Broadcom opened VMware Explore 2026 in Las Vegas today with the clearest articulation yet of where VMware Cloud Foundation is heading: the Private AI Cloud. After two years of positioning VCF 9 as the modern private cloud platform, the message this year is that private cloud and private AI infrastructure are no longer separate disciplines — they are converging into a single platform. Here's my rundown of everything announced today, and what it means for those of us designing and running these environments.

VMware Private AI Cloud: the new umbrella

The headline announcement is VMware Private AI Cloud, Broadcom's integrated approach to running AI workloads — inference, agentic applications, and traditional enterprise workloads — side by side on infrastructure you own. It brings together VCF 9 as the foundation, the new VMware AI Factory, the VMware Tanzu Platform for agent foundations, vDefend for hypervisor-level microsegmentation, and the brand-new AgentMinder for AI agent governance.

Ram Velaga, president of Broadcom's Infrastructure Software Group, called it "the inflection point where enterprise private cloud and private AI infrastructure stop operating as separate disciplines." The timing is backed by Broadcom's own Private Cloud Outlook 2026 research: 56% of enterprises now plan to run production AI inferencing on their private cloud.

VMware AI Factory: bare metal to inference in hours

For infrastructure architects, VMware AI Factory is the most significant piece. It's a software-defined foundation — included in VCF at no additional cost — that automates the full path from bare metal to a running model. Broadcom claims deployment times drop from weeks to hours. The building blocks:

• Automated AI-ready infrastructure — provisioning and unified lifecycle management that spans hardware and software, with heterogeneous bare metal automation through a MetalSoft integration.
• Model Runtime and a unified Model Gallery — one interface for managing model inference and RAG workflows across VMs, containers, and GPU resources, with new multi-tenant model sharing via isolated namespaces.
• GPU resource pooling — multiple models sharing the same accelerators to improve utilization and control what Broadcom is now calling "AI tokenomics."
• AI Gateway — unified governance across on-prem and cloud endpoints with intelligent prompt routing, token limiting, and rate limiting.
• Secure AI sandboxes — isolated container spaces for executing agent-generated code with proper control layers.

Just as notable is the hardware story. AI Factory pairs VCF not only with NVIDIA but now formally with AMD Instinct GPUs and the ROCm ecosystem, running on certified AI ReadyNodes from Cisco, Dell Technologies, Lenovo, and Supermicro. VMware Chief Product Officer Paul Turner summed it up: customers get "a software-defined foundation that automates infrastructure deployment, unifies lifecycle management, and lets them choose their preferred hardware and vetted models."

Leading AI models, validated on VCF

Broadcom also announced a set of frontier and enterprise models tested and validated to run on VCF:

• NVIDIA Nemotron 3 — multimodal, hybrid Mamba-Transformer architecture, 1M-token context window.
• Google DeepMind Gemma 4 — open-weight multimodal model built for local execution.
• NEC cotomi — Japanese-language model with roughly 40% better token efficiency.
• Alibaba Qwen 3.7-Max — proprietary multimodal model with a one-million-token context window.
• Z.ai GLM 5.2 — open-source model aimed at coding and reasoning agents.

On top of that, VCF supports more than 150 open-source models through the vLLM runtime, with MLPerf Inference v5.1 results showing performance "on par with bare metal" — and mixed compute across AMD, Intel, and NVIDIA hardware. Chris Wolf framed it as "a clear path to data sovereignty and cost-effective AI at scale," which is exactly the conversation I'm having with regulated and government customers every week.

AgentMinder: governance for autonomous agents

AgentMinder — generally available today — is Broadcom's answer to the question every security team is asking: what happens when AI agents start acting on enterprise systems? AgentMinder treats every agent as an enterprise identity with a declared mission, permitted intents, approved tools, and authorized resources. A cloud-native gateway authenticates and authorizes every tool call at runtime against a dynamic policy engine, and OpenTelemetry-based observability delivers compliance-grade audit trails — a "chain of custody" for agent actions.

Broadcom has been running it internally at serious scale: 36 million customer-related and 7 million workforce-related API calls per day, across more than 20 million customer identities. Clayton Donley described it as "a traffic controller" for agents — verifying what they're doing, providing guardrails, and tracking their work.

Tanzu Platform: AI-ready data foundations

The Tanzu team is tackling what Purnima Padmanabhan called the real blocker: "Enterprises do not have an AI ambition problem. They have an agent trust problem." Coming in Fall 2026, the Tanzu Platform adds hardened agent sandboxes with a deny-by-default runtime model (isolating credentials and blocking prompt-injection and unauthorized network access), AI-ready data foundations that process structured and unstructured data on-prem to give agents high-precision context (fewer hallucinations, lower token spend), and governance across three dimensions: access, context, and lineage — so every agentic decision can be traced back to its data sources. A curated marketplace of vetted models and tools plus a Tanzu AI gateway for monitoring, rate limiting, and credential management round it out.

