For the last decade, the story of enterprise AI has sounded like a template: point at the biggest public cloud, call the APIs, rent the GPUs, and scale. Broadcom’s Private Cloud Outlook 2026 says that template is breaking. In last year’s study, 56% of enterprises used public cloud as the primary environment for production AI inference. This year, that number fell 15 percentage points to 41%. At the same time, 56% of enterprises are now running or planning to run production inferencing in a private cloud.

That “tipping point” swing is not a small optimization decision. The report frames it as infrastructure homecoming that is already showing up in production workloads, capital budgets, and board-level priorities, and it supports that claim with a second shift that matters more to builders: 43% of enterprises that are actively repatriating workloads are moving AI training, large language models, and inference out of public cloud, in a category that did not exist in last year’s study. Meanwhile, the broader repatriation trend accelerated sharply too, with 83% of enterprises considering repatriation, up from 69% in 2025, and half having already moved at least some workloads, a 15-point jump in a single year.

So why now? The report’s answer is that the forces pushing companies into private cloud for other “grown-up IT” reasons did not become more important because of AI. Those reasons were already there. Security, control, cost, and governance were already drivers for storage and regulated or security-sensitive applications. What AI changed was the cost of being wrong. When workloads run at production scale, security gaps, unclear accountability, and uncontrolled spend stop being paper cuts and start being recurring bills, audit headaches, and latency failures.

In practice, IT leaders say the workloads that land in private cloud are the same workload types that historically demand stricter handling: high-security, latency sensitive, business critical, or data-intensive environments. The report links that placement logic to the operational reality that AI inference and training are not “light touch” workloads. They are data-intensive. They are expensive. They are sensitive. And they create new governance obligations on top of existing regulatory demands.

Cost is the clearest blunt instrument in the report’s findings. For the first time, cost overtakes security as the top concern about public cloud. The numbers are unusually direct: 97% of IT leaders surveyed believe some portion of their public cloud spend is wasted, and more than half (52%) say that waste exceeds 25% of their total spending. On top of that, 62% of IT leaders are very or extremely concerned about AI infrastructure costs, with generative AI and agentic workloads compounding the pressure. Put differently: forecasting and managing public cloud spend was already hard, and AI makes the problem scale faster than enterprise budgeting cycles.

Enterprises are responding with capital planning that reads like a vote of no confidence in variable consumption economics. Net intent to increase private cloud investment over three years rose from 51% to 72%, and private cloud investment is now growing at more than twice the rate of public cloud. Cost predictability is also called out as the second biggest driver of the shift, cited by 39% of organizations. If you built AI ambitions on consumption-based pricing, the report suggests you are recalculating now.

But cost is only one leg of the stool. Sovereignty has moved into the infrastructure conversation with the force of a board agenda item. Eighty-six percent of IT leaders say geopolitical and regulatory factors now directly affect IT strategy and operations. Data sovereignty and residency requirements are the top concern at 54% of respondents, followed by jurisdiction-specific compliance requirements at 51%. For enterprises operating across borders, the report makes the link explicit: where data lives can determine where workloads can run. And for AI workloads that process sensitive, regulated, or proprietary data, you need governance and control “from the ground up,” not bolted on after deployment.

The report is also specific about how AI changes the compliance stack. Security and compliance remain the single most important factor in workload placement decisions, cited by 32% of respondents. Then AI adds new obligations: data protection and privacy at 37%, and security and control at 36% are named as the leading infrastructure requirements that AI imposes. The operational implication is that private cloud is not just where you host models. It is the governance architecture you use to meet those obligations by design.

There is a final friction point executives sometimes underestimate: running this stuff is a skills and operations fight, not only an infrastructure procurement exercise. The report lists the top skills gap as AI infrastructure and operations, cited by 40% of respondents. Next comes cloud security operations at 38% and Kubernetes operations at 37%. To close the gap, 81% of enterprises now fully outsource or use professional services for their cloud-related needs.

If you are an executive, that outsourcing rate is the second-order signal: even with the “right” platform, teams need help executing reliably at scale. The report argues that a platform-centric approach can reduce fragmentation and surface area by standardizing on a unified, well-governed private cloud platform. Translation: fewer specialists chasing edge cases, clearer organizational accountability, and less operational chaos.

The takeaway in the report is blunt: the tipping point is here, and private cloud is the preferred platform for production AI because it addresses what AI at scale demands: security, cost predictability, data sovereignty, and governance that enterprises cannot treat as optional. In this partner contribution, the named platform is VMware Cloud Foundation 9.1, positioned as a unified platform for running AI and traditional workloads together with performance, cost controls, and security capabilities for production AI at enterprise scale. For peers weighing budgets and platform strategy, the strategic stakes are clear: if your production AI remains anchored to variable hyperscaler inference economics without governance-by-design, you risk paying in unpredictable cost, compliance friction, and operational overload as the market’s default moves.