A VMblog Q&A with Prashanth Shenoy, VP of Product Marketing, VCF Division at Broadcom
For years, the default assumption was that enterprise AI would live in the hyperscaler cloud. The APIs were available, the GPU capacity was scaling up, and a decade of public cloud momentum pointed in one direction. Then the bills arrived — and the governance gaps became impossible to ignore.
Broadcom’s newly released Private Cloud Outlook 2026 report, based on a blind global survey of 1,800 senior IT leaders across eight countries, documents what may be the most dramatic single-year infrastructure shift in recent memory: production AI inference is moving decisively to private cloud. Public cloud’s share of production AI workloads dropped 15 percentage points in a single year, while private cloud now leads as the preferred environment for running AI at scale.
VMblog sat down with Prashanth Shenoy, Vice President of Product Marketing for the VCF Division at Broadcom, to unpack the findings, explore what’s driving the shift, and discuss how VMware Cloud Foundation 9.1 is built for this new reality.
VMblog: Broadcom just published the second edition of the Private Cloud Outlook report. What’s the headline finding that stood out most to you?
Prashanth Shenoy: The headline is that enterprise AI has found its infrastructure home — and it’s private cloud. Last year, 56% of enterprises were running production AI inference in the public cloud. This year, that number dropped to 41%. Meanwhile, 56% of enterprises are now running or planning to run production inferencing on private cloud. That’s a complete reversal in just 12 months, and it’s one of the most dramatic year-over-year swings we’re seeing in the second year of this research.
What makes it even more striking is that 43% of the enterprises that are actively repatriating workloads are specifically pulling AI training, large language models, and inference out of public cloud. That category didn’t even exist as a distinct data point in last year’s study. The AI conversation has moved from experimentation to production infrastructure decisions, and enterprises are voting with their budgets.
VMblog: Why is production AI landing on a private cloud rather than continuing to scale in the public cloud? Is this purely a cost story?
Prashanth Shenoy: Cost is a big part of it — and for the first time in this study, cost has actually overtaken security as the number one concern about public cloud. That’s a significant signal. But I wouldn’t reduce it to just cost. It’s really about three forces converging at once: cost, control, and complexity.
On cost, the data is striking. Ninety-seven percent of IT leaders believe some portion of their public cloud spend is wasted, and more than half say that waste exceeds 25% of their total public cloud budget. Now layer AI workloads on top of an already difficult-to-forecast cost environment, and 62% of IT leaders say they are very or extremely concerned about AI infrastructure costs. Enterprises that built AI ambitions on variable, consumption-based pricing are recalculating fast. Private cloud, with predictable economics and direct IT control over infrastructure, increasingly is where the budget decisions are landing.
Control – delivered through security and compliance – matters too, especially for the types of workloads that AI touches. High-stakes workloads like security-sensitive applications, latency-sensitive systems, business-critical processes, and data-intensive operations already had a preference for private cloud before AI came along. AI doesn’t change that calculus; it accelerates it. Security and compliance is the top driver for workload placement decisions, and continues to be the top reason IT organizations are repatriating from public to private cloud.
VMblog: Data sovereignty comes up prominently in the report. How much is geopolitics now shaping infrastructure decisions?
Prashanth Shenoy: More than most people outside of large enterprise IT organizations might expect. Four out of five IT leaders in our survey say geopolitical and regulatory factors are now directly affecting their IT strategy and operations. That is not a compliance checkbox issue anymore — it’s a board-level conversation.
For the first time this year, data sovereignty and residency requirements — cited by 54% of respondents — overtook jurisdiction-specific compliance at 51% as the leading geopolitical factor shaping infrastructure decisions. When you’re an enterprise operating across borders, decisions about where data lives carry direct implications for where workloads can run. And AI workloads are uniquely sensitive here, because they process sensitive, proprietary, and regulated data at a scale that multiplies the governance risk.
Industries like financial services, public sector, healthcare, and life sciences are at the leading edge of this shift. For these organizations, the combination of AI-driven data volumes, cross-border governance complexity, and the rising cost burden of public cloud is making an increasingly compelling case for private cloud infrastructure that puts data governance under organizational control from the start — not bolted on after the fact.
