VentureBeat Pulse research from a single Q2 2026 (June) wave, surveying 107 enterprises with 100+ employees, finds that organisations are accelerating investment in AI infrastructure even though many cannot yet measure or control the underlying compute economics. Most respondents run AI on familiar hyperscalers and model-provider APIs today, but planned evaluations point toward specialized compute platforms they rarely use now.
Many organisations are not yet at production scale
Only 21% of respondents report running AI in production at scale, while 76% are either experimenting or only running some workloads in production. That maturity gap matters because the infrastructure and purchasing plans reported here largely reflect organisations still building out their AI footprint, whose compute demand — and costs — are likely to rise.
Current stacks: hyperscalers and model APIs dominate
The present deployment mix is concentrated in general-purpose cloud providers and major model APIs. Google Cloud appears in 48% of stacks reported; together Google, Microsoft, AWS, Oracle and model APIs such as Gemini, OpenAI and Anthropic account for essentially all current deployments in this sample. The specialist GPU "neocloud" providers (CoreWeave, Lambda, Crusoe, Nebius and peers) register at or near zero among these enterprises today. Only 6% run on-prem GPU clusters and 4% use a custom open-source stack.
(Note: respondents could select multiple providers — an average of 2.1 selections each — so these shares indicate presence in the stack rather than spend-weighted market share.)
The next dollar is aimed at infrastructure they mostly don’t run yet
Despite the current usage pattern, the top area enterprises plan to evaluate over the next 12 months is AI‑specialized clouds (45%). Other planned evaluations include non‑Nvidia accelerators (32%), next‑generation Nvidia silicon (28%), decentralized compute networks (16%), and sovereign compute (11%). Every infrastructure approach shows net expansion, but AI‑specialized clouds carry the strongest momentum (+24), edging even the hyperscalers (+22). This suggests many organisations plan to move a meaningful share of AI compute off general-purpose clouds.
Two separate VentureBeat waves (April–May and June) gave a consistent picture: specialised GPU clouds remain marginal in current usage, yet are the most commonly evaluated option for the year ahead.
A switching wave is building: many expect to change providers soon
A clear majority (64%) plan to switch or add an infrastructure provider within 12 months; 38% expect to do so within the next quarter. For such a foundational category, this is unusually high churn intent. The providers most considered for switching are incumbents — Microsoft Azure and Google Cloud (33% each), OpenAI (30%), and Gemini (22%) — indicating much near-term movement will reshuffle share among major vendors more than it will immediately shift customers to neocloud entrants.
Buyers prioritize integration and TCO over headline token price
When choosing infrastructure providers, respondents ranked integration with their existing stack highest (41%) and total cost of ownership (TCO) second (35%). The headline unit metric — cost per million tokens — was the deciding factor for only 8%. Buyers are optimizing for operational fit and real economic impact rather than advertised unit prices.
GPUs are mostly underutilised
Among enterprises operating GPUs, 83% report utilization at 50% or less; 49% report 25% or lower utilization. Only 12% exceed 50% utilization, and 8% do not measure utilization at all. Idle or underused accelerators are expensive to own, and this low utilisation is the clearest measure of the report’s so‑called compute gap: organisations plan additional GPU and specialised compute spend even while much of their existing capacity sits unused.
Measurement lags spending
Fewer than half (44%) rigorously track the cost and return of AI compute; 39% track it only partially, 20% cannot yet quantify it, and 6% have not prioritised it. This lack of economic visibility is consequential given that TCO is a top buying criterion. Reported satisfaction with current infrastructure is moderately positive (average 4.0 on a five‑point scale), with ease of implementation (3.8) and value for money (3.9) slightly lower, reflecting softness around cost assessment.
The next bottleneck — memory over compute — is not a mainstream focus yet
As large‑scale inference shifts from a compute‑bound problem toward constraints around memory bandwidth and KV‑cache capacity, enterprises’ plans are fragmented. When asked which approach they would rely on as inference constraints move to memory, 31% pointed to Dell, 16% to Nvidia; other answers included storage vendors, open‑source tooling, and model‑level efficiency techniques. Roughly 18% either do not recognize the emerging constraint or have not begun to address it.
Bottom line: a concrete compute gap that faster spending may widen
In this directional, mid‑market‑skewed sample of 107 enterprise respondents from a single Q2 2026 wave, the clear pattern is that appetite to spend on AI infrastructure is running well ahead of the instrumentation to measure and control those costs. GPUs already purchased are largely underutilised, fewer than half of organisations rigorously track compute economics, and many plan to evaluate specialised clouds and alternative accelerators within the year. The compute gap is not only a capacity issue; it is chiefly a visibility problem: organisations need to see what their hardware already costs before buying more of it. The open question for future waves is whether enterprises build that visibility before re‑platforming — or acquire the next layer of infrastructure with the same blind spots as the last.
Methodology: VentureBeat fielded this Pulse Research wave in June 2026 among organisations with 100+ employees (n=107). Because this is a single‑wave, cross‑sectional sample that skews mid‑market and self‑selected, the findings should be treated as directional rather than a precise population estimate. Respondents include managers, individual contributors, VPs/directors and C‑suite execs, with 45% final decision‑makers and 30% recommenders or influencers for AI purchasing. Industry representation includes Technology/Software (26%), Healthcare/Life Sciences (15%), Financial Services (13%), and Retail/E‑commerce (12%).