Security hardening: vDefend, Avi, and TrueSource

A few security announcements deserve their own call-out:

• vDefend gains detection of unauthorized shadow AI usage inside the private cloud — a topic close to my heart, as regular readers know.
• Avi Load Balancer adds Agentic Threat Defense, with MCP tool restrictions and real-time threat isolation to prevent agent tool misuse and data exfiltration.
• TrueSource is a new verified open-source portfolio, available today: Trusted Artifacts (secure builds for Java, Python, and Node.js), Data Services (PostgreSQL, RabbitMQ, MySQL, Valkey), alongside the existing Spring Enterprise offering.

Ecosystem momentum

Beyond the product news, colocation provider Flexential announced plans to build AI offerings on VMware AI Factory launching early 2027, combining its infrastructure footprint with hosted Private AI services — an early sign that the AI Factory model will extend beyond enterprise datacenters into the provider ecosystem.

My take

What strikes me most about this year's Explore is coherence. Last year, Private AI Foundation with NVIDIA was an add-on you architected alongside VCF. Today, AI Factory is simply part of the platform — the same VCF domains, the same lifecycle management, the same operational model, now extended from bare metal all the way to model endpoints. For organisations where data sovereignty is non-negotiable, the combination of validated models running at near bare-metal performance, agent governance with a real audit trail, and shadow-AI detection makes the private cloud case stronger than it has ever been.

The multi-vendor accelerator story (NVIDIA and AMD), the model-vendor breadth (from Nemotron to Gemma to Qwen), and the fact that AI Factory ships in VCF at no additional cost all point in the same direction: Broadcom wants the private AI conversation to start — and stay — on VCF. I'll be digging into the AI Factory architecture, Model Runtime, and AgentMinder in detail in upcoming posts and videos as the bits become available.

Broadcom Introduces VMware Private AI Cloud (Aug 31, 2026)

Broadcom Announces VMware AI Factory (Aug 31, 2026)

VMware Cloud Foundation Brings Leading AI Models to the Private AI Cloud (Aug 31, 2026)

Broadcom Unveils AgentMinder (Aug 31, 2026)

Broadcom Unveils AI-Ready Data Foundations in VMware Tanzu Platform (Aug 31, 2026)

VMware Explore 2026 Las Vegas Media Kit — Broadcom

Network World — Private AI cloud, agentic infrastructure dominate VMware Explore

Athom recently released the Homey Self-Hosted Server as an official Docker image (ghcr.io/athombv/homey-shs), which means you can now run the Homey smart home platform on your own hardware instead of a Homey Pro appliance. That raises a fun question for those of us running VMware Cloud Foundation at home or in the lab: where does a container like this land in a VCF 9.1 environment? A vSphere Pod on the Supervisor, a VKS cluster, or a plain Docker VM?

The short answer: put it in a small Linux VM in a workload domain. The longer answer is more interesting, because this little container is a nice case study in why some workloads fundamentally don’t fit vSphere Pods — and why that is by design.

What the Homey container actually needs

The requirements from Athom’s documentation are modest: a Linux-based OS (Debian or Ubuntu), an Intel/AMD or Arm 64-bit CPU, 1 GB of RAM and 1 GB of storage. But two lines in the Docker Compose file do all the architectural heavy lifting:

services:

  homey-shs:

    image: ghcr.io/athombv/homey-shs:latest

    container_name: homey-shs

    restart: unless-stopped

    network_mode: host

    privileged: true

    volumes:

      - ~/.homey-shs:/homey/user

network_mode: host and privileged: true — Homey wants to sit directly on the LAN with a dedicated IP address, because it relies on mDNS multicast for device discovery, and it wants near-unrestricted access to the host kernel. Athom explicitly notes that virtual machines should use bridged networking so the server is reachable on the LAN. The server listens on four ports: 4859 (HTTP + Socket.io), 4860 (HTTPS + Socket.io), and 4861/4862 for the Homey Bridge server (v1/v2). All persistent state lives in a single volume.

Keep those two Compose lines in mind — they decide the placement discussion below almost on their own.

