Enterprise AI spending outpaces measurement: 64% plan to switch infra within a year
A VentureBeat Pulse Research survey of 107 enterprises (100+ employees, Q2 2026) surfaces a specific and consequential gap: organizations are committing to AI infrastructure faster than they can account for what that infrastructure costs. Sixty-four percent plan to switch or add a provider within twelve months, yet fewer than half rigorously track AI compute cost and return. The compute gap the survey names is real, but it is worth separating what the data shows from what the framing implies.
The directional finding is consistent across the survey. Current deployment concentrates on hyperscalers and model APIs — Google Cloud at 48%, with Microsoft, AWS, Oracle, and the major model providers accounting for essentially all current usage. Specialized GPU clouds that dominate infrastructure headlines, including CoreWeave, Lambda, and Crusoe, register at or near zero. Yet the single largest planned evaluation area is AI-specialized clouds at 45%, a category almost none of these enterprises run today. Nearly a third intend to evaluate non-Nvidia accelerators, and 28% next-generation Nvidia silicon. This is not incremental; the survey frames it as a re-platforming in progress.
The switching intent reinforces the reading. Sixty-four percent plan to switch or add a provider within twelve months; 38% within the next quarter. For a category as foundational as compute, that is unusually high churn intent. The providers drawing the most switching consideration — Microsoft Azure and Google Cloud at 33% each, OpenAI at 30% — are incumbents, which the survey reads as reshuffling among the majors rather than defection to new entrants. The neocloud interest in the evaluation data is a 12-month thesis; the near-term movement is incumbents trading share.
What the survey cannot establish is whether the stated switching intent reflects actual procurement behavior or survey-grounded aspiration. Organizations that say they plan to switch within a quarter do not always execute on that timeline, and the 107-respondent sample is directional rather than definitive.
The measurement gap the survey documents is more concrete and more useful. Eighty-three percent of enterprises operating GPUs report utilization at 50% or less, and 49% run at 25% or below. Only 12% clear the 50% utilization mark. This is a well-known inefficiency in the industry, but the specific quantification matters: the capacity enterprises already own is substantially idle, and the efficiency headroom is large. The survey also finds that only 44% rigorously track AI compute cost and return. The majority track partially, cannot yet quantify it, or have not prioritized measurement. This creates a specific tension with Finding 5, where total cost of ownership was the second-ranked buying criterion at 35% — enterprises are choosing providers on an economic basis they mostly cannot yet measure. The stated priority and the measured capability are out of step.
The decision criteria themselves are informative. Integration with the existing stack leads at 41%, followed by TCO at 35%. Headline price — cost per million tokens — finishes last at 8%. This is coherent: buyers are optimizing for fit and real operating cost rather than the advertised unit rate. It is also fortunate, given the measurement gap, because token price is the one metric that requires the least internal instrumentation to compare.
Satisfaction data adds texture without resolving anything. Overall satisfaction averages 4.0 on a five-point scale, with ease of implementation (3.8) and value for money (3.9) trailing slightly. The softness lands, as the survey notes, on cost — the dimension hardest to judge without measurement. Whether the moderate satisfaction reflects genuine adequacy or low expectations in a category where benchmarks are immature is not answered by the data.
The final finding addresses a constraint that will shape the next round of inference economics: the shift from GPU compute to memory bandwidth as scale increases. Dell leads at 31%, Nvidia follows at 16%, and the rest fragments across storage vendors, open-source tooling, and model-level efficiency techniques. Roughly 18% either do not recognize the constraint or have not begun to address it. The survey frames this as an early and unsettled market. That framing is defensible, but the data cannot tell us whether the fragmentation reflects rational diversity of approach or genuine uncertainty about which solution will win.
Three caveats apply throughout. The sample is self-selected, skews mid-market (36% of respondents are organizations with 101–250 employees), and concentrates in Technology/Software, Healthcare, Financial Services, and Retail. It is best read as the view from organizations actively building out AI infrastructure rather than from the largest hyperscale operators. The survey measures stated preferences and self-reported utilization — both subject to the usual self-report biases, and the latter without external validation. And the single Q2 wave reads cross-sectionally rather than as a trend, so the consistency with the April-May survey on neocloud evaluation is suggestive but does not establish trajectory.
The compute gap the survey names is real in a specific sense: spending velocity exceeds measurement capability across the cohort. Whether that gap closes before the re-platforming arrives — or whether organizations commit to specialized clouds as blind to their economics as they are to the hyperscaler spend already on the books — is the question the survey poses without answering. At 107 respondents the direction is readable; the magnitude of the eventual re-platforming is not.