On May 5, 2026, Broadcom announced VMware Cloud Foundation 9.1 — the latest major release of the platform that anchors private clouds at hundreds of the world’s largest enterprises, including most of our own customers. Reading the press materials, it would be easy to file this as another point release. It is not.
VCF 9.1 is a deliberate response to the three forces that are reshaping every infrastructure roadmap right now: production AI is moving back on-premises faster than the public cloud playbook predicted, hardware budgets are tighter than they have been in a decade, and the security perimeter has officially shifted from the network edge into individual Kubernetes pods. This release is engineered around all three of those problems at once, and the resulting feature list is the most consequential we’ve seen since the move from VCF 4 to VCF 5.
Here’s what’s actually in it, what it changes for the teams who run VMware infrastructure, and where the genuine business value lands.
The thesis: efficiency, AI delivery, resilience
Broadcom organized the 9.1 announcement around three outcomes — higher infrastructure efficiency, faster AI and application delivery, and stronger cyber resilience. That framing isn’t marketing. Each pillar maps to specific engineering work in the release, and each one targets a measurable problem most enterprise IT teams are dealing with today.
The market context Broadcom cited makes the timing clear. Their Private Cloud Outlook 2026 report shows 56% of organizations are now running or planning production inferencing in private cloud, while public cloud use for production inference dropped 15% year over year. At the same time, 62% of IT leaders surveyed reported being very or extremely concerned about generative AI infrastructure costs. AI is coming home, and it has to be cheaper to operate when it gets there.
VCF 9.1 is built for that moment.
The headline numbers — and what they actually mean
Broadcom’s official efficiency claims for 9.1 are:
- Up to 40% reduction in server costs through intelligent memory tiering on clusters running mixed AI and non-AI workloads
- Up to 39% lower storage TCO through enhanced compression and deduplication for AI data pipelines
- Up to 46% reduction in Kubernetes operational costs for AI workloads at scale
- 4× faster cluster upgrades, with parallel upgrade capacity now supporting up to 256 clusters simultaneously
- 9 Tbps threat inspection performance for distributed inference workloads
These are vendor-published “up to” numbers — they reflect best-case conditions, not promises. But the underlying engineering is real, and even at half of the claimed values, the unit economics of running mixed VM-plus-AI estates on existing hardware change materially. The right way to read these numbers is not as a guarantee, but as a signal of where the platform’s investment is concentrated. When Broadcom claims 40% server-cost reduction, they’re telling you the architecture changes are aimed at extracting more workload density from servers you already own.
That matters because the alternative — building more infrastructure to keep up — has gotten genuinely expensive. GPU server lead times are still measured in months. Power and cooling costs in major metros have outpaced inflation for three straight years. Software-defined density wins right now have a harder dollar value than they did at any prior point in the platform’s history.
Memory and storage: the heart of the efficiency story
The single most important new capability in 9.1 is Enhanced NVMe Memory Tiering. The platform now operates a unified memory model in which hot data lives in DRAM and colder pages get tiered to NVMe automatically. From the application’s perspective nothing changes; from the infrastructure team’s perspective, effective memory capacity grows substantially without buying more DIMMs.
This is the feature behind the 40% server-cost claim. For memory-bound workloads — large in-memory databases, AI inference, analytics platforms — it is genuinely transformative. VCF Operations now also includes capacity recommendations specifically for memory tiering, plus a What-If Analysis tool that quantifies cost savings and VM density improvements before you commit to a configuration change.
Alongside it, Extended vSAN Dedup and Compression drives the 39% storage TCO reduction. The compression engine has been retuned for the data shapes typical of AI pipelines — embeddings, model weights, and tokenized datasets compress very differently than traditional VM disks, and 9.1’s algorithms are specifically tuned for that. Topology Aware Scheduling rounds this out by placing workloads with NUMA and accelerator locality in mind, which removes a common source of unexplained performance variance on multi-socket GPU hosts.
AI and Kubernetes: a real platform, not a bolt-on
VCF 9.1 is the release that finally treats Kubernetes and AI as first-class citizens of the platform rather than something running alongside it. The most important additions:
- Private AI Model and GPU Metrics expose utilization, memory pressure, and model-level visibility on the same VCF Operations console used for the rest of the estate. MLOps teams no longer need a separate observability stack for the AI portion of their environment.
- VKS and VM Fast-Deploy materially shortens time-to-running for both Kubernetes pods and traditional VMs — important because deploy latency has become one of the most-cited complaints from internal application teams in large estates.
- VKS cost showback and chargeback follow the FinOps Open Cost and Usage Specification (FOCUS) framework, giving infrastructure teams the financial visibility their CFOs are starting to demand for Kubernetes spend.
- VKS Reference Architectures with cloud-native ISVs provide validated starting points so platform teams aren’t designing every new Kubernetes deployment from a blank canvas.
- Native Object Storage arrives in tech preview, bringing S3-compatible storage natively into the platform for the first time. This is significant — most AI data pipelines depend on object storage, and being able to host that natively in VCF rather than an external system removes a real friction point.
