Half-Empty GPU Utilization Rattles Wall Street's AI Bet

Enterprise GPU utilization is the AI buildout’s ugliest tell: 86% of enterprises running their own GPUs say those chips operate at 50% capacity or less, putting CIOs, CFOs, and infrastructure vendors on the defensive at the same time.

That figure comes from VentureBeat Research’s June survey of 573 technical leaders at companies with 100 or more employees, according to VentureBeat. The survey lands directly in the middle of Wall Street’s AI capex argument. The sharper read is not that enterprise AI demand has vanished. It’s that many companies bought or reserved compute before they had the controls to make it productive, measurable, and safe.

Wall Street's AI capex fight just ran into half-empty enterprise GPU utilization

The survey supports a clean thesis: the next phase of enterprise AI spending will be judged less by who controls the most compute, and more by who can prove workload-level utilization, cost per agent, output quality, and security control.

86% of enterprises that run their own GPUs report utilization of 50% or less.

That is a demand-quality problem. Enterprises are still evaluating more AI infrastructure. 45% said the emerging compute option they are most likely to evaluate in the next 12 months is an AI-specialized cloud such as CoreWeave, Lambda, Crusoe, or Nebius. Yet under 2% report using one of those neoclouds today.

The contradiction matters. If the existing GPU estate is half-used or worse, what exactly is the next contract meant to solve?

XOOMAR analysis: This does not prove a collapse in AI infrastructure spending. The survey does not say enterprises will cut compute budgets. It does show a harder procurement environment, where new GPU, accelerator, and cloud deals will need to survive questions that were easier to ignore during the scramble for capacity.

Builders face a control debt bill after shipping agents first

VentureBeat’s central finding is that enterprises deployed AI agents ahead of the control systems needed to manage them, and they did so knowingly.

The five control layers in the survey are:

The bill is already showing up. 54% of companies reported an agent security incident or near-miss caught before harm in the past 12 months. 27% manage agent spend reactively, meaning they learn what an agent costs when the invoice arrives, with no per-agent budget or ceiling.

Roughly six in 10 enterprises plan to switch or add vendors in each of the five control layers within the next 12 months. Across specific layers, 57% to 64% plan vendor movement, including 64% in infrastructure and evaluations, 59% in agent security, and 57% in retrieval and context.

The builder question is blunt: did teams ship agents because the workflow was ready, or because the dashboard needed an AI win?

XOOMAR analysis: The survey points to a control-debt cycle. Enterprises shipped first, then started budgeting for the identity, evaluation, telemetry, context, and orchestration layers that should have bounded production risk from the start. That is not the same as saying raw compute spending is over. It means control software now has a stronger claim on incremental AI budget.

Buyers are discovering that most AI agents are still chatbots with bigger invoices

The agent adoption story looks thinner when enterprises are asked what their deployed systems actually do.

71% of enterprises said a quarter or fewer of their deployed “agents” can complete multi-step tasks on their own. Only 10% said true agents are the majority of what they run. The respondents were not casual observers: 81% said they recommend or decide AI purchases at their companies.

That cuts against the way agent adoption is often marketed. VentureBeat cites Gartner’s prediction that 40% of enterprise applications will be integrated with task-specific AI agents by the end of 2026, up from less than 5% in 2025. Gartner also warned about “agentwashing,” the practice of calling AI assistants agents when they are not actually autonomous. VentureBeat also notes that Zapier’s enterprise survey found 72% reported deploying or testing autonomous agents, while Writer’s 2026 survey had 97% of executives saying their company deployed AI agents in the past year.

Definitions now drive budgets. A single-prompt chatbot with a human reading every answer does not need the same identity, evaluation, cost, and orchestration stack as an agent that can execute multi-step work.

The buyer question is simple: are you funding autonomy, or are you funding a chatbot label?

This is where internal pressure can distort procurement. Boards see high agent adoption figures. Technical leaders then get pushed to match the headline, even when the real production bar is lower and the control work is slower. For related XOOMAR context on the cost side of agent rollout, see Enterprise AI Agents Turn Safe Pilots Into Cost Traps.

Security teams inherit automated evals, shared credentials, and bad context

The most alarming finding is not just that controls are incomplete. It’s that enterprises are already moving toward more autonomy despite distrusting the mechanisms that approve it.

34% already allow an AI agent to push a code or system change to production based only on automated evaluation results, with no human review. Another 33% are engineering pipelines to allow that within the next 12 months. Only 5% fully trust those automated evaluations.

