The fact: A VentureBeat survey of 107 enterprises reveals that AI infrastructure spending is accelerating far ahead of organizations' ability to see or control its economics. GPUs run at 50% utilization or less, and fewer than half of enterprises rigorously track what their compute actually costs.
Context: The survey shows a paradoxical landscape. Most organizations run their AI on familiar hyperscalers and model-provider APIs, yet the next dollar of investment is heading toward specialized compute that almost none use today. A majority intend to switch or add providers within the year, many within a quarter. Buying decisions turn on integration and total cost of ownership rather than headline token price — which is fortunate, because most enterprises cannot yet see their unit economics clearly.
Analysis: This is the classic 'grow before you know the cost' problem — at unprecedented scale. Unlike conventional cloud computing, where FinOps tools (AWS Cost Explorer, Azure Cost Management) are mature, AI compute carries extra variables: volatile GPU pricing, competition for H100/B200 cluster access, inference costs that scale unpredictably with usage, and the asymmetry between training (predictable) and inference (explosive) costs. GPUs running at 50% or less is a sign of massive inefficiency: either companies are over-provisioning out of fear, or they lack the visibility to size correctly. The planned provider switching within the year suggests dissatisfaction with current cost models, but switching without clear metrics is rearranging deck chairs.
What to watch: Watch the growth of AI FinOps startups focused on GPU and inference cost observability. The compute broker market (CoreWeave, Lambda, RunPod) will expand, but the differentiator will be cost visibility, not just capacity. Also pay attention to CFO pressure — as AI spending grows as a percentage of IT budgets, the demand for precise ROI metrics will intensify. Companies that do not implement rigorous measurement within 6–12 months will be flying blind.
Source: VentureBeat
