86% of enterprises run AI agents at half GPU capacity or below

The survey reveals a fundamental mismatch in enterprise AI deployment: companies prioritized speed-to-production over governance infrastructure. The fact that 86% of enterprises run GPUs at half capacity or below suggests not only underutilized hardware investment, but also that the rush to deploy agents left room for only partial workloads while management systems were retrofitted. This pattern—where enterprises knowingly deployed without controls in place—points to a broader industry dynamic in which proof-of-concept and competitive pressure to ship AI features outweighed operational discipline.

The remediation timeline the research uncovered is telling. That roughly 6 in 10 enterprises plan to change vendors within a 12-month window across all five control layers indicates not a one-time tuning problem, but a systematic rebuilding effort. The fact that roughly a third plan to move within the quarter suggests some enterprises face urgent governance gaps. This workload, spread across identity, evaluation, cost tracking, context, and orchestration, suggests enterprises are not addressing a single pain point, but rather building out an entire governance architecture they should have designed before deployment. The budgeting commitment implied by vendor-switching plans signals that enterprises now recognize the cost and risk of this approach—and are willing to pay to correct it.