# Enterprise AI buyers admit their GPUs are mostly underused > Source: > Published: 2026-07-13 14:04:22+00:00 [Matt Marshall](https://venturebeat.com/author/matt-marshall?ref=runtimewire)'s [VentureBeat Research survey](https://venturebeat.com/orchestration/wall-street-is-debating-the-ai-buildout-enterprises-just-answered-86-say-their-gpus-run-at-half-capacity-or-less?ref=runtimewire) landed on July 10th with a number that should slow every AI infrastructure purchase order: 86% of enterprises operating their own GPUs say those chips run at 50% utilization or less. That is the buyer-side answer to the market's AI buildout debate. The public argument has centered on whether hyperscalers, neoclouds and model labs are building too much capacity. VentureBeat's data points to a more immediate enterprise problem: many companies already bought the scarce asset and still cannot keep it busy. The research was fielded in June 2026 across 573 technical leaders at companies with 100 or more employees, split across five parallel surveys covering orchestration, reliability and evaluations, security and identity, infrastructure and compute, and context or RAG. VentureBeat says 81% of respondents recommend or decide AI purchases at their companies, which makes the survey a procurement signal rather than a consumer sentiment poll. The infrastructure slice had 107 respondents, so the exact percentage should be read as directional. The pattern is harder to dismiss. Enterprises are moving ahead anyway. VentureBeat found that 45% of GPU-operating enterprises say the emerging compute option they are most likely to evaluate in the next 12 months is an AI-specialized cloud such as [CoreWeave](https://coreweave.com/?ref=runtimewire), [Lambda](https://lambda.ai/welcome/ai-cloud?ref=runtimewire), [Crusoe](https://www.crusoe.ai/cloud?ref=runtimewire) or [Nebius](https://nebius.com/about?ref=runtimewire). Fewer than 2% say they use one of those neoclouds today. Another 32% named non-Nvidia accelerators such as [AWS Trainium](https://aws.amazon.com/ai/machine-learning/trainium/?nc2=h_prod_ai_trn&ref=runtimewire), [Google Cloud TPUs](https://cloud.google.com/tpu?ref=runtimewire) or [AMD Instinct](https://www.amd.com/en/products/accelerators/instinct/mi300.html?ref=runtimewire) as the emerging compute option they are most likely to evaluate, while 28% named next-generation Nvidia GPUs. That mix is good news for cloud and silicon challengers. It is also a warning. If enterprises cannot measure workload-level cost and utilization on assets they already own, switching vendors can turn an accounting problem into a multi-cloud version of the same problem. ### The underused GPU is a management failure VentureBeat's most useful finding sits underneath the 86% headline: only 44% of the surveyed enterprises say they rigorously track what their AI compute costs and returns. That leaves a majority estimating ROI on one of the most expensive line items in modern technology budgets. In older software cycles, rough capacity planning was survivable because compute was elastic and comparatively cheap. AI infrastructure punishes that sloppiness. A reserved cluster, an underfilled H100 pool or a poorly scheduled inference workload burns money whether the model improves or sits idle. For founders building AI infrastructure software, the buying motion is clearer. The durable wedge is proof: utilization by workload, cost by agent, cost by business unit, and budget controls before the invoice arrives. VentureBeat found 27% of enterprises have only reactive control of agent spend, meaning they learn what an agent costs after the bill shows up. That is the budget owner equivalent of flying blind. The survey also explains why enterprises keep shopping while capacity sits underused. AI teams are under board pressure to show progress, application teams are embedding agents faster than governance teams can keep up, and infrastructure leaders do not want to be the blocker when a business unit asks for more compute. Buying another option can feel safer than admitting the current stack is poorly instrumented. ### Many deployed agents are still chatbots with a bigger label The compute story connects directly to the agent story. VentureBeat found that 71% of enterprises say a quarter or fewer of their deployed "agents" can complete multi-step tasks without a person driving each step. Only 10% say true agents make up a majority of what they run. That cuts against the way enterprise AI adoption is often marketed. [Gartner](https://www.gartner.com/en/newsroom/press-releases/2025-08-26-gartner-predicts-40-percent-of-enterprise-apps-will-feature-task-specific-ai-agents-by-2026-up-from-less-than-5-percent-in-2025?ref=runtimewire) predicted in August 2025 that 40% of enterprise applications would feature task-specific AI agents by the end of 2026, up from less than 5% in 