Nvidia: AI Data Centers Waste 40% of Power, Vertiv Deal Fixes It

How much power do AI data centers waste?

Nvidia says AI data centers waste 30–40% of the power they draw. For a one-gigawatt facility, about 20% is lost to cooling, power distribution, and conversion, while another 10% goes to incorrectly configured racks, failures, and restarts.

Nvidia Vice President Rod Evans shared the figures at Vertiv's annual EMEA event, noting that a data center contracted for 100 megawatts might only deliver 93 megawatts in practice. The losses are not just technical quirks; they represent a massive operational cost and a growing sustainability problem.

The remaining 60–70% is what actually reaches GPUs for AI processing. Nvidia argues that better design can recover a significant portion of the wasted power, making existing grid connections go further. This focus on AI data center power efficiency is central to the company's strategy.

• Cooling, power distribution, and conversion: ~20% loss
• Incorrectly configured racks, failures, and restarts: ~10% loss
• Total waste: 30–40% of power drawn from the grid

Why is power now the biggest bottleneck for AI data centers?

Chip shortages once limited AI data center growth, but power availability is now the main constraint. Many facilities wait years for grid connections, and some locations offer power without permits, or permits without power.

Vertiv EMEA president Paul Ryan explained that location is a major factor: you might get power in one place but no permit, or a quick permit but no power. Sometimes both are available, but the site is so remote that equipment delivery is difficult.

This bottleneck means AI capacity is increasingly determined by energy infrastructure rather than GPU supply. Nvidia and Vertiv aim to address this by designing data centers that use every watt more effectively. The shift echoes concerns raised by DeepMind CEO Demis Hassabis about AI's climate impact, though he remained silent on data center power in a recent talk.

What does the Nvidia-Vertiv partnership aim to achieve?

Nvidia and Vertiv have been collaborating since early 2026 to co-design next-generation data centers. Vertiv provides power and infrastructure management for Nvidia GPUs, with the goal of building sustainable facilities that maximize output per watt.

Rod Evans said the partnership focuses on getting the most out of every watt, not on using as much power as possible. This includes rethinking power and cooling configurations to match actual usage rather than peak load.

The collaboration could help data center operators reduce waste and deploy AI capacity faster, especially in regions where grid connections are scarce. It also aligns with broader industry pressure to make AI more energy-efficient, a topic that has gained urgency as AI models grow larger.

Can the wasted power be recovered?

Yes, Nvidia believes a significant portion of the 30–40% waste can be recovered. Data centers are typically designed for peak load and outages, not typical usage, so unused capacity can be tapped and cooling and power systems can be optimized.

By revising configurations and matching infrastructure to real workloads, operators can free up power for AI processing without building new grid connections. This is especially important as AI demand surges and grid capacity lags.

Nvidia's collaboration with Vertiv is a step toward standardized, efficient designs that could become a blueprint for future AI factories. However, the exact efficiency gains will depend on implementation and site-specific factors. For a deeper look at how AI is reshaping hardware, see our review of the Lenovo ThinkPad T14s Gen 6, which packs 50 TOPS of AMD AI performance.

Bottom line

Nvidia's claim that AI data centers waste up to 40% of power highlights a critical inefficiency in the AI boom. The partnership with Vertiv aims to recover that lost power and ease grid bottlenecks, making every watt count. For operators, prioritizing efficiency is no longer optional—it's essential for scaling AI sustainably.