Analysis
VentureBeat Research's June survey of 573 technical leaders at companies with 100 or more employees found that 86% of enterprises running their own GPU infrastructure report utilization of 50% or less -- meaning the most expensive hardware many companies own sits idle or underused more often than not. The finding is notable specifically because it's a buy-side measurement, coming directly from the enterprises making GPU purchasing and deployment decisions, rather than a sell-side or vendor estimate of demand.
The data lands squarely in the middle of an active Wall Street debate over whether the roughly $700 billion in hyperscaler AI capital expenditure projected for 2026 is being justified by genuine enterprise demand or is running ahead of it. A separate VentureBeat finding from the same research package shows enterprises are running AI agents ahead of the governance controls needed to manage them, and are doing so knowingly -- with roughly six in ten planning to switch or add vendors across each of five distinct control layers within the next twelve months, suggesting current tooling isn't meeting the bar enterprises actually need.
The underutilization finding sits in tension with continued strong demand signals elsewhere in the AI supply chain: SK Hynix's $26.5 billion IPO was more than seven times oversubscribed this same week, and hyperscalers continue committing to multi-year compute buildout plans. That tension -- enterprises reporting genuinely low utilization while public markets keep rewarding infrastructure suppliers -- suggests the gap between AI compute supply and demand hasn't yet been fully priced by either enterprise budget owners or public investors.
For enterprise technology leaders, the 86% figure is a strong argument for auditing actual GPU utilization before committing to further infrastructure spend, and for exploring more flexible cloud-based or shared-capacity alternatives to owned hardware that sits idle most of the time. For infrastructure investors, the finding is a leading indicator worth tracking closely: if enterprise utilization doesn't improve, the current pace of GPU and memory capital spending eventually has to reconcile with actual workload demand.
The bear case: utilization surveys can undercount genuine future-proofing rationale -- some enterprises deliberately overprovision GPU capacity to avoid supply shortages during demand spikes, meaning low average utilization doesn't necessarily indicate wasted capital. What to watch next: whether hyperscaler capex guidance for 2027 reflects any moderation in response to utilization data like this, and whether more enterprises shift toward on-demand cloud GPU consumption rather than owned infrastructure.
