Nvidia Claims Up to 40% of Data Centre Power Is Lost to Failures

What happened

Nvidia says data centres can use as little as 60% of the power they draw for actual AI work, with failures and restarts consuming around 10% of every watt, according to TechRadar's reporting. The claim comes alongside a new partnership between Nvidia and Vertiv, which aims to improve data centre efficiency and support the construction of more sustainable facilities.

The central point is the gap between the power a facility draws and the power that reaches productive AI computation. TechRadar attributes part of that gap to failed jobs and restarts, which take energy without delivering finished work.

The source material available for this briefing is limited to the headline and summary. It does not set out a methodology, timeline or the detailed terms of the partnership, so the efficiency figures should be read as Nvidia's claims until they are independently verified.

Why it matters

For organisations building or buying AI capability, compute efficiency is no longer a back-office concern. Wasted power translates directly into cost per training run, cost per inference and how much capacity is available for new AI services. The infrastructure layer now shapes what AI products can be delivered, how quickly, and at what price.

Restarts and failures also matter beyond the data centre. Each interruption at the infrastructure level can delay the output that customers, clinicians, analysts or citizens ultimately receive. Leaders planning AI-enabled services should treat infrastructure reliability as part of the service they promise, not as a purely technical detail.

By the numbers

• 60% — the share of data centre power that Nvidia says can be used for actual AI work, in the best case cited.
• 10% — the share of every watt that Nvidia says failures and restarts consume.

The Renascence take

Most coverage will focus on the headline percentages. The more useful question is what sits behind them: how much work is lost, and how often.

The real story here is not the watts, it is the wasted work. Every failed job and restart is a service interruption that someone pays for in time, money and trust, even if the customer never sees it. Loss aversion and visibility both point the same way: a gap that is measured and shown will get managed, and one that is hidden will be quietly tolerated. Operators who want to lead should ask vendors for useful-work-per-watt and restart rates in their service agreements, and treat them with the same seriousness as uptime.