Survey finds enterprises investing in AI infrastructure faster than they can measure costs

A VentureBeat Pulse Research survey of 107 enterprises with more than 100 employees finds that AI infrastructure spending is accelerating faster than organizations’ ability to measure its economics. The survey, conducted in Q2 2026, reveals a widening "compute gap" driven by heavy investment in AI infrastructure that outpaces visibility into utilization, costs, and returns.

GPU utilization is a stark indicator of inefficiency: 83% of enterprises report utilization at or below 50%, with nearly half (49%) operating at 25% or below. Only 12% report utilization above 50%, and 8% do not measure utilization at all. This idle capacity compounds as organizations plan to expand their compute footprints, including specialized AI clouds that 45% intend to evaluate over the next year.

Production maturity lags spending ambitions. Only 21% of enterprises report running AI in production at scale, while 76% are still experimenting or running limited workloads in production. This early-stage posture shapes infrastructure decisions, which increasingly prioritize integration with existing stacks (41%) and total cost of ownership (35%) over headline pricing like cost per million tokens (8%).

A majority of enterprises plan to change or add infrastructure providers within a year. Sixty-four percent intend to switch or add a provider within 12 months, including 38% within the next quarter. The most-cited providers for switching consideration are incumbents such as Microsoft Azure (33%), Google Cloud (33%), OpenAI (30%), and Gemini (22%), suggesting near-term movement may consolidate spend among major players rather than drive defection to new entrants.

Despite prioritizing total cost of ownership in vendor selection, fewer than half (44%) of enterprises can rigorously track the cost and return of their AI compute infrastructure. Satisfaction with current infrastructure is moderate, averaging 4.0 out of 5, with the lowest scores on value for money (3.9) and ease of implementation (3.8).

The survey highlights an emerging technical constraint that remains under-addressed: the shift from GPU compute to memory bandwidth as inference scales. When asked how they would address this bottleneck, responses were fragmented, with 18% of enterprises either unaware of the constraint or yet to address it. This suggests the next phase of infrastructure decisions may be made without full visibility into the operational realities of large-scale inference.