Enterprise AI infrastructure spending is accelerating faster than organizations can track its costs, according to a June 2026 survey of 107 mid-market companies by VentureBeat Pulse Research. While only 21% of respondents run AI in production at scale, 64% plan to switch or add an infrastructure provider within the next twelve months, and 38% intend to do so within the next quarter. The data points to a widening gap between investment appetite and operational visibility.
The Compute Gap: Buying Blind
The survey's central finding is a stark mismatch between deployment ambition and financial control. Enterprises are preparing to evaluate AI-specialized clouds like CoreWeave, Lambda, and Crusoe at a rate of 45%, even though these providers barely register in current stacks. Meanwhile, the hardware already in place is running cold: 83% of GPU-operating enterprises report utilization at 50% or less, with nearly half running at 25% or below. Only 12% exceed the 50% threshold, and 8% do not measure utilization at all.
This inefficiency carries a direct cost. Idle accelerators burn budget without producing value, yet fewer than half of enterprises (44%) can rigorously track what their AI compute actually costs. The majority either track partially (39%), cannot quantify spend (20%), or have not prioritized measurement (6%). The result is a fleet of expensive hardware that is both underused and poorly accounted for.
Why Enterprises Switch
Provider churn intent is unusually high for a category as foundational as compute. When asked what drives selection, integration with the existing stack topped the list at 41%, followed by total cost of ownership at 35%. The metric vendors compete on most aggressively, cost per million tokens, ranked dead last at just 8%. Buyers want infrastructure that fits and operates efficiently, not the lowest headline rate.
The catch is that most cannot verify whether they are getting that efficiency. The same enterprises that rank TCO as a top priority lack the instrumentation to measure it. Satisfaction scores reflect this tension: overall infrastructure satisfaction averages 4.0 out of 5.0, but value for money trails at 3.9 and ease of implementation at 3.8. The softness lands exactly where measurement is weakest.
A Re-platforming in Slow Motion
Current stacks are dominated by familiar names. Google Cloud leads at 48%, with Microsoft Azure at 29%, AWS at 22%, and Oracle Cloud at 22%. Specialized AI clouds and on-prem GPU clusters barely appear. Yet the direction of travel points elsewhere. Beyond neoclouds, 32% of enterprises plan to evaluate non-Nvidia accelerators including AWS Trainium, Google TPU, and AMD Instinct, while 28% are looking at next-generation Nvidia silicon.
This is not a sudden defection. It is a gradual re-platforming driven by the mid-market's need for purpose-built AI infrastructure. The April-May wave of the same survey showed an identical pattern: specialized clouds drew the most evaluation interest despite near-zero current usage. Two waves of data, two different question framings, one consistent signal.
The Memory Frontier Nobody Is Watching
The next architectural constraint is already visible to specialists but not to most buyers. As inference scales, the bottleneck is shifting from raw GPU compute to memory bandwidth, specifically KV-cache capacity. When asked how they would address this, enterprises scattered across vendors: Dell led at 31%, Nvidia followed at 16%, and the remainder split across Hammerspace, DDN, VAST Data, WEKA, and open-source tooling. Most revealing, roughly one in five respondents either did not recognize the constraint or had not begun to address it.
This mirrors the broader visibility problem. A technical shift that will reshape inference architecture is arriving while many enterprises still lack a view of their current compute economics. The instrumentation gap that makes TCO opaque today will make the memory transition harder tomorrow.
What This Means for the Market
The survey sample skews mid-market, with 63% of respondents from organizations between 101 and 1,000 employees. These are not hyperscale operators with dedicated infrastructure teams. They are the companies building out AI capabilities from earlier stages, and their decisions will shape vendor fortunes over the next two years. The data suggests a market in flux: incumbents will trade share in the near term while specialized providers win evaluation cycles. The winners will be those that solve the visibility problem first, because buyers are optimizing for fit and true cost, they just cannot yet see it clearly.
The open question is whether enterprises build measurement and governance before the next wave of infrastructure arrives, or whether they buy the next layer as blind to its economics as the last. The next Pulse Research wave will show which direction holds.