Enterprise AI GPUs Underutilized; Companies Aware of Control Gaps
A recent survey of 573 technical leaders at companies with 100 or more employees reveals that 86% of enterprises running their own GPUs for AI workloads report utilization rates of 50% or less. This finding from VentureBeat Research directly contrasts with ongoing debates on Wall Street regarding AI buildout overcapacity, indicating that the most expensive hardware within these organizations is significantly underused. Compounding this issue, only 44% of companies rigorously track the actual costs and returns of their AI compute, with most relying on estimates. Despite this, a substantial portion of enterprises, 45%, are considering AI-specialized cloud providers like CoreWeave or Lambda within the next year, while nearly a third are exploring non-Nvidia accelerators as a potential hedge against Nvidia's dominance. The research suggests a priority on measuring current GPU utilization and per-workload costs before investing in new compute options.
Furthermore, the survey highlights that most deployed "agents" are essentially single-prompt chatbots rather than true autonomous agents capable of multi-step tasks, with 71% reporting that a quarter or fewer of their agents can complete such tasks independently. This discrepancy, termed "agentwashing," inflates adoption figures and impacts billing, as true agents require more robust control layers. Enterprises are also knowingly deploying agents ahead of necessary controls, leading to security incidents, with 54% reporting an incident or near-miss in the past 12 months. Control over agent spending is reactive for 27% of companies, who only learn costs upon receiving an invoice. The research also indicates that 66% of enterprises allow agents to push changes to production based on automated evaluations, yet only 5% fully trust these evaluations, with half of companies experiencing customer-facing failures due to agents that passed internal tests. Credential sharing among agents is prevalent, with 69% of companies allowing it, and these organizations face security incidents at a higher rate. Finally, 57% of enterprises have traced incorrect agent responses to missing or inconsistent business context, underscoring the need for better data governance and control layers.
The findings suggest a significant disconnect between the perceived urgency of AI adoption and the operational realities within enterprises. While substantial investments are being made in high-performance hardware like GPUs, their underutilization points to a potential overestimation of current AI capabilities or a lag in developing the necessary software infrastructure and workflows to leverage them effectively. The prevalence of "agentwashing" and the deployment of agents ahead of robust control mechanisms indicate a market driven by hype rather than measured progress. Enterprises appear to be retrofitting control layers and security measures after initial deployments, a reactive approach that incurs costs and risks. The exploration of alternative compute options and a hedging strategy against dominant hardware providers like Nvidia reflects a strategic effort to manage costs and mitigate supply chain risks. Moving forward, a more disciplined approach focusing on measuring ROI, enhancing agent autonomy through proper context and evaluation, and prioritizing security and cost controls will be crucial for realizing the full potential of AI investments.
AI-generated to prompt reflection — not editorial opinion, not advice, not a statement of fact. How this works.