Enterprise technology teams spent years testing AI in isolated experiments. Now the invoices have landed. And the infrastructure choices look nothing like the original plans.
Broadcom’s survey of 1,800 senior IT leaders across eight countries shows a decisive turn. Last year 56% ran production AI inference primarily in public cloud. This year that share dropped to 41%. Private cloud now claims 56% for production or planned production workloads. The shift marks what the report calls the AI tipping point. (Broadcom News, June 9, 2026)
Prashanth Shenoy, vice president of product marketing for the VCF division at Broadcom, put it plainly. “Enterprise AI has found its infrastructure home. And it is private cloud.” The data backs him. Nearly half the organizations actively repatriating workloads are pulling AI training, large language models, and inference back from public providers. This category barely registered in last year’s study.
Costs explain much of the reversal. For the first time cost overtook security as the top worry about public cloud. Ninety-seven percent of respondents believe some of their public cloud spend is wasted. More than half put that waste above 25% of total spending. Sixty-two percent express very or extremely high concern about AI infrastructure costs specifically. Variable pricing that once seemed flexible now feels unpredictable at the volumes required for serious inference and agentic systems.
But money tells only part of the story. Data sovereignty, regulatory pressure, and governance needs have moved from compliance checklists to boardroom mandates. Eighty-six percent of IT leaders say geopolitical and regulatory factors now shape their technology decisions directly. Fifty-four percent flag data sovereignty and residency as their foremost concern, followed closely by jurisdiction-specific rules.
These pressures hit harder with AI. Models trained on proprietary data or handling regulated customer information cannot tolerate the shared responsibility models common in hyperscale environments. Private clouds deliver control from the foundation. Security and compliance rank as the single most important factor in workload placement for 32% of organizations. Requirements around data protection and privacy (37%) and security and control (36%) top the list of new demands AI places on infrastructure.
The Infrastructure Reckoning
Repatriation momentum has accelerated fast. Eighty-three percent of enterprises now consider moving workloads back from public cloud, up from 69% the prior year. Half have already done so, a 15-percentage-point increase. For AI workloads the pattern holds stronger. Forty-three percent target training, models, and inference for the move. High-security, latency-sensitive, business-critical, and data-intensive applications consistently land in private environments.
Investment plans reflect this reality. Net intent to grow private cloud spending over the next three years jumped from 51% to 72%. Private cloud budgets are expanding at more than twice the pace of public cloud outlays. Cost predictability drives 39% of these decisions. Performance and compliance follow closely.
The Register covered these findings in detail, noting that the consequences of weak governance or runaway bills become far more painful once AI reaches production scale. (The Register, June 18, 2026)
Skills shortages compound the operational load. Forty percent of leaders cite AI infrastructure and operations as their biggest talent gap. Cloud security operations (38%) and Kubernetes management (37%) follow. Eighty-one percent now outsource or bring in professional services for cloud needs. A unified private cloud platform reduces the number of specialists required and limits operational sprawl. That matters when every new agentic workflow adds complexity.
Deloitte’s separate 2026 State of AI in the Enterprise report, based on more than 3,200 global leaders, shows parallel movement from pilots toward scale. Worker access to AI tools rose 50% in 2025. The share of organizations expecting at least 40% of AI projects in production is projected to double within six months. Twice as many executives now report transformative business impact compared with the previous survey. Yet only about one-third say they are truly reimagining operations. Most still apply AI to surface-level tasks or process tweaks. (Deloitte)
Agentic systems add another layer. Twenty-three percent of organizations use them at moderate scale today. That number is expected to climb sharply. Governance, however, lags. Only one in five companies claims mature oversight for autonomous agents. The gap between experimentation and reliable production remains wide even as ambition grows.
Recent coverage reinforces the pattern. A March 2026 NVIDIA blog post on its own state of AI findings noted enterprises shifting from assessment phases to scaled deployments across industries. Productivity and efficiency gains appear in 66% of Deloitte respondents. Insights and decision-making improve for 53%. Cost reduction shows for 40%. But measurable financial returns still prove elusive for many. One industry analysis put the portion of CFOs seeing clear ROI at just 14% despite trillions in collective spending. (NVIDIA Blog, March 2026)
Stanford researchers documented similar dynamics in their Enterprise AI Playbook. While 88% of organizations deploy AI somewhere, only one-third have scaled programs enterprise-wide. Pilot failure rates for generative AI once approached 95% according to earlier MIT work cited in the playbook. Success now hinges on structured workflows, clear human oversight levels, and agentic frameworks that handle multi-step tasks with error recovery.
Yet the private cloud preference documented by Broadcom and VMware points to a practical path. VMware Cloud Foundation appears repeatedly in the report as a platform designed to run AI and traditional workloads together with built-in cost controls, security, and density optimization for GPUs. The ability to pool resources, enforce consistent policies, and maintain sovereignty without sacrificing performance addresses the exact pain points executives cite.
Geopolitics will keep tightening these choices. Companies operating across borders face conflicting data rules and rising scrutiny on where models train and infer. Private clouds let them keep sensitive workloads inside approved jurisdictions. That flexibility carries real weight when 86% already report regulatory factors altering strategy.
So the narrative has flipped. Early AI strategies assumed public cloud dominance because of easy access and rapid scaling. Production demands changed the math. Predictable costs. Stronger governance. Direct control over data and infrastructure. These factors now outweigh the convenience of APIs and seemingly unlimited capacity.
Leaders who treat AI infrastructure as an afterthought risk watching budgets balloon while value stays trapped in pilots. Those who standardize on platforms built for mixed workloads and sovereign requirements stand to pull ahead. The data from 2026 shows the split is already underway.
Private cloud did not win by default. It earned the role because AI at true enterprise scale exposes every weakness in cost models, security postures, and operational maturity. The tipping point isn’t coming. The numbers say it has arrived.
WebProNews is an iEntry Publication