Infrastructure and compute: Enterprises are buying AI compute for speed while flying blind on what it costs

Across a sample of 170 enterprises, the adoption of AI infrastructure for production purposes has advanced significantly. Two-thirds of these organizations (66%) currently run AI workloads live, while 29% describe their AI deployments as operating at scale. Only a small fraction (4%) have not yet initiated any AI workloads. This maturity in deployment is reflected in the stack configurations, with the average enterprise utilizing three different infrastructure platforms.

Major players in the AI infrastructure landscape include OpenAI (49%), Google Gemini (48%), Microsoft Azure (47%), and Google Cloud (42%). When asked to identify their primary platform, Azure emerged as the top choice for 26% of respondents. The key shift in decision-making processes is evident in how enterprises evaluate their choices.

Integration with existing cloud and data stacks remains the primary factor (40%), followed by performance (35%) and GPU availability (24%) - both surpassing total cost of ownership (22%). Measurement priorities also reflect this operational focus, with uptime and reliability being the primary success metric for 51% of enterprises, and developer productivity ranking at 39%, ahead of cost per million tokens (31%).

Enterprises operating under production pressure are prioritizing speed, availability, and cost, indicating a shift in focus away from traditional cost considerations. However, this reordering of priorities raises concerns regarding the lack of rigorous tracking of AI compute costs, with only 69% of enterprises reporting GPU utilization of 50% or less, and only 23% effectively measuring utilization.

Additionally, only 47% of enterprises rigorously track what their AI compute costs, and this number drops to 56% among those operating AI workloads at scale. The survey highlights a deficiency in the ability to assess the economics of AI compute, with value for money being the weakest satisfaction score at 3.87, while overall satisfaction stands at 4.14.

This discrepancy emphasizes the challenges in measuring and improving the economic aspects of AI infrastructure. Looking ahead, the majority of enterprises (44%) plan to evaluate or invest in AI-specialized clouds, with CoreWeave and Lambda registering the highest usage at 3.5%. Non-Nvidia accelerators also show promise, accounting for 39% of the market.

Notably, 62% of enterprises intend to switch or expand their AI provider within the next 12 months, though the majority are expected to continue using the same incumbents they currently rely on. The methodology behind this study involved VentureBeat Pulse Research, which surveyed a cohort of organizations with over 100 employees, drawn from a single July 2026 wave.

While the sample size provides a valuable directional insight, it should be considered a directional signal rather than a precise measurement, as the survey is self-selected and not a probability sample.