[Survey: 79% of Enterprises' GPUs Are Idle for at Least Half the Time, 70% Still Plan to Expand AI Platforms] According to a Q3 2026 survey compiled by VentureBeat Intelligence, 79% of enterprise respondents indicated that their GPU computing power is idle for at least half the time. However, 70% of the surveyed enterprises still plan to introduce, expand, or replace their AI computing platforms within the next 12 months. Return on investment (ROI) assessments are also lagging. The survey shows that among the 64 respondents planning to adopt new AI platforms, 58 lacked ROI data audited by their finance departments, accounting for 91%. Furthermore, while underutilizing their own GPUs, these companies continue to call closed-source models from companies like OpenAI and Anthropic via APIs, resulting in both idle hardware resources and model call expenses. With the increasing prevalence of AI agents, the token consumption and inference costs associated with multi-step tasks are further raising concerns. Gartner predicts that by 2028, the inference cost of a single agent workflow will increase more than fivefold. The report points out that enterprise AI competition is shifting from vying for computing power to improving resource utilization and controlling the cost of individual tasks. Specific measures include migrating routine tasks such as classification and summarization to open models, optimizing context caching, and reducing unnecessary multi-agent collaboration. However, GPU idleness may also stem from enterprises reserving computing power for peak demand or planning to replace older platforms; the survey data itself cannot directly prove that an AI bubble has formed. [Original Link]