Survey Finds Low Enterprise GPU Use as Firms Plan AI Compute Platform Changes

A September survey found low reported GPU use among some enterprises running AI in production, while many qualified respondents planned to adopt a different or additional compute platform within the following year. The results describe respondents’ reported use and plans, but do not establish an industrywide surplus of data-center capacity.

Among respondents whose organizations used AI in production and who provided a utilization figure, 81% reported average GPU or accelerator use of 50% or less. The survey excluded respondents who did not measure utilization; three enterprises that said they did not operate GPUs were also not counted in that finding.

Within a year of the survey, 64 of 102 qualified respondents (63%) expected to take up an AI compute platform that was new, additional or a replacement. The figure covers platform plans, not how much computing capacity respondents intended to add or whether they planned to build data centers.

Only 10 of the 102 respondents had an AI infrastructure return-on-investment figure reviewed and signed off by their finance organization. The survey received 178 responses, of which 102 qualified. Respondents were employees of organizations with at least 100 workers or consultants with direct access to such organizations’ AI systems. Consultants made up 43% of qualified respondents.

David Sherman, AI and financial inclusion strategist at io.net, said companies could make better use of GPUs already in service before expanding infrastructure. He described the approach as “making better use of what we already have.”

The results show reported underuse and planned platform adoption among the surveyed respondents. They do not measure GPU utilization at hyperscalers, distinguish owned GPUs from rented capacity or access through APIs, or establish how much data-center capacity respondents planned to build. The survey therefore does not determine whether current industry capacity exceeds future demand.