AI data centers waste up to 40% of power, with shortages now the bigger bottleneck than chips

Power inefficiencies in AI data centers

According to TechRadar, Nvidia claims that AI data centers use only 60-70% of their power consumption for actual AI workloads. The remaining 30-40% is lost to inefficiencies such as cooling, power distribution, conversion, failures, and restarts.

In a one-gigawatt data center, as much as 20% of power is lost through cooling systems, power distribution, and conversion processes. An additional 10% is wasted due to failures, restarts, and incorrectly configured racks.

Nvidia and Vertiv are collaborating to address these inefficiencies by designing sustainable AI data centers that maximize power efficiency. Rod Evans, Nvidia’s Vice President of Supercomputing, AI Cloud Infrastructure, and AI/HPC for EMEA, and Paul Ryan, Vertiv’s EMEA President, are reportedly leading this effort.

Power shortages emerge as the new bottleneck

Power availability has overtaken chip shortages as the primary challenge for AI data center expansion, according to reports from TechRadar and Mint (Technology). Some data centers wait years to secure adequate power supply, delaying their ability to operate at full capacity.

Morgan Stanley estimates that US data center developers could face a 32-gigawatt (GW) power shortfall through 2028, representing roughly 34% of projected demand. This shortfall could widen to 68 GW by 2029 if current trends persist.

The shortage is concentrating power access among large, financially robust companies. Smaller developers may struggle to secure electricity, potentially leaving them at a competitive disadvantage.

Geographic and infrastructure challenges

Data center location plays a critical role in power availability, permitting, and equipment transportation. Remote or constrained sites can delay projects by years, adding complexity to expansion plans.

Mint (Technology) reports that 14 GW of demand from US hyperscalers may shift to Asia by 2030 due to power constraints in the US. However, this relocation is unlikely to fully resolve the US power shortage.

Major technology companies and specialist cloud providers are projected to account for 94% of global data center capacity additions in 2026, further consolidating the industry.

Broader context in AI infrastructure

The shift toward AI-driven infrastructure has intensified demand for power and efficient data center design. Companies like Nvidia have previously invested in European AI startups and supported initiatives to improve AI safety and computational capacity.

Power efficiency and sustainability have become key focus areas as data centers scale to meet AI workloads. This mirrors broader industry trends, where access to electricity and infrastructure is increasingly critical to competitiveness.

What this means

Lazyfounder analysis — our interpretation, not reported fact.

For founders and operators in the AI and data center space, this story underscores a critical shift: power, not silicon, is now the binding constraint. The inefficiencies highlighted—cooling losses, misconfigured racks, and stranded capacity—are not just technical problems but business risks. A data center that wastes 30-40% of its power is effectively leaving revenue on the table, especially as electricity costs rise and availability shrinks.

The consolidation of power access among hyperscalers and well-capitalized players should be a wake-up call for smaller startups. If your growth plan assumes unfettered access to data center capacity, you may need to revisit it. Partnerships with infrastructure providers like Vertiv, or co-location strategies with credit-worthy anchor tenants, could become table stakes.

Finally, the geographic shift in demand—particularly to regions with more available power—suggests that founders should evaluate infrastructure as a first-order variable in their expansion plans. A data center in the wrong location isn’t just inefficient; it could become a stranded asset.

Key takeaways

• Power efficiency is now a competitive advantage for AI data centers, with inefficiencies costing up to 40% of total consumption.
• Founders scaling AI infrastructure must prioritize power availability and location as core factors in their growth strategy.
• Smaller players risk being squeezed out of the market as power shortages concentrate access among large, credit-worthy companies.
• Collaboration with infrastructure partners can help mitigate inefficiencies, but long-term solutions require systemic changes in power supply and data center design.
• Geographic flexibility—such as evaluating sites in Asia—may become necessary for US-based hyperscalers facing power constraints.

FAQ

Why are AI data centers only using 60-70% of their power for AI work?

The remaining 30-40% is lost to inefficiencies such as cooling, power distribution, conversion, failures, restarts, and misconfigured racks. These losses are a combination of technical limitations and operational challenges.

How severe is the US power shortfall for data centers?

Morgan Stanley projects a 32-GW shortfall through 2028, which could widen to 68 GW by 2029. This represents a significant constraint on data center expansion and AI growth.

Will shifting demand to Asia solve the US power problem?

No. While 14 GW of US hyperscaler demand may shift to Asia by 2030, this relocation is unlikely to address the broader US power shortage.

Related on Lazyfounder

• Nvidia
• More technology news