For most of the last decade, the enterprise default was simple: send it to the public cloud. In 2026, that default has quietly broken. CFOs are staring at cloud bills that grew faster than the workloads behind them, and the arrival of production AI has turned a manageable line item into a board-level problem. Training and inference on rented GPUs—billed by the hour, marked up by the hyperscaler, and surrounded by egress fees—have exposed the uncomfortable truth that the public cloud was never engineered to be the cheapest place to run steady-state, high-utilization infrastructure.
The result is a structural shift that analysts and vendors are now calling private cloud repatriation: the deliberate movement of predictable, high-volume, and AI-intensive workloads back into infrastructure the enterprise owns and controls. And the platform sitting at the center of that shift is VMware Cloud Foundation. With the general availability of VCF 9.1, Broadcom has reframed the conversation entirely—from “how do we cut our cloud bill” to “how do we run a public-cloud-grade private cloud at a fraction of the cost.” This article breaks down why VCF 9.1 has become the reference architecture for repatriation, what changed under the hood, and how to build the financial case for the move.
The 2026 Inflection Point: Why Public Cloud Economics Broke for AI Workloads
The repatriation trend is not anti-cloud sentiment. It is arithmetic. Bursty, unpredictable workloads still belong in the public cloud, where elasticity is worth the premium. But the workloads that now dominate enterprise compute—databases, virtual desktops, and especially production AI inference—are the opposite of bursty. They run continuously, at high utilization, on infrastructure that benefits enormously from being owned rather than rented.
Broadcom’s preview of its Private Cloud Outlook 2026 report quantifies the swing. More than half of organizations surveyed (56%) are running or planning to run production inferencing in a private cloud, while public cloud use for production inference was 41%, down 15% year over year. That is a remarkable reversal in a single year, and it tracks directly with the cost of GPU capacity. When an enterprise is committed to running inference 24/7, paying hyperscaler hourly rates plus egress to move data in and out becomes indefensible against the alternative of owning the silicon.
Three forces are compounding the pressure in 2026:
- AI infrastructure sticker shock. GPU instances are the most expensive line items in any cloud bill, and inference is a permanent, not temporary, cost.
- Data sovereignty and compliance. Regulated industries increasingly cannot place model training data or inference pipelines in multi-tenant public environments.
- Hardware scarcity economics. During a period of constrained GPU supply, squeezing more workload density out of owned hardware is more valuable than ever.
What Private Cloud Repatriation Means—and Why VMware Cloud Foundation Leads It
Repatriation only works if the destination is genuinely better than what enterprises fled in the first place. Lifting workloads out of AWS or Azure only to land them on brittle, manually-operated legacy virtualization is not a strategy. This is precisely the gap VCF is built to close.
VMware Cloud Foundation is a private cloud platform that bundles compute (vSphere), software-defined storage (vSAN), networking and micro-segmentation (NSX), and management tooling into a single, integrated stack delivered as one subscription. It delivers a unified private cloud platform that combines the flexibility of public cloud environments with the security, compliance, and control of your own data centers. The strategic positioning of the 9.x line is explicit: “VCF 9.0 was the architectural shift, VCF 9.1 is the optimization layer.” In other words, 9.0 rebuilt the foundation; 9.1 is the release that makes the economics work at enterprise scale.
Inside VMware Cloud Foundation 9.1: The Architectural Shifts That Rewrite the TCO Equation
VCF 9.1 is not a cosmetic point release. It introduces four architectural changes that, taken together, materially lower the cost and risk of running a private cloud.
Intelligent NVMe Memory Tiering: Up to 42% Lower Server Cost
Memory is one of the most expensive components in any server, and AI and database workloads are notoriously memory-hungry. VCF 9.1’s Enhanced NVMe Memory Tiering attacks this directly. The hypervisor tiers active pages to DRAM and cold pages to local NVMe, expanding effective memory capacity without additional DRAM. Critically for production architects, 9.1 adds software-based mirroring for tiered memory—closing the availability gap that made earlier tiering risky for stateful workloads.
