Enterprise AI Survey: 86% of Companies Running GPUs at Half Capacity While Agents Already Caused Security Incidents

Since prior coverage this week focused on labor markets and geopolitical risks, a large-scale enterprise AI survey published by VentureBeat Research in June 2026 adds a concrete data layer to the broader AI buildout debate.

The survey covered 573 technical leaders at companies with 100 or more employees, fielded across five parallel surveys examining what VentureBeat calls the "agentic stack" — the identity, evaluation, cost, context, and orchestration layers that govern how AI agents operate inside businesses.

The Hardware Problem Wall Street Missed

The AI infrastructure buildout has been a dominant market narrative for two years. The enterprise data cuts against the bull case, at least in the short term.

According to VentureBeat Research, 86% of enterprises running their own GPUs report utilization of 50% or less. That is not a short-seller's projection. That is the companies that bought the hardware saying they are running it at half capacity.

The measurement problem compounds it. Only 44% of these enterprises rigorously track what their AI compute actually costs and returns. The other 56% are estimating, which means half the industry is flying blind on the single most expensive line item in their AI budgets.

Despite the idle hardware, purchasing appetite has not stopped. According to VentureBeat, 45% of surveyed enterprises say the compute option they are most likely to evaluate in the next 12 months is an AI-specialized cloud — providers like CoreWeave, Lambda, Crusoe, or Nebius. Under 2% are actually using one of those providers today.

Roughly 32% named non-Nvidia accelerators — AWS Trainium, Google TPUs, AMD — as their most likely evaluation target, compared to 28% who named next-generation Nvidia GPUs. That is a meaningful hedge signal, even if it has not yet translated into purchasing decisions.

The Security and Governance Gap

Enterprises did not accidentally deploy agents without adequate controls. They did it knowingly, according to VentureBeat, and are now retrofitting.

54% of companies experienced an agent security incident or near-miss caught before harm in the past 12 months. 27% exercise only reactive control of agent spend, meaning they learn what an agent cost when the invoice arrives, with no per-agent budget or ceiling in place.

Roughly six in ten enterprises plan to switch or add vendors across all five control layers within the next 12 months. About a third plan to move within the current quarter, depending on the layer.

The Evaluation Gap: Agents Passing Tests, Failing in Production

A separate June 2026 VentureBeat Pulse survey of 157 qualified enterprise respondents — smaller sample, directional rather than precise by the publication's own disclosure — drills into a specific failure mode.

Half of enterprises have deployed an AI agent or LLM feature that passed internal evaluations and still caused a customer-facing failure. One in four experienced that outcome more than once.

The rational response would be to slow automation. Enterprises are doing the opposite. 66% of respondents already permit some production deployment without human review, or are building systems intended to do so within the next 12 months. Only 5% say they fully trust the automated evaluations driving those release decisions.

VentureBeat describes this as the "evaluation gap": the autonomy ceiling is rising faster than the assurance beneath it.

The core technical problem is that traditional software testing checks whether a defined input produces an expected output. Agent testing does not work that way. An agent chooses its own sequence of steps, calls tools, retrieves data, and can produce different results from one run to the next. It can make five individually correct decisions and still reach the wrong outcome, updating the wrong database field, sending a refund request without approval, or leaking sensitive information on step six of a six-step workflow.

Enterprises recognize this. The most common reason for distrusting automated evaluation, cited by 29% of respondents, is poor alignment with real-world outcomes — not speed or cost. Bias or inconsistency follows at 21%, lack of explainability at 18%, and data leakage or privacy concerns at 17%.

NIST made a similar point in its Generative AI Profile, noting that measurements from controlled environments may not transfer to deployment because agent behavior shifts with prompts, users, context, and operating conditions. NIST's guidance calls for field testing, post-deployment monitoring, and clear escalation processes for failures.

The Strongest Counterargument

The case for optimism is real and worth stating plainly. Enterprises deploying ahead of controls is not a new phenomenon — it happened with cloud adoption, mobile device management, and SaaS sprawl. In each case, a retrofit cycle followed, vendors built governance tooling, and the technology matured. Roughly six in ten enterprises plan to switch or add vendors across all five control layers within the next 12 months, suggesting the market is self-correcting rather than stalled. Low GPU utilization could also reflect early-stage learning curves and workload consolidation rather than permanent overcapacity.

That context does not make the current exposure disappear. A 54% security incident rate and a 27% reactive-spend rate are problems right now, not hypothetically.

What Happens Next

VentureBeat is hosting VB Transform 2026 on July 14-15 in Menlo Park, where evaluation methodology — specifically what Amazon and Waymo test instead of standard benchmarks — is on the agenda. The open question the survey does not answer is whether the retrofit cycle enterprises are budgeting for will close the evaluation gap before the next wave of agent autonomy expansion makes it materially larger.