 to 50 kilowatts, pushing operators toward liquid systems that cool the chip directly.
That shift also costs more. AI-optimized data centers spend about $4.6 million per megawatt on cooling, versus $2.4 million at traditional sites, the report found.
Electrical infrastructure costs have climbed too. AI-optimized facilities spend roughly $3.6 million per megawatt on grey space electrical work, covering transformers and switchgear inside the building. Traditional facilities spend about $2.2 million on that same category, according to the report.
How Power Moves Through the Building
Electricity does not arrive at a server ready to use. It enters at medium voltage from the grid, steps down through transformers, and passes through switchgear and backup systems. It then travels through distribution units before reaching the rack. Each handoff adds cost, delay, and lost energy.
One fix gaining ground is a shift to 800-volt direct current distribution, which sends power to the rack in fewer steps. The approach cuts copper requirements by more than 40% and lifts efficiency to 92% — 95%, according to the report. That compares with 75% to 85% for conventional systems.
Small shipments are expected to begin in late 2026, though the industry has not settled on a single standard. Nvidia favors native 800V, while hyperscalers including Meta Platforms, Inc. (META) and Alphabet Inc. (GOOGL) favor a different design, the report noted.
Roughly one-third of planned U.S. data center capacity is expected to include on-site power generation, the report found. That includes gas turbines and fuel cells. The equipment helps developers skip utility interconnection queues that can stretch two to four years. But it adds another layer to the building’s electrical stack.
Infrastructure Firms Feel the Strain
Equipment backlogs show where the strain is concentrated. Eaton Corp. (ETN) reported data center orders up 240% in the Americas, with total backlog up 31% year over year, according to the report. Eaton, Vertiv Holdings Co. (VRT) and Schneider Electric have each flagged the same trend. Equipment content sold per megawatt is nearing double traditional levels for AI-optimized deployments.
The Thornburg American Opportunities Fund (TAOZ) and the Thornburg Focus Growth Fund (TFGZ), both launched April 1, 2026, are actively managed strategies. Rather than track a fixed index, the funds aim to capture that kind of shift directly.
TFGZ counts Vertiv Holdings Co. and Argan, Inc. (AGX), a power infrastructure contractor, among its top ten holdings, according to the fund’s factsheet. TAOZ holds Coherent Corp. (COHR), an optical components maker, at 4.3% of its portfolio, according to VettaFi.
TAOZ managed $8.75 million in assets and TFGZ managed $7.51 million as of August 10, according to VettaFi.
At least 13 U.S. states have introduced legislation to pause or restrict new data center projects, the report found. Roughly 34 gigawatts of planned capacity is now classified as stranded or delayed. Virginia, home to a dense cluster of data centers, recently passed a per-kilowatt-hour electricity tax aimed specifically at AI facilities.
Register here for a VettaFi webcast with Thornburg Investment Management to learn more.
Originally posted on ETF Trends
For more news, information, and strategy, visit our Portfolio Strategies Content Hub.
A message from Advisor Perspectives and VettaFi: Discover something new! Click here to register for our upcoming webcasts.
More Mutual Funds Topics >
