Most companies running AI in production say their chips sit underused, yet a majority still plan to add or switch compute platforms within a year
Enterprises are buying AI compute the way some people buy gym memberships. The commitment is real. The attendance is optional.
New data from VentureBeat Intelligence shows that 79% of 218 respondents who shared GPU utilization figures said their GPUs are busy half the time or less. The finding comes from three monthly VB Pulse surveys spanning July through September. These are companies already running AI in production, not ones still kicking the tires.
Underuse is not slowing the shopping spree. Among the companies reporting low utilization, 70% said they plan to adopt a new, additional or replacement AI compute platform within the next 12 months.
The September wave of the VB Pulse survey found that 81% of 68 production AI enterprises reported GPU utilization at or below 50%. Earlier monthly surveys landed in a similar range, with 80-83% reporting utilization of 50% or less.
Some 31% of respondents said their GPUs do useful work only 25% of the time or less.
In the September data, 71% of respondents with low GPU utilization said they plan to adopt new AI compute platforms within 12 months. Across all 102 qualified respondents in that wave, 63% signaled intent to switch or add platforms.
Only 10% of the 102 qualified respondents have a finance-reviewed return on investment for their existing infrastructure.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Only 12% of production AI enterprises reported using specialized AI clouds such as CoreWeave or Lambda.
One reading is that the 70% figure reflects dissatisfaction rather than appetite. The survey question covered new, additional or replacement platforms. A company sitting on underused hardware might be looking to swap it for something that fits its workloads better, not stack more capacity on top.
With only 10% of respondents holding a finance-reviewed ROI, some of this spending may be running on strategic conviction rather than measured results.
The research notes that the gap between low usage and continued investment may create an opening for tools that optimize existing deployments or offer alternative compute structures.
The 12% figure for specialized AI clouds is worth watching. Providers like CoreWeave and Lambda currently serve a small slice of production AI enterprises in this sample, and a market where 70% of underutilizing companies plan to change platforms is not a static one.
The research flags that the 10% finance-reviewed ROI result raises questions about how AI performance gets measured. Companies that build utilization tracking and cost review into their AI programs before adding more compute may avoid paying twice for idle capacity.
The thing to monitor is whether utilization climbs in future VB Pulse waves. If the 80% range holds while purchase intent stays near 70%, the disconnect between what enterprises own and what they use will only widen.