2025, while warning that many AI assistants are being mislabeled as agents. [Zapier](https://zapier.com/blog/ai-agents-survey/?ref=runtimewire) and [Writer](https://writer.com/blog/enterprise-ai-adoption-2026/?ref=runtimewire) have also published enterprise surveys showing broad agent deployment or testing. Those adoption figures matter because they shape budgets. If a company says it has agents in production, security leaders need identity controls for non-human actors, engineering leaders need evaluations that hold up after deployment, finance needs spend ceilings, and infrastructure teams need workload-level cost accounting. If the "agent" is a single-prompt assistant with a human reading every output, the risk and cost profile is different. VentureBeat's survey pushes buyers to separate the label from the behavior. The operational question is whether software can plan and execute a multi-step task across systems with limited human intervention. That distinction decides whether a company needs an agent control plane or a better chatbot approval workflow. ### The budget opening is in control layers The survey is strongest when it moves past adoption language and into controls. VentureBeat says roughly six in 10 enterprises plan to switch or add vendors in each of five control layers over the next 12 months: identity for agents, evaluation of agent output, cost telemetry, context, and orchestration. That is the founder opportunity inside the mess. Enterprises are not waiting for a perfect reference architecture. They already deployed agents, found the gaps, and are now backfilling the missing control layers with budget attached. The risk profile is already visible. VentureBeat found 54% of companies had an agent security incident or near-miss caught before harm in the past 12 months. The evals data is sharper: 34% already allow an AI agent to push a code or system change to production based on automated evaluation results alone, with no human review, and another 33% are engineering pipelines to allow that within 12 months. Only 5% fully trust the automated evaluations that would make that decision. That is a procurement contradiction security and reliability founders can sell into. Enterprises are automating release authority faster than they trust the tests governing that authority. VentureBeat says half of enterprises shipped an agent that passed internal evaluations and then caused a customer-facing failure in the past year, and a quarter saw that happen more than once. The right product answer will not be a dashboard that says "AI governance" across the top. Buyers need controls close to the runtime: identity bound to specific agent actions, evals tied to real-world outcomes, spend caps per agent or workflow, context pipelines that track provenance, and orchestration software that can stop or reroute work before a bad action reaches production. ### The neocloud pitch now has to include utilization The survey gives neocloud providers a demand signal and a sales complication. CoreWeave, Lambda, Crusoe and Nebius can credibly argue that specialized AI clouds make more sense than undermanaged in-house GPU fleets for training, fine-tuning and inference. CoreWeave markets itself as an AI-native GPU cloud, Lambda says it offers cloud GPUs for training and deploying models and agents, Crusoe pitches an AI cloud spanning training and inference, and Nebius describes itself as an AI cloud company. VentureBeat's data says the next enterprise buyer will ask a more expensive question: will this provider improve utilization, or just move idle time from the company's data center to a cloud contract? That question cuts both ways. Specialized clouds can make scheduling, cluster management and access to newer hardware easier. They can also hide waste behind a cleaner invoice if the buyer lacks per-workload telemetry. The companies that win the next phase of enterprise AI infrastructure will have to sell utilization evidence, not capacity alone. VentureBeat plans to present the full research set at [VB Transform 2026](https://venturebeat.com/vbtransform2026?ref=runtimewire) on July 14th and 15th in Menlo Park, where the agenda centers on agentic orchestration, evals, infrastructure, context and security. The timing is useful. The enterprise AI market has moved past the easy proof-of-concept story. The hard part is turning demos into governed systems that can run for hours or days, spend real money, touch production systems, and still be accountable to someone. For now, the cleanest read from the survey is this: the AI buildout is not failing because enterprises lack appetite. It is running into the same operating discipline every infrastructure cycle eventually demands. Measure what is used, meter what it costs, control what can act, and buy more only after the existing fleet proves it deserves the next dollar.