The financial impact is the headline. Broadcom reports potential server cost reduction of up to 42% and storage TCO reduction of up to 39% when combining NVMe tiering with vSAN ESA deduplication. For an enterprise repatriating workloads, this means running more VMs on the same physical footprint, deferring expensive hardware refreshes during a period of constrained supply. VCF Operations now quantifies the benefit before you commit, with memory tiering recommendations for NVMe clusters that quantify cost savings and VM density improvements, plus What-If Analysis support for NVMe memory tiering.
VCF Private AI Services: NVIDIA Blackwell and Intel Xeon 6 Acceleration
This is the feature that makes repatriation viable for AI, and it is the strategic heart of the release. VCF 9.1 delivers an AI and Kubernetes native private cloud platform with integrated security and mixed compute infrastructure support across AMD, Intel, and NVIDIA, enabling enterprises to deploy inference and agentic AI applications with significantly lower costs and the freedom to choose best-of-breed GPU and CPU hardware.
On the GPU side, VCF Private AI Services brings the latest NVIDIA silicon into the platform. VCF now supports the latest NVIDIA Blackwell series GPUs, including the RTX PRO 6000 Blackwell Server Edition, NVIDIA HGX B200, and RTX PRO 4500 Blackwell Server Edition, as well as the NVIDIA HGX platform with Blackwell GPUs and NVLink Switch for massive-scale AI deployments. High-speed connectivity is handled through NVIDIA ConnectX-7 NICs and BlueField-3 with Enhanced DirectPath I/O, with support for GPUDirect RDMA—essential for distributed training and data-transfer-heavy pipelines.
Just as important is the CPU story, because not every inference workload needs a GPU. VCF 9.1 is further optimized for Intel Xeon 6 processors, with native integration of Intel QuickAssist Technology (QAT) that accelerates Encrypted vMotion while freeing valuable compute resources. Offloading cryptography to dedicated silicon removes the historical performance penalty on end-to-end encryption, returning cores to the workloads that pay the bills. The platform even supports CPU-based inferencing using Llama.cpp and Model Context Protocol (MCP) support with governance, letting teams connect AI assistants to internal systems like ServiceNow, GitHub, and SQL Server without building custom connectors.
Unified Operations: One Control Plane Replaces the Aria/SDDC Silos
Operational complexity is a hidden tax on every private cloud. Historically, VCF customers juggled separate Aria and SDDC Manager tooling, each with its own interface and lifecycle. VCF 9.1 collapses that fragmentation. The release adds VCF management services—a common runtime and set of components that unifies the architecture of lifecycle and operational capabilities. Administrators now work through a redesigned VCF Operations interface, a single command center to build, protect, and operate the private cloud.
This unification extends to existing estates. Repatriating teams do not have to rebuild from scratch: using VCF Operations, organizations running vSphere 8.0 Update 3 or higher can integrate their existing vCenter as a workload domain in VCF 9.1, maintaining business continuity without downtime or data migration. The result is a genuinely consolidated control plane—the operational simplicity enterprises expected from public cloud, delivered on owned infrastructure.
In-Platform Ransomware Recovery and Continuous Compliance
Security is no longer a bolt-on. In VCF 9.1, resilience is treated as an architectural property: vSAN for Recovery and On-prem Ransomware Recovery provide a sovereign, in-platform path to recover from destructive attacks without dependence on external escrow. For repatriating regulated workloads, the compliance posture is equally significant. Continuous Compliance Enforcement transforms compliance from a quarterly audit exercise into a runtime guarantee, Live Patching for TPM-Enabled Hosts eliminates a major source of unplanned downtime, and a CrowdStrike EDR integration adds endpoint detection to the recovery workflow.
Massive Scale: Up to 5,000 ESXi Hosts Per Instance
Repatriation at enterprise scale requires a control plane that can govern thousands of hosts without proliferating management silos. VCF 9.1 doubles down—literally. Expanded management scale allows a single instance to support up to 5,000 ESXi hosts, which is up 2X from the previous release.
The scale story extends to modern, containerized, and AI workloads:
- 500 Kubernetes clusters per Supervisor, with the re-engineered supervisor architecture increasing parallel upgrade throughput from 64 to 256 clusters simultaneously.
