79% of Companies Report GPUs Idle Half the Time, Yet Most Plan New AI Platforms

Quick Facts

• 79% of surveyed companies running AI in production report GPU utilization at 50% or less, across three monthly surveys from July through September 2026.
• 70% of those same companies plan to adopt a new, additional, or replacement AI compute platform within 12 months.
• 91% of companies planning a new AI compute platform had no finance-reviewed return-on-investment figure.

Enterprise GPU fleets are sitting idle across the country, yet the spending keeps coming. A pooled analysis of three monthly VB Pulse surveys covering July through September 2026 found that 79% of 218 respondents who reported GPU utilization said their hardware runs at 50% capacity or less. Among that group, 70% plan to add or switch AI compute platforms within the next 12 months.

The September wave of the survey was the sharpest. Of 68 respondents running AI in production who provided a utilization figure, 55 — or 81% — reported average utilization of 50% or less. The range held consistently: the idle GPU share fluctuated between 77% and 81% across all three months.

A deeper cut of the numbers is more striking. Thirty-one percent of respondents said their GPUs do useful work only 25% of the time or less.

Flying Blind on Costs

The utilization problem is compounded by a measurement gap. Among the 155 enterprises operating their own GPU infrastructure, 69% report utilization of 50% or less, and fewer than half — 47% — rigorously track what their AI compute costs and returns. Even among enterprises running AI in production at scale, that figure only reaches 56%.

The most alarming cost data point: 91% of all companies planning a new AI compute platform had no return-on-investment figure their finance team had reviewed. Only about one in five enterprises — 21% — run AI in production at scale, yet spending intentions are running well ahead of that maturity.

Fear of Missing Out Drives Purchases

The VB Pulse research points to fear of being left behind as the primary driver of continued purchases. Reports of GPU shortages in 2024 and 2025 pushed enterprises to join hyperscaler waitlists, sometimes waiting weeks or months for allocation callbacks, only to receive offers of fewer GPUs than requested on expensive multi-year commitments.

Dan Herbatschek, CEO of Ramsey Theory Group, described the pattern directly: "C-suites fear being compute-strained when agentic AI systems go live. We work with enterprise organizations, and they are buying ahead of demand. But it is really difficult to justify that investment now, when most companies lack production-ready use cases."

Debo Ray, founder of DevZero, put the behavioral root cause plainly: "If you have one bad outage, teams overprovision. If you have one missed GPU reservation, leaders panic-buy capacity."

Independent Data Backs the Survey

The VB Pulse findings align with a separate technical dataset. Cast AI's 2026 State of Kubernetes Optimization Report, drawn from direct measurements across tens of thousands of production clusters, found average GPU utilization of just 5%. CPU utilization fell to 8%, down from 10% the prior year. Memory dropped from 23% to 20%.

Cast AI co-founder Laurent Gil noted: "These are not survey responses or estimates. They come from direct measurements of production clusters and millions of compute resources, before any optimization was applied."

At 5% utilization, enterprises are effectively running with roughly 20 times the GPU capacity their workloads consume at any given moment. The financial math is severe: at current AI infrastructure spend levels estimated at roughly $401 billion annually, that figure implies approximately $381 billion in annual spend producing no computational output.

A Pricing Shift and a Macro Warning

GPU costs are moving in the wrong direction for underutilizing buyers. In January 2026, AWS raised H200 Capacity Block prices by 15%, citing supply and demand. It was the first GPU price increase since Amazon launched EC2 in 2006, breaking a two-decade trend of falling compute costs. A CPU core sitting idle costs cents per hour; a GPU sitting idle costs dollars.

The International Monetary Fund's October 2026 Global Financial Stability Report, released October 13, flagged weaker or less durable AI investment and profitability as one of three risks that could abruptly tighten financial conditions worldwide. The IMF warned that a correction in AI-linked equities and credit remained a salient threat.

For founders and executives evaluating infrastructure spend, the data presents a clear challenge: most organizations are buying compute faster than they can build workloads to fill it, at prices that are climbing rather than falling, with no finance-reviewed return on the investment.