Enterprise AI compute costs trail production deployment, survey finds

Enterprise AI compute costs are receiving less attention than performance and availability in a July survey of 170 organizations, even as 66% of respondents said they have AI workloads in production. VentureBeat Pulse Research found that 47% rigorously track AI-compute cost and return, a gap that limits what this respondent group can say about the economics of its live systems.

The results are a directional view of organizations with more than 100 employees, rather than a measure of the full enterprise market. The survey was self-selected and non-probability sampled, and was fielded once in July 2026. Still, its respondents are operating rather than merely testing AI: 29% said workloads were in production at scale, while 4% had not begun running AI workloads.

Why are enterprise AI compute costs hard to assess?

The survey does not establish why respondents lack rigorous cost-and-return tracking. It does show that their AI stacks are distributed: organizations used three infrastructure platforms on average. OpenAI was used by 49% of respondents, Google Gemini by 48%, Microsoft Azure by 47% and Google Cloud by 42%. Azure was the primary platform for 26%.

That mix sits on cloud infrastructure that respondents evaluate chiefly for fit and live-workload performance. Integration with an existing cloud and data stack was the most-cited selection factor, at 40%. Performance, including latency and throughput, followed at 35%, then GPU availability at 24%. Total cost of ownership ranked fourth, at 22%.

The success-metric question produced a different set of measures, but cost-related measurement also trailed other listed priorities. Uptime and reliability was the primary success metric for 51%, developer productivity for 39%, and cost per million tokens for 31%.

Utilization points to unused GPU capacity

Among the 155 respondents that operated their own GPUs, 69% reported utilization of 50% or less. Another 23% reported utilization above 50%, and 12% said they did not measure utilization. Utilization and cost-and-return tracking are separate survey measures, so the results do not identify the cause of either low utilization or incomplete cost visibility.

For operators, the findings separate procurement preferences from financial controls. In this directional respondent group, performance and GPU availability ranked above total cost of ownership as selection factors, while reliability and developer productivity ranked above cost per million tokens as primary success metrics. The survey does not say that organizations are unable to manage costs, only that fewer than half reported rigorous tracking of compute cost and return.

Specialized clouds draw interest, not proven migration

The next procurement cycle could add more vendors. Forty-four percent of respondents planned to evaluate AI-specialized clouds, the highest share for a planned evaluation area, and 62% intended to switch or add an infrastructure provider within 12 months. Those are evaluation and intent figures, not spending commitments or market-share data.

Current use of specialist providers remained limited in the sample: CoreWeave and Lambda were each used by 3.5% of respondents. The contrast suggests that buyers are considering alternatives to incumbent platforms while their existing deployments remain concentrated among hyperscalers and model APIs.

Several survey questions allowed multiple answers, so percentages in those questions need not total 100%. The July findings also should not be read as a trend against VentureBeat’s separate June survey, which had a different sample and reported different results.

This story draws on original reporting from VentureBeat.