The failure data explains why. Half of enterprises shipped an agent that passed internal evaluations and later caused a customer-facing failure in the past year. A quarter saw that happen more than once. When asked for the biggest weakness in current evaluations, 29% chose “poor alignment with real-world outcomes.”

Production monitoring is also thin. Only 23% run real-time quality checks on live agent answers. Another 51% monitor system health only, such as uptime, request traces, and gateway logs. That tells teams the agent is running. It does not tell them whether the answer is right.

The security question is harder: can you stop a bad agent fast enough if you cannot clearly identify it?

Credential sharing makes that question concrete. 69% of companies allow agent credential sharing somewhere in the fleet, meaning multiple agents operate under one API key or service account. Those companies reported a security incident or near-miss at a 63.5% rate, compared with 40.9% where every agent has its own scoped identity.

Bad context adds another failure path. 57% of enterprises traced at least one confident, wrong agent answer in the past six months to missing or inconsistent business context, such as wrong metrics, stale definitions, or absent documents. 25% already run a governed semantic layer in production, 34% are building one, and 41% have not started.

Vendors are chasing the same control budget as compute choices widen

The vendor fight is not only about models or GPUs. It is about who owns the control plane around enterprise agents.

Hyperscalers and model providers have the convenience advantage. VentureBeat says most enterprises default today to built-in tools from Anthropic, OpenAI, Google, Microsoft, and AWS across guardrails, evaluations, and retrieval. Provider-native or hyperscaler controls are the primary agent security layer for 82% of respondents.

Specialists see the opening. No layer has an established incumbent. In evaluations, the most common tooling is tied between the model provider’s built-in evals and no dedicated tooling at all, at 17% each.

Compute choices are also widening. 32% of enterprises named non-Nvidia accelerators, including AWS Trainium, Google TPUs, and AMD, as the emerging compute option they are most likely to evaluate in the next 12 months. 28% named next-generation Nvidia GPUs.

The vendor question is now: can bundled tools keep accounts locked in when buyers want portability?

VentureBeat’s orchestration data shows that concern rising. In the spring survey wave, the top concern about provider-controlled orchestration was security and permissioning limits at 32%. By June, vendor lock-in led at roughly a third, with security limits at 28%. For adjacent XOOMAR coverage on model choice and agent portability, see Model Lock-In Cracks as Vercel AI Agents Pick Labs.

Enterprise buyers should measure before they sign the next GPU or agent deal

The practical lesson is not “stop buying AI.” It is “measure before expanding.”

Enterprises need to know:

• GPU utilization: Which workloads are consuming capacity, and which chips sit idle.
• Per-agent cost: What each agent costs before the invoice arrives.
• Budget ceilings: Where spend automatically stops or escalates.
• Scoped identity: Which agent did what, with what permission.
• Live answer quality: Whether production outputs are correct, not just whether systems are online.
• Business context control: Whether agents read from governed definitions.

The procurement question should be direct: will this contract expose operational weakness, or hide it behind another tool?

A sensible priority order follows from the survey. Start with scoped identities for agents touching production systems. Then validate evaluations against production outcomes, not internal benchmarks alone. After that, instrument answer quality in real time. Finally, govern the metrics, entities, and definitions that agents use to answer business questions.

Buying orchestration, retrieval, or neocloud capacity as a shortcut risks repeating the same mistake. Without telemetry and quality controls, more tooling can create a cleaner-looking version of the same unmanaged agent fleet.

Hybrid control planes become the 2026 test for agent productivity

By the end of 2026, 51% of enterprises expect their primary control plane for enterprise agents to be hybrid, combining provider-native tools with external orchestration. That is up from 34% in the spring survey wave. Pure reliance on provider-managed agent services fell from 12% to 7%.

That is the forward signal to watch. If VentureBeat’s Q3 survey shows higher GPU utilization, more scoped identities, production-tested evals, and shipped semantic layers, the thesis strengthens: enterprises are turning agent enthusiasm into governed infrastructure.

If those numbers do not improve, the AI buildout debate gets harsher. Half-empty GPUs are survivable when workloads are maturing. They are harder to defend when agents remain mostly chatbots, costs arrive after the fact, and production controls lag autonomy.

The Bottom Line

• The survey challenges the idea that more AI infrastructure automatically means more productive AI usage.
• Vendors may face tougher scrutiny as buyers demand proof of utilization, cost efficiency, quality, and security.
• Enterprise interest in specialized AI clouds remains high, but current adoption is still minimal.