- Up to 46% lower Kubernetes operational costs at scale, with 4x faster cluster upgrades and 2x increased fleet capacity to rapidly scale AI infrastructure.
- Up to 70% faster provisioning via linked-clone technology, now extended to standard VMs through VM Fast-Deploy.
The combination of higher density per host and a single instance governing far larger estates is what makes the unit economics of a private cloud platform decisively beat rented public capacity for steady-state workloads.
Understanding VMware VCF 9.1 Licensing: The Subscription Math CIOs Must Model
No private cloud business case is honest without confronting the commercial model head-on. VMware VCF 9.1 licensing is subscription-only and per-core. The per-CPU and perpetual options are gone; organizations must license all physical CPU cores in scope, with two mandatory minimums shaping the process: 16 cores per CPU and a 72-core minimum per license instance. Every physical core on any host running the hypervisor must be in scope—there is no longer an exemption for unlicensed dev or test capacity.
The bundle itself is comprehensive. NSX, vSAN, and Aria are not separately sellable in 2026—they live inside VCF, which includes 1 TiB of vSAN capacity per licensed core. List pricing references put VCF at roughly $350 per core per year, with the 16-core-per-CPU minimum raising the effective starting point to about $5,600 per socket per year, though realized post-discount pricing at enterprise scale typically lands between $185 and $275 per core.
The strategic takeaways for infrastructure leaders are clear:
- First quotes are negotiable. Advisory data shows opening quotes of 2x–5x prior cost frequently settling at 1.3x–2x with an accurate core audit and a credible alternative on the table.
- Density is now a licensing lever, not just a performance one. Because you pay per core, the 9.1 efficiency gains—NVMe tiering, deduplication, higher VM density—directly reduce the number of cores you need to license. The technical and commercial cases reinforce each other.
- Consolidate before you renew. Modeling workloads onto fewer, denser hosts can recover a meaningful share of the subscription increase before any discount.
Building the Business Case for Private Cloud Repatriation
When you assemble the pieces, the financial logic of moving steady-state and AI workloads onto VCF 9.1 becomes hard to ignore. You eliminate hourly GPU markups and egress fees. You drive up to 42% lower server cost and up to 39% lower storage TCO through tiering and deduplication. You retire separate management tooling into a single control plane that scales to 5,000 hosts. And you keep sovereign control over AI data and recovery. The per-core subscription is a real and rising cost—but it is a predictable one, and the 9.1 efficiency gains directly offset it by letting you license fewer cores for the same workload.
For workloads that are predictable, high-utilization, regulated, or AI-intensive, the repatriation math in 2026 increasingly favors a private cloud platform you own over capacity you rent.
Your Next Step: Evaluate Your vSphere and vSAN Estate for VCF 9.1 Conversion
The enterprises winning the repatriation shift are not the ones that waited for a contract renewal to force the conversation. They are the ones that modeled it early. Your next move should be concrete:
- Run a full core and capacity audit. Inventory every physical core, CPU, and host in scope, then calculate billable cores including the per-CPU and per-instance minimums—this is the single biggest variable in your VCF 9.1 cost.
- Model consolidation onto denser hosts. Quantify how NVMe memory tiering and vSAN deduplication reduce your required core count, and run VCF Operations’ What-If Analysis to project the density and TCO gains on your actual workloads.
- Identify repatriation candidates. Map your steady-state, high-utilization, and AI inference workloads currently in public cloud—these are where the economics swing hardest in favor of private cloud.
- Prepare the commercial conversion. If you are on vSphere 8.0 Update 3 or later, plan the non-disruptive path to integrate existing vCenters as VCF workload domains, and open renewal or subscription conversations nine to twelve months ahead with the audit data in hand.
The 2026 cloud cost crisis is not a reason to panic—it is a reason to plan. VMware Cloud Foundation 9.1 gives enterprise infrastructure leaders the platform to turn repatriation from a cost-cutting reaction into a long-term strategic advantage. Start the evaluation now, while you still control the timeline.