- SQL Server DBaaS elevates the most-deployed enterprise database to a first-class citizen of the private cloud control plane.
For the operational side, Tanzu Marketplace integration gives platform teams a curated path to certified middleware and data services. Organizations no longer have to self-validate every container image they want to bring into the environment.
Security: zero-trust extends into Kubernetes
For the first time, VCF 9.1 extends distributed IDS/IPS protection into Kubernetes AI workloads. This closes the most awkward gap in VMware’s lateral-security story — Kubernetes pods that previously had to be protected by a separate set of tools can now be inspected by the same vDefend distributed firewall that protects VMs. Broadcom’s quoted inspection throughput of 9 Tbps for distributed inference is meaningful; the prior generation of east-west inspection wasn’t designed for the traffic volumes AI inference produces.
A few other security additions deserve attention:
- Confidential Computing reaches general availability with Intel TDX and AMD SEV-SNP. Workloads running in Trust Domains (Intel) or Confidential VMs (AMD) get hardware-enforced memory isolation from the hypervisor itself. This is the third pillar of data protection — at-rest, in-transit, and now in-use — and it’s the version regulators are starting to require for sensitive workloads.
- Live Patching now extends to TPM-enabled hosts by default. With nearly 90% of new server hardware shipping with TPMs, this means critical security patches can land on production infrastructure without VM evacuations or host downtime.
- Self-service security with automation introduces centralized tagging, predefined security profiles, and delegated firewall configuration so application teams can secure their own workloads without filing tickets — exactly the model that makes zero-trust survivable at scale.
Networking and operations: a real platform, simplified
A few less-publicized changes have outsized operational impact:
- Unified EVPN with Arista, Cisco, and SONiC delivers a single overlay fabric across the three dominant data center networking stacks. Network teams get one operating model regardless of switch vendor — which dramatically shortens the time to onboard new sites or absorb acquired data centers in M&A scenarios.
- VPC Policy-based Connectivity introduces three Community types (Regular, Isolated, and Shared) that let admins enforce VPC isolation and shared services without rewriting firewall rules every time a new department needs a network.
- Native Infoblox IPAM integration provides a single source of truth for IP and DNS, automatically preventing IP conflicts across the environment.
- VCF Management Services on a shared runtime is a significant architectural change. Lifecycle management, the software depot, log management, and real-time data services no longer run as separate appliances; they’re distributed on a unified runtime that simplifies the management plane considerably.
- 4× parallel upgrade capacity — the ability to upgrade 256 clusters simultaneously — is the change that anyone who has nursed a large estate through a quarterly maintenance window will appreciate most.
What this means for your upgrade roadmap
Not every customer should rush to 9.1 on day one. Here’s the honest assessment of who benefits from this release fastest, and where the value lands:
Move now if: You’re running production AI workloads or planning to within the next two quarters. The Kubernetes-native security, GPU observability, and memory tiering features in 9.1 are the most significant uplift to AI-on-VMware in years. The cost-per-inference math changes meaningfully.
Move soon if: You operate a large VCF estate (more than ~30 clusters) and have felt the pain of serial upgrade windows. The 4× parallel upgrade capacity alone changes how you plan maintenance for the next several years.
Plan deliberately if: You’re on a stable VCF 5.x or 4.x footprint with no near-term AI ambitions and no networking M&A in the pipeline. Many of the headline features (memory tiering, EVPN unification, native object storage) require either new hardware or new architectural decisions, and there’s no harm in moving on a measured timeline that lines up with your normal hardware refresh cycle.
The common thread across all three is that the value of 9.1 compounds when you operate at scale. Smaller environments will get the security and operational improvements; larger estates will see the cost-reduction and AI-platform benefits more dramatically.
How 27 Virtual helps customers move to and operate VCF 9.1
We’ve spent the last decade specializing in VMware Cloud Foundation deployments, and 9.1 sits squarely in our wheelhouse. The features that make this release powerful — memory tiering optimization, AI workload integration, EVPN fabric unification, large-scale parallel upgrades — are also the features that benefit most from experienced design and migration work. The math on a 40% server-cost reduction looks great in a slide; capturing it in your specific environment requires careful capacity modeling, the right NVMe configurations, and a workload-placement strategy that holds up under production load.
For existing VCF customers planning a 9.1 path, we offer a free upgrade-readiness assessment that maps your current architecture against the 9.1 feature set and identifies where the realized savings actually land in your environment. For organizations evaluating VCF for the first time — particularly those building production AI infrastructure — we run a similar exercise to model what a 9.1-based private cloud would look like compared to your current state.
If you want to see what VCF 9.1 means for your specific environment, schedule a consultation. We’ll walk through the release, the math, and the architecture decisions that move the most value into your column.
27 Virtual is a Broadcom Pinnacle Partner and one of North America’s leading VMware specialists. Our engineers have led more than 350 VCF deployments and we’re authorized to license, design, deploy, and educate across the full VMware portfolio.