Survey Links Idle GPUs With Continued AI Platform Plans
VentureBeat Intelligence’s three-round VB Pulse survey found that 79% of surveyed enterprise respondents reported GPU capacity sitting idle at least half the time. Among companies with average utilization of no more than 50%, 70% planned to introduce, expand or replace an AI computing platform within 12 months. Of 64 respondents planning a new platform, 58, or 91%, lacked finance-reviewed ROI data. Among companies planning to procure a new computing platform whose GPUs were idle, 90% favored public clouds and model API providers, while 8% considered expanding local private clusters. The report also cited an NVIDIA guide reporting that 85%–97% of context could be reused in a Claude Code scenario and a Vals AI test in which multi-agent setups cost 1.8–5.1 times as much as a single agent, with a statistically significant performance gain in only one of four configurations.
The survey points to a possible gap between resource investment and actual demand at many respondents, while raising questions about infrastructure utilization, per-task cost, context caching and financial ROI review. The report cautions that the findings do not prove an industry-wide AI bubble or overcapacity.
AI 军备竞赛打到现在,场面似乎开始变得魔幻……
近日,科技媒体 VentureBeat 旗下研究机构 VentureBeat Intelligence 汇集的 2026 年第三季度连续调研数据显示,79% 的企业受访者表示,其 GPU 算力至少有一半时间处于完全闲置状态。
这一数据在三个月中始终稳定维持在 77% 至 81% 的高位。换句话说,在这批受访企业中,GPU 使用不充分已经成为一个持续存在的现象。
但更令人匪夷所思的是,在这批 GPU 平均利用率不超过 50% 的企业里,70% 的公司仍计划在未来 12 个月内引入、扩充或更换新的 AI 计算平台。
#VentureBeat Intelligence#VB Pulse#NVIDIA#Anthropic#Claude Code#OpenAI#Google#Vals AI#Gartner