86% of Enterprise GPUs Operating at Half Capacity or Below

A survey revealing the current state of AI investment has been released. According to a VentureBeat Research survey conducted in June 2025, 86% of enterprises operating their own GPUs reported utilization rates of 50% or below. The survey targeted 573 technology leaders from companies with 100 or more employees, conducted as five parallel studies spanning each architectural layer of AI systems.

What this figure means is that many enterprises are purchasing and installing expensive AI hardware while actually utilizing less than half of its capacity. AI GPUs have very high deployment costs, representing significant capital expenditure for enterprises. The low utilization rate suggests that investments may not be yielding commensurate returns. While Wall Street continues to debate whether AI investment is excessive, this survey provides data directly measured by the purchasing enterprises themselves, carrying considerable weight.

What makes the problem even more serious is insufficient cost tracking. Only 44% of enterprises answered that they are "strictly tracking" AI computational costs and return on investment, meaning the remaining enterprises operate based on rough estimates. Additionally, the survey revealed that 27% of enterprises only become aware of AI agent operational costs after receiving invoices—a "reactive management" state. Operations continue without budget caps or real-time monitoring in place.

Nevertheless, enterprise appetite for AI investment remains strong. Among new computational platforms being evaluated within the next 12 months, 45% of enterprises cited AI-specialized cloud services (CoreWeave, Lambda, Crusoe, Nebius, and others). However, fewer than 2% of enterprises are actually using these emerging cloud providers at present. There is a significant gap between interest and current reality.

A shift toward reducing Nvidia dependency is also evident from the survey. When asked about emerging computational platforms likely to be evaluated within the next 12 months, 32% of enterprises mentioned accelerators other than Nvidia, including AWS Trainium, Google TPU, and AMD. This slightly exceeds the 28% citing next-generation Nvidia GPUs, suggesting a certain level of caution regarding Nvidia, which maintains an overwhelming market share in GPUs.

The survey also highlights that enterprises are moving forward with AI agent deployment before establishing governance frameworks. Fifty-four percent of enterprises experienced either security incidents related to agents or near-misses detected before actual harm occurred in the past 12 months. Additionally, approximately 60% of enterprises indicated plans to switch or add vendors across various governance layers within the next 12 months, with around 30% in certain layers considering migration within a quarter.

The overall picture presented by this survey reveals a structural misalignment between the speed of AI adoption and the maturity of governance frameworks. Before deciding on new hardware or cloud service investments, enterprises should prioritize accurately assessing current GPU utilization rates and workload-specific costs. For enterprises to justify AI expenditures, building mechanisms to visualize return on investment has become equally or more important than technology selection.