VentureBeat's research arm has pooled three VB Pulse surveys fielded from July to September, and the results point to widespread idle AI compute in the enterprise, with no sign that purchasing plans are slowing. The sample covers 385 qualified respondents at organizations with 100 or more employees, including consultants answering on behalf of a main client; 218 of them reported GPU utilization.
79% (172 of the 218) said their GPUs are busy half the time or less, with the monthly figures ranging between 77% and 81%. Among those very companies with idle GPUs, 70% (120 of 172) plan to bring in new or replacement AI compute platforms over the next 12 months.
Idle or busy, the urge to buy is about the same
Among the 46 companies whose GPUs are busier, 74% plan to add a new platform. The article judges the gap between 70% and 74% to be statistically insignificant, meaning utilization has almost no bearing on appetite for new purchases.
Where the new platforms would come from is also highly concentrated. Of the 120 idle-GPU companies planning a new platform, 90% named a cloud platform or a model provider, and only 8% intend to go with on-premises or hosted clusters or a self-managed open-source stack.
The contrast shows up on the build-it-yourself side. Among the 32 companies running on-premises or hosted clusters or a self-managed open-source stack, 56% have GPUs idle at least half the time; for the other 186 companies, the figure is 83%. The survey did not ask why, and the author's explanation is explicitly labeled as speculation.
The claim that "nine in ten investments never passed a finance audit" comes from September data alone: of the 64 companies planning a new platform, 58 (91%) could not produce a return-on-investment figure reviewed by their finance department. That is a sample of one month and a few dozen companies, and the other months did not ask the same question.
A rough estimate of what idleness costs
Utilization affects unit cost inversely. Suppose renting a high-end GPU costs $2 per card-hour; an eight-card server then runs about $140,000 a year in rent (rough math: 8 cards × 8,760 hours × $2). At 80% utilization, each card-hour of real work costs about $2.50; at 40%, it becomes $5. Getting the same amount of effective compute would mean paying roughly $140,000 more a year, the equivalent of renting a second server.
The bill is rising on the other side as well. The article cites a Gartner forecast that by 2028 the inference cost of each agent workflow will rise more than fivefold, and Vals AI found that multi-agent coding teams spend 1.8 to 5.1 times as much as a single agent, with only one of four team configurations posting a statistically significant gain on benchmarks.
One way to save is caching. An NVIDIA guide shows that 85% to 97% of the context in follow-up calls within a Claude Code session can be reused from cache, and Anthropic charges between one-tenth and one-fortieth of the normal input price for cached tokens, depending on the model. Uber chief technology officer Praveen Neppalli Naga told Fortune in August:
“we're coming to the end of the so-called tokenmaxxing era”
Limits of the sample
These figures should be discounted. Respondents are VentureBeat readers and volunteers from its survey panel, so they do not represent all companies. The survey did not ask whether the GPUs are owned or rented, whether the company can schedule them directly, or whether idle capacity was deliberately held back for traffic peaks; a planned new platform may also simply replace an old one. Respondents who ticked several utilization options in July, and companies that do not measure GPU usage, were left out.
The report also doubles as a promotion for VentureBeat's own agent-infrastructure summit, whose sponsors include WEKA, Cisco's Outshift and OneTrust.
Sources: VentureBeat Intelligence survey report, Gartner, CocoLoop, Vals AI; we checked the GPU utilization sample sizes, the percentages for each group and the basis for the statistical-significance claims.