Enterprises are staring at a “compute gap”: they’re stacking up AI infrastructure spending while many cannot rigorously account for the compute economics beneath it. In the VentureBeat Pulse Research survey of 107 organizations (Q2 2026, June), 83% report GPU utilization of 50% or less. Even more revealing, fewer than half, 44%, can rigorously track what their AI compute costs.

That mismatch is happening while AI deployment maturity is still early. Only about one in five (21%) run AI in production at scale. Three-quarters (76%) are experimenting or running only some workloads in production. In other words, the organizations most aggressively planning next steps are often not the ones that have already wrung out repeatable cost controls.

So why does this matter now? Because the next dollar is not going to the same place as the last dollar. Most enterprises run their AI on the familiar base of hyperscalers and model-provider APIs, but the next wave of planned evaluation is aimed at infrastructure they almost none of them use today. The report’s sharpest tension is that the single most-cited planned evaluation area over the next 12 months is AI-specialized clouds at 45%. Yet when asked what they use today, specialized “neocloud” GPU providers barely register near zero. If you are an operator, a CFO, or a board member, this is the moment where “we’ll optimize later” can become expensive, because procurement is moving faster than measurement.

When enterprises do switch or add providers, it is not usually because of a headline token price. Buying decisions are shaped by integration and total cost of ownership. Integration is the top driver for switching at 41%, and total cost of ownership is next at 35%. Cost per million tokens is the deciding factor for just 8%. That pricing indifference would be comforting if most enterprises already had clean unit economics. They don’t. With GPUs sitting at half utilization or less for most orgs, the gap between the theoretical cost model and the real bill can widen quickly, especially as they expand training and inference.

The churn intent is also unusually high for something so foundational. A clear majority (64%) plan to switch or add an infrastructure provider within twelve months, and 38% within the next quarter. This is “direction-of-travel” behavior, not incremental tinkering. The survey shows that every infrastructure approach is net-expanding in intent, but specialized AI clouds carry the highest net momentum (+24), edging out even the hyperscalers (+22). So the practical risk is simple: if you are preparing to move a meaningful share of compute off general-purpose cloud, but you are still unable to rigorously track compute costs today, you could be doing a fast re-platforming without a working dashboard for the economics.

Context matters here. The report emphasizes that the current stack is hyperscaler-and-API today: Google Cloud leads at 48%, and general-purpose clouds together with major model APIs account for essentially all current deployment. Specialized GPU providers such as CoreWeave, Lambda, Crusoe, Nebius and peers register at or near zero among these enterprises today. Only 6% run their own on-prem GPU clusters and 4% use a custom open-source stack. That suggests many organizations are evaluating a move that changes not just performance, but the whole measurement and governance surface area.

There is also a second-order technical constraint that could turn into a financial constraint. The report points to a shift from GPU compute toward memory bandwidth as inference scales, noting it is barely on the radar. Roughly one in five enterprises are either unaware of it or yet to address it. When enterprises are already struggling with utilization and cost tracking, this kind of “next bottleneck” is exactly how today’s savings plans can get steamrolled by tomorrow’s performance realities.

Finally, the survey’s methodology makes the signal directional but still actionable: it is self-selected, skewed toward the mid-market and earlier-stage adopters, and not a probability sample. At n=107, it is large enough to read directionally but should not be treated as precise market-wide census data. Even so, the internal logic of the findings holds together. With only 21% at production scale, 83% at 50% or less GPU utilization, and only 44% able to rigorously track compute costs, the compute gap is not a theoretical problem. It is a live operating condition that can drive overspend, vendor thrash, and governance headaches for any organization ramping AI.

If you are a founder, finance leader, or investor watching enterprise AI budgets, the takeaway is blunt: the organizations most likely to re-platform in the next year are often least equipped to measure unit economics today. The winners will be the ones who close the loop between procurement decisions and measurable cost behavior before the migration becomes irreversible.