79% of enterprises say GPUs sit idle half the time, yet 70% plan new AI buys

机器之心 · wechat · 2026-10-24

VentureBeat Intelligence's three-round survey of 385 enterprises (100+ employees) finds 79% report their GPUs are fully idle at least half the time (stable at 77–81% over three months), yet 70% of those low-utilization companies still plan to adopt or replace AI platforms within 12 months — and 91% of planned adopters have no finance-approved ROI data.

The mismatch stems from two diverging bills: panic-driven GPU stockpiles have become sunk assets, while actual workloads run on closed frontier-model APIs from OpenAI, Anthropic and Google, with token costs devouring budgets (Uber famously burned its full-year AI budget in months of "token maxxing"). Among idle-GPU buyers, 90% look to public cloud and API providers; only 8% would expand private clusters. Agentic AI will amplify this: Gartner predicts 5x+ growth in per-workflow inference cost by 2028, and ValsAI tests show multi-sub-agent setups cost 1.8–5.1x a single agent with statistically significant gains in only one of four configurations.

Suggested remedies: measure cost per completed task rather than per token; cache repeated context (NVIDIA found 85–97% of context in Claude Code sessions is cache-reusable, with cache reads at 1/40–1/10 of input price); add agents only for parallelizable tasks; migrate high-frequency routine work to open models on idle hardware; use spare GPU time for batch inference, evals and fine-tuning.

Original post →

79% of enterprises say GPUs sit idle half the time, yet 70% plan new AI buys

More from Companies & People