美國數據中心的電力質素問題影響 AI 工作負載
AI Workloads Hit by Power-Quality Issues at Data Centers
Fluke 對美國 500 名數據中心營運者及業界人士的調查顯示,近三成人稱電力質素問題削弱其部署或運行高密度 AI 和 GPU 工作負載的能力。僅 36% 認為電力基建已完全準備好支援更高密度 AI 工作負載,74% 則指電力或電網容量不足正中度或嚴重限制擴張。另有 79% 表示過去一年電力質素問題曾導致延誤、服務中斷、非預期停機或嚴重險情;報道説明,該數字不等於近八成人曾停機。

Grid access remains a major barrier to growth, with 74% of survey respondents saying insufficient power and grid capacity are moderately or severely limiting expansion.Getty Images
Electrical problems are constraining AI deployments at US data centers, with nearly a third of respondents to a Fluke survey saying power-quality issues have reduced their ability to deploy or operate high-density AI and GPU workloads.
The survey, which polled 500 data center operators and professionals across the United States, suggests facilities are facing internal infrastructure constraints alongside the broader struggle to secure electricity from utilities. Just 36% of respondents said their electrical infrastructure is fully prepared for higher-density AI workloads, while 74% said insufficient power and grid capacity are moderately or severely limiting expansion.
“The biggest risk is whether a data center’s internal infrastructure can reliably distribute and support that power at scale,” Michael Stuart, senior manager and technical fellow at Fluke, told Data Center Knowledge.
Power, Cooling, and Cabling Failures Create Risks
About three-quarters of respondents (76%) reported experiencing electrical or power-system issues. Separately, 79% said power-quality problems had contributed over the past year to delays, service disruptions, unplanned downtime, or serious near-misses.
Related:Rethinking High-Voltage Commissioning for AI Data Centers
The figure groups operational failures together with incidents that did not result in an outage. It does not mean nearly four in five respondents experienced downtime.
Stuart said weaknesses often emerge where power, cooling, and cabling systems intersect. Problems can include harmonics, voltage sags and swells, thermal hotspots, deteriorating cooling systems, and infrastructure connections that have not been adequately documented or verified.
“Since each data center is so complex and diverse, determining weaknesses within facilities should be done on a case-by-case basis,” he said.
Those conditions can limit the computing capacity a facility can support even after an operator has secured sufficient electricity.
“If electrical, mechanical, and network systems aren’t performing to specification, operators may be unable to fully utilize the capacity they’ve already secured from the grid,” he added.
Testing and Monitoring Before GPU Rollouts
For operators preparing to install high-density GPU racks, Stuart recommends measuring power quality as close as possible to the GPU loads along the electrical distribution chain.
The appropriate measurement points will vary by facility and may change during commissioning as actual load conditions become clearer. Electrical distribution and cooling also need to be evaluated together to establish whether they can sustain the intended workloads.
Related:New Approaches to Data Center Power Procurement
“If measurements reveal that either system cannot reliably support the intended GPU workloads, operators should address those issues before proceeding with deployment,” Stuart said.
The survey also identified gaps in operators’ ability to detect developing problems. Some 78% of respondents said they lacked comprehensive, continuous, real-time visibility across critical electrical systems. Another 15% said electrical problems were identified only after performance degradation, equipment failure, or downtime had already occurred.
Stuart said operators should assess existing monitoring coverage to identify where additional visibility would most improve reliability. Potential areas include power distribution, backup power, batteries, and switchgear, although priorities will depend on each facility’s infrastructure and current monitoring capabilities.
Labor Shortages Add Pressure to Maintenance
Recruiting qualified electrical technicians has become harder for 57% of respondents, adding pressure to the teams responsible for installing, testing, and maintaining infrastructure.
That hiring difficulty has not uniformly translated into slower expansion. Some 65% of respondents said skills shortages had not delayed infrastructure upgrades or new capacity.
Related:AI’s Expansion Hits the Limits of Local Power
Still, Stuart said workforce constraints can contribute to localized delays, inconsistent installation practices, and additional work to correct problems or verify quality.
“Not every problem is visible during an initial inspection; some only surface once electrical systems are operational,” he said.
Stuart said qualified electricians should perform and validate the work, supported by testing and diagnostic tools to identify faults that may emerge after installation and help ensure long-term reliability.
About the Author
Contributor
Nathan Eddy covers data center trends and technologies across multiple industries. A graduate of Northwestern University’s Medill School of Journalism, he is also a documentary filmmaker specializing in architecture and urban planning. He currently lives in Berlin, Germany.
來源:Data Center Knowledge · datacenterknowledge.com
.webp%3Fwidth%3D400%26auto%3Dwebp%26quality%3D80%26disable%3Dupscale&mode=full&exp=1795564800&sig=2aa79c35fbd12918)