In a VentureBeat Pulse Research survey of 107 enterprises (Q2 2026, June), the “compute gap” is not a theoretical concept. It is visible in two numbers that practically fight each other: 83% of organizations report GPU utilization of 50% or less, while fewer than half (44%) can rigorously track what their AI compute costs. In plain English, a lot of AI infrastructure is running, but the people paying for it cannot always say what it truly costs them per unit of work.

That mismatch matters because spending intentions are moving ahead of operational maturity. Only about one in five (21%) run AI in production at scale, while 38% are still experimenting (proofs of concept) and 37% have some workloads in production, but not across the organization. Yet the next year’s evaluation priorities point to infrastructure changes that will affect both performance and cost. Enterprises are signaling they want to steer the economics, but right now they are not measuring them tightly.

This is happening in a familiar stack, which makes the coming shift more consequential. Today, most enterprises run AI on hyperscalers and model-provider APIs. Google Cloud is the most-used platform overall at 48% (Microsoft Azure 29%, AWS 22%, Oracle Cloud 22%). On the model side, 41% use Google’s Gemini models, with OpenAI close behind at 40% and Anthropic at 12%. Meanwhile, specialized GPU clouds barely register today: only 6% run their own on-prem or co-located GPU clusters, 4% use a custom open-source self-managed stack, and under 2% each use specialized AI clouds such as CoreWeave, Lambda, Crusoe, Nebius, Together, Fireworks, and peers.

So the tension is obvious. A lot of the infrastructure people are already using is “standard” by enterprise procurement habits, but the infrastructure they plan to evaluate is drifting toward compute layers they do not yet run. Over the next 12 months, 45% plan to evaluate AI-specialized clouds (CoreWeave, Lambda, Crusoe, Nebius), which is a layer almost none of these enterprises use today. A further 32% plan to evaluate non-NVIDIA accelerators (AWS Trainium, Google TPU, AMD Instinct, Intel Gaudi, and in-house ASICs), while 28% cite Nvidia Blackwell (GB300) or next-generation GPUs. Other evaluation areas include decentralized or distributed compute networks (16%) and sovereign or region-specific compute (11%). Nine percent say none of the above.

If you are a board member or CFO, this is where the “compute gap” becomes a governance problem, not just an engineering issue. Enterprises are not settled on infrastructure vendors. A clear majority (64%) plan to switch or add an infrastructure provider within twelve months, and 38% plan to do it within the next quarter. That churn intent is unusually high for a category this foundational. And when these buyers decide, they claim integration and total cost of ownership matter more than headline pricing: integration with the existing stack (41%) and total cost of ownership (35%) are top decision factors. Cost per million tokens drives decisions for just 8%.

That is a revealing footnote. Token pricing is the easiest number to compare across vendors, but it is not the only driver of real compute economics. Utilization rates and how well teams can trace compute charges to actual workloads are what turn “we chose the best price” into “we understood our unit costs.” The survey’s finding that GPU utilization is often low (83% at 50% or less) suggests capacity is being bought faster than it is being used well. Add the finding that only 44% can rigorously track compute costs, and you get the practical shape of the gap: money flows into infrastructure upgrades faster than measurement systems, tagging discipline, and cost attribution mature.

There is also a hardware shift lurking beneath the surface that could widen the gap if it stays unaddressed. As inference scales, the frontier constraint moves from GPU compute toward memory bandwidth. According to the survey, this is barely on the radar: roughly one in five enterprises are either unaware of it or have yet to address it. If teams cannot model where bottlenecks will appear, they risk optimizing for the wrong bottleneck and paying for underperforming configurations, especially during a period when many are actively switching providers or adding specialized compute layers.

Zoom out and the strategic stakes become clear: these are not yet “optimized operators.” Only 21% run AI in production at scale, and the survey itself notes it is cross-sectional, self-selected, and not a probability sample. It also skews toward mid-market organizations and earlier-stage adopters that are actively building out AI infrastructure. Translation: the decisions being made today are often being made while the foundation is still being installed. For leaders across finance, engineering, and procurement, the challenge is to close the loop between buying and understanding, before rapid spend turns into expensive blind spots.