Enterprises eye AI compute buys without clear cost controls

Enterprises are preparing to shift AI infrastructure suppliers and evaluate specialized compute, but most still lack hard cost visibility, according to VentureBeat Pulse Research. The survey of 107 organizations with more than 100 employees did not disclose spending totals, which is part of the point: buyers are moving faster than their measurement systems.

VentureBeat describes the result as a “compute gap,” with infrastructure decisions running ahead of production maturity and cost controls. Only 21% of respondents said they operate AI in production at scale. Most are still experimenting or running limited workloads, yet many are already considering new providers and accelerator strategies.

Specialized clouds draw interest, not usage

Current AI workloads remain concentrated with familiar suppliers. Google Cloud was present in 48% of respondent stacks, while the broader mix was dominated by hyperscalers and major model APIs, including Microsoft, AWS, Oracle, Gemini, OpenAI and Anthropic. Specialized GPU cloud vendors such as CoreWeave, Lambda, Crusoe and Nebius barely appeared in current usage. Only 6% reported running their own on-premise GPU clusters, and 4% used a custom open-source stack.

The planned evaluation list looks different. AI-specialized clouds were the top category respondents expect to assess over the next 12 months, at 45%. Non-Nvidia accelerators followed at 32%, with next-generation Nvidia silicon at 28%. Decentralized compute networks drew 16%, and sovereign compute 11%.

That gap between current deployment and planned evaluation suggests specialized compute vendors are getting serious consideration, but not yet production trust from this cohort. VentureBeat also said similar interest appeared in its April-May survey wave, where specialized cloud usage was still marginal.

Provider churn intent is high

Sixty-four percent of respondents said they plan to switch or add an AI infrastructure provider within 12 months, including 38% within the next quarter. For core infrastructure, that is a high level of stated churn intent. It should still be read as intent, not completed migration.

The near-term beneficiaries may still be incumbents. Microsoft Azure and Google Cloud each drew switching consideration from 33% of respondents, followed by OpenAI at 30% and Gemini at 22%. That points to redistribution among large platforms before a broader move to newer GPU cloud providers.

Buying criteria also cut against some vendor messaging. Integration with existing systems was the top factor at 41%, followed by total cost of ownership at 35%. Cost per million tokens, a common pricing comparison in AI services, was the deciding factor for only 8%.

Utilization and cost tracking lag

The weaker part of the market is operational discipline. Among enterprises operating GPUs, 83% reported utilization at 50% or less, and 49% reported utilization at 25% or less. Only 12% said utilization was above 50%, while 8% did not measure it.

Cost controls are also incomplete. VentureBeat found that 44% of enterprises rigorously track the cost and return of AI compute. Others reported partial tracking, no ability to quantify costs yet, or that measurement had not been prioritized. Respondents rated overall infrastructure satisfaction at 4.0 on a five-point scale, with ease of implementation at 3.8 and value for money at 3.9.

The survey also found limited preparation for memory bandwidth and KV-cache capacity becoming a larger inference bottleneck. Dell was the most cited approach at 31%, Nvidia followed at 16%, and about 18% of respondents either did not recognize the issue or had not started addressing it.

The sample was self-selected, drawn from one June wave in Q2 2026, and skewed toward mid-market companies and earlier-stage AI adopters. That makes the findings directional rather than a precise market census. Even with that caveat, the signal is clear enough: enterprises are evaluating more AI compute before they can fully account for the compute they already use.

This story draws on original reporting from VentureBeat.