VMblog: The repatriation trend has clearly accelerated. What are you hearing from enterprises about what’s actually driving workloads back to the private cloud?
Shenoy: The numbers tell a clear story: 83% of enterprises are now considering repatriation, up from 69% in 2025, and 50% have already moved at least some workloads — a 15-point jump in a single year. That’s not a fringe trend anymore; it’s mainstream enterprise behavior.
The top three drivers are security and compliance at 51%, cost predictability at 39%, and performance at 39%. What’s most notable there is cost predictability. That wasn’t always in the top tier of repatriation drivers — its explosive rise year-over-year is one of the most significant changes in the whole report. It underscores just how severely public cloud economics have deteriorated in the AI era. When we see net intent to increase private cloud investment rising from 51% to 72% in a single year — and private cloud investment growing at more than twice the rate of public cloud — we know this isn’t a temporary correction. Enterprises are making a structural shift in where they’re putting their infrastructure dollars.
VMblog: The report also touches on the skills and complexity challenge. Running AI at enterprise scale isn’t just an infrastructure problem — it’s an operations problem too. What does the data say there?
Shenoy: That’s exactly right, and it’s something I think gets underweighted in a lot of the AI infrastructure conversation. The number one skills gap cited by IT leaders in our survey is AI infrastructure and operations, named by 40% of respondents — ahead of cloud security operations at 38% and Kubernetes operations at 37%. And 81% of enterprises now fully outsource or use professional services for cloud-related needs, at least in part.
The implication I draw from that is that operational simplification isn’t a nice-to-have — it’s a competitive necessity. Enterprises that have to manage fragmented, multi-vendor infrastructure stacks are burning their most scarce resource: skilled people’s time. A platform-centric approach that standardizes on a unified, well-governed private cloud environment addresses the AI skills challenge with fewer specialists, less operational fragmentation, and clearer accountability. You reduce the surface area that teams have to manage, and that’s where you get real operational leverage.
VMblog: Where does VMware Cloud Foundation 9.1 fit into all of this? How is it specifically built for the AI era you’re describing?
Shenoy: VCF 9.1 is our answer to exactly the environment the research describes. It’s a unified platform for running AI and traditional workloads together — you get the performance, cost controls, security capabilities, and governance architecture that production AI at enterprise scale actually demands.
The key word there is “unified.” One of the core enterprise pain points that drives up cost and complexity is infrastructure fragmentation — separate stacks for VMs and containers, piecemeal security tooling, different operational models across workload types. VCF eliminates that. You get virtualization, CNCF-conformant Kubernetes, built-in security, and automation in one integrated platform. For AI workloads specifically, that means you can run inference and training alongside your existing business-critical workloads without introducing new management complexity or governance gaps.
We’ve also built the platform with the sovereignty and compliance requirements in mind — private cloud governance from the ground up, not retrofitted. For the enterprises in regulated industries or operating across multiple jurisdictions, that architecture matters enormously.
VMblog: Looking ahead, what should enterprises be doing right now to position themselves well for production AI on private cloud?
Shenoy: The message from the research is that the window for getting this right is now. The AI experimentation phase is over for most large enterprises. Production workloads are being placed today, capital budgets are being reallocated today, and the infrastructure decisions getting made now will govern how well organizations can compete with AI over the next several years.
The enterprises I see succeeding are the ones that approach this as a platform decision rather than a point solution problem. Don’t try to build AI infrastructure by stitching together a dozen different tools and hoping the seams hold at scale. Start with a platform that gives you a consistent operational model for both traditional and AI workloads, security that’s built in rather than bolted on, and cost visibility and control from day one.
The question for enterprise IT leaders is no longer whether private cloud is the right home for production AI — the data is clear on that. The question now is how fast you can build the foundation to get there.
Editorial Note: The Private Cloud Outlook 2026 is based on a global survey of 1,800 senior IT decision-makers at enterprise organizations with 1,000 or more employees, across eight countries in North America, Europe, and Asia-Pacific. The survey was conducted in February–March 2026 by Radius Tech in partnership with Broadcom.
