Enterprise AI Investment Outpaces Cost Management Capabilities

Enterprise AI infrastructure investment significantly outpaces the ability to understand and manage costs. A VentureBeat Pulse Research survey of 107 companies reveals the reality of this "compute gap." While only 21% of enterprises are able to run AI systems at full scale in production, infrastructure spending expansion appetite remains high, creating a mismatch between investment maturity and expenditure pace.

The severity of the problem becomes clear when examining GPU (graphics processing unit) utilization. 83% of surveyed companies reported GPU utilization rates of 50% or below. In other words, the vast majority of enterprises are unable to fully utilize more than half of their purchased processing capacity. Furthermore, fewer than 44% of companies said they can precisely track and understand the costs of their AI compute. This indicates that while accelerating investment, many enterprises lack mechanisms to measure return on investment.

Infrastructure procurement sources also show little signs of consolidation. 64% of surveyed companies indicated an intention to change or add infrastructure providers within the next 12 months. Furthermore, 38% said they plan to do so within the next three months. Such high provider-switching intent in such a fundamental category suggests the entire industry is in a phase of procurement review.

The criteria enterprises use when selecting providers are also noteworthy. 41% of companies said they prioritize ease of integration with existing systems, while 35% prioritized total cost of ownership (TCO). Conversely, only 8% of enterprises placed primary emphasis on surface-level token pricing (the per-unit cost when AI processes text). This indicates a growing trend toward prioritizing overall operational costs and usability after implementation over surface-level pricing.

When asked which infrastructure they plan to evaluate and assess over the next year, 45% of companies cited "AI-specialized cloud services" as their top choice. A discrepancy between plans and reality was confirmed: these services are currently used by almost no enterprises, yet they attract the most interest as a next investment target. Additionally, regarding the industry-wide shift where memory bandwidth (the speed at which data is transferred) becomes a bottleneck rather than GPU compute power as AI inference processing scales, approximately one in five companies either remains unaware or has not yet begun considering responses.

The picture presented by this survey can be framed as an "invisible cost" problem in AI investment. While spending expansion itself reflects enterprise enthusiasm for AI adoption, unutilized GPUs and untracked costs pose risks of diminishing investment efficiency. Establishing cost visibility mechanisms alongside infrastructure development may become a key factor determining the competitiveness of future AI adoption.

For enterprises rushing to procure AI infrastructure, the stage has arrived where "how to fully utilize and measure what you already have" is as important as "what to buy." Whether to invest in the "management foundation"—such as GPU utilization rates and operational cost visibility—will significantly impact the effectiveness of future infrastructure investments. This survey result deserves consideration as a question posed to the entire industry.