Compute#Cloud#Market research#AI chips
VB Pulse survey: 79% of enterprises run GPUs at half capacity or less — yet 70% plan new compute platforms
VentureBeat's VB Pulse surveys: 79% of enterprises report GPUs at half capacity or less, yet 70% plan a new or replacement AI compute platform within a year.
VentureBeat Intelligence, the reader-survey research arm of VentureBeat, asked production-AI operators the same question in three monthly VB Pulse questionnaires from July through September 2026: how busy are your GPUs. The pooled answer, published October 23, is awkward for an industry that keeps buying silicon. Of the 218 companies that gave a utilization figure, 79% (172) said their GPUs run at half capacity or less — and 70% (120) of those same companies still plan to adopt a new, additional or replacement AI compute platform within 12 months. One caveat belongs in the opening paragraph rather than a footnote: respondents are self-selected VentureBeat readers and panelists, and 43% are consultants answering for a client organization. This is a self-portrait of one outlet’s enterprise audience, not an industry measurement.
Key points
- “Idle” means self-reported utilization at or below 50%, not dead hardware. The share moved between 77% and 81% across the three months, and VentureBeat’s own tests call the month-to-month differences statistically insignificant. In September alone, 31% said their GPUs do useful work a quarter of the time or less (base: 71 production enterprises, excluding the 4% that don’t measure utilization).
- 70% is the idle group’s number. Companies whose GPUs are busy want more capacity too: 34 of 46 (74%) plan a new platform, and VentureBeat finds the gap with 70% insignificant. The headline contrast holds, but it joins two different cuts of the same panel.
- 91% of buyers have no finance-reviewed ROI. In September, 58 of 64 companies planning a platform purchase within 12 months had no finance-reviewed ROI figure; across all respondents, only 10% do.
How the survey was run
The methodology notes are candid about the limits. September drew 178 responses, 102 of them qualified; across the three months, 385 qualified respondents came from organizations with at least 100 employees. Each month’s group is self-selected and different, so the results describe “those respondents rather than every company,” as the report puts it. Utilization is self-reported, and the questionnaire never asked how the numbers were produced. Only differences passing a Fisher’s exact or chi-square test at p < 0.05 count as significant — which is why so many of the survey’s own comparisons get labeled insignificant. Subsamples stay small: the September utilization question has a base of 71 production enterprises, or 68 excluding the 4% that don’t measure utilization, and 81% (55 of 68) put it at 50% or less. That figure is where any “idling climbed to 81%” line comes from; in reality the pooled share sat in the same 77–81% band all three months.
The numbers in detail
The pooled numbers, split by cut (per the October 23 analysis and its report page):
| Question | Pooled (Jul–Sep) | September only |
|---|---|---|
| Self-reported utilization ≤50% | 79% (172/218) | 81% (55/68) |
| Utilization ≤25% | — | 31% |
| Idle firms planning a platform within 12 months | 70% (120/172) | — |
| Busy firms planning a platform | 74% (34/46) | — |
| Idle buyers naming a cloud platform or model provider | 90% (108/120) | — |
| Idle buyers naming self-built or self-managed stacks | 8% (9/120) | — |
| Buyers with no finance-reviewed ROI | — | 91% (58/64) |
September’s remaining details: 63% of enterprises plan a new AI compute platform within 12 months, 75% of them within six. Among production enterprises, 68% use Google Cloud, Gemini or both, against 41% for OpenAI, 36% for Microsoft Azure and 24% for Oracle Cloud. Among the enterprises planning a platform, Google Cloud and Oracle Cloud each sit at 31% under consideration and OpenAI at 29%. Specialized AI clouds such as CoreWeave, Lambda, Crusoe and Nebius run in production at just 12% (18% are under consideration), though 29% name them the emerging option they would most likely evaluate. The only month-over-month change that passes a significance test is cost awareness: the share naming cost per million tokens as a factor in choosing AI infrastructure (multiple answers allowed) rose from 11% in August to 23% in September.
Idle but still buying
VentureBeat’s explanation is two bills coming due at once. Enterprises’ own GPUs cannot run closed frontier models, and inference spending keeps climbing: Gartner projects the inference cost per agentic workflow will rise more than fivefold through 2028, and Uber CTO Praveen Neppalli Naga told Fortune in August that “we’re coming to the end of the so-called tokenmaxxing era” — after Uber burned through its 2026 AI budget within months. On the thriftier side of the ledger, Nvidia engineers measured 85%–97% of context served from cache after the first call in a Claude Code session, and Anthropic charges 1/40 to 1/10 of the normal input price for cached tokens; Vals AI found that coding-agent teams — a lead agent plus up to five subagents — cost 1.8 to 5.1 times a single agent, with only one of four team setups significantly improving benchmark scores. Together those figures explain where the next dollar goes: 90% of idle buyers name a cloud platform or model provider, 8% a self-built cluster or self-managed open-source stack. The self-managed group reports the least idling of all — 56% (18 of 32) at or below half utilization, versus 83% (154 of 186) of everyone else — the survey didn’t ask why; the article’s own guess is that a team running its stack sees every idle hour directly and has the engineers to fill it. The survey leaves two gaps open. It never asked whether the GPUs are owned, rented or reached purely through an API, and it cannot rule out deliberate idling: companies reserve GPUs for demand spikes, and a “new” platform may simply be a replacement for an old one.
What others say
WebPulse, a third-party read-through site (AI-assisted, crediting VentureBeat as the original report), turns the survey into a pre-purchase checklist: a self-selected sample with self-reported numbers cannot say why utilization is low. “Picture a factory running at half capacity. Before building a second one, a sensible owner asks why the first is half empty,” it writes, and its first question for any buyer is “What are the idle GPUs waiting for?” The cache findings back that skepticism — average utilization is a blunt figure, one that cannot tell wasted cycles from cache hits that save money. WebPulse also sets the finding against Jensen Huang’s October 21 message at GTC Berlin that companies should stop merely “consuming the AI” and start “producing” it. Chinese coverage came through Jiqizhixin’s full compilation on October 24, whose “completely idle for at least half the time” phrasing overshoots the questionnaire, which asked about average utilization. And on the supply side, Andy Jassy said the same day that AWS is sold out for 2026 and largely sold out for 2027 (see our October 24 entry): vendors are rationing while enterprises report half-idle fleets, and the two datasets belong on the same desk.
Our take
For anyone budgeting compute, the useful part of this survey is its cross-section of the buying process: 91% of buyers have no finance-reviewed ROI, and of the 22% who say they track returns rigorously, fewer than half (45%) hold a finance-reviewed figure. The 79% headline travels far on its own; these two numbers are what should slow a purchase, because most platform decisions have not been through finance. If someone is pushing to rent another block of GPUs, walk WebPulse’s checklist in reverse: measure your own utilization, price a million tokens, and only then argue about a new platform. Our October 18 entry covered the other pole — AWS’s CFO claims AI servers pay back in under three years, with $496B in Q2 backlog (see AWS CFO: AI servers pay back in under three years). Vendors prove demand with backlogs; this survey proves idleness with self-reports. Both can be true at once, and the gap between them is exactly the arithmetic a purchasing decision has to do.
Three doubts are worth keeping. First, the sample is self-selected VentureBeat readers and 43% consultants answering for clients — cite none of it as an industry census. Second, utilization is self-reported, and an average cannot separate deliberate headroom, cache effects and genuine waste. Third, every third-party reading we found builds on the same survey; there is no independent measurement on the other side, and that absence is part of the story.