The recommended option: a small Linux VM in a workload domain

This is by far the most practical route in VCF 9.1, and it takes about fifteen minutes:

1. Deploy a small VM (Ubuntu or Debian, 2 vCPU, 2 GB RAM, 20 GB disk is plenty) on an NSX segment — overlay or VLAN-backed — that has Layer 2 reachability to your IoT and smart home devices. mDNS discovery works within the subnet, so placement on the right segment matters more than sizing.
2. Install Docker and start the container with the Compose file above (docker compose -f homey-shs.yml up --detach).
3. Open ports 4859–4862 in the NSX distributed firewall (and gateway firewall if applicable), then pair the server with your Athom account via my.homey.app.
4. For Zigbee, Z-Wave and Thread you need radios. The clean way in a virtual environment is a Homey Bridge on the network, which talks to the server over ports 4861/4862. The alternative — a USB dongle with ESXi USB passthrough — works, but pins the VM to a host and breaks vMotion and DRS, which defeats much of the purpose of running it on VCF in the first place.
5. Put the ~/.homey-shs volume in your backup scope, or simply protect the whole VM with your VADP-based backup product. That one directory holds the entire configuration.

The Kubernetes option: VKS

Could you run this on a vSphere Kubernetes Service cluster instead? Technically, yes. VKS worker nodes are ordinary Linux VMs with a full kernel, so a pod with hostNetwork: true and privileged: true, plus a PersistentVolumeClaim through the vSAN CSI driver, will run fine. But you gain very little: Homey is a single-instance stateful appliance, it would share the node’s IP address, and mDNS discovery becomes harder to reason about. Kubernetes adds operational surface without adding value here. If you already run everything on VKS as a matter of principle, it works — otherwise, keep it simple.

Why vSphere Pods are off the table

The Supervisor’s native vSphere Pods might seem like the elegant answer — no VM to manage, no VKS cluster to maintain. But privileged containers and vSphere Pods fundamentally don’t mix, and it is worth understanding why.

On regular Kubernetes, a pod runs as a set of processes on a node: a Linux machine with one kernel shared by dozens of pods. privileged: true means “give this container near-total access to that node kernel” — all capabilities, direct access to /dev, the ability to load kernel modules and manipulate the host network stack, with seccomp and AppArmor out of the way. That is exactly what Homey uses (together with host networking) to sit directly on the LAN.

A vSphere Pod works fundamentally differently. The Supervisor runs every pod through the CRX runtime as its own stripped-down virtual machine directly on ESXi, with a dedicated, slimmed Photon kernel per pod. There is no shared node kernel to get privileged access to — the “host” is the ESXi hypervisor itself, and that is deliberately never exposed to a workload. This is precisely the selling point of vSphere Pods: VM-grade isolation for every single pod. Allowing privileged access would break that security model, so the Supervisor simply doesn’t support it.

For the same reason, every pod spec that refers to a “node” is unavailable:

• hostNetwork — there is no host network stack to join; the pod’s vNIC attaches to an NSX segment
• hostPath volumes — there is no node filesystem
• hostPID and hostIPC — no shared process or IPC namespaces
• Device passthrough — no way to hand a USB Zigbee dongle to a CRX pod

Homey Self-Hosted Server asks for almost exactly the list of things vSphere Pods deliberately don’t offer. That is not a limitation to work around; it is two security models that point in opposite directions.

Conclusion

VCF 9.1 gives you three ways to run containers, and this workload sorts them out neatly. vSphere Pods are excluded by design — their per-pod micro-VM isolation is incompatible with a privileged, host-networked container. VKS can do it, but brings Kubernetes overhead to a single-instance appliance that doesn’t benefit from it. A small Ubuntu VM on the right NSX segment, running Docker with the official Compose file, gets you a fully supported Homey server with vMotion, DRS, HA and your regular backup tooling — exactly the qualities you run VCF for.

The cornerstone of the VMware Cloud Foundation curriculum is getting its 9.1 refresh: VMware Cloud Foundation: Build, Manage, Secure [v9.1] is about to be released. I've been reviewing the course materials, and this five-day, hands-on training remains the definitive starting point for anyone who needs to design, deploy, and run a VCF private cloud — and the natural entry point into the rest of the 9.1 curriculum.

What the course is about

VMware Cloud Foundation delivers public cloud scale and agility with on-premises security, resilience, and performance. This course takes you through the complete lifecycle of a VCF 9.1 private cloud: planning and preparing the deployment, bringing up the environment with the VCF Installer, performing post-deployment tasks, and then managing the platform day to day — fleet management, workload domains, networking, storage, certificates, lifecycle management, and security.

The management experience is built around VCF Operations throughout. From navigating the redesigned console and verifying its integration with the management domain, to licensing, identity, password management, and security posture — you learn to run the fleet the way it is actually operated in 9.1.

Course outline

Twelve modules over five days:

1. Course Introduction
2. VCF Private Cloud: Overview and Architecture
3. VCF Private Cloud Deployment
4. VCF Post-Deployment Tasks
5. VCF Fleet Management
6. VCF Workload Domain
7. VCF Networking
8. VCF Storage Management
9. VCF Certificate Management
10. VCF Life Cycle Management
11. VCF Security
12. VCF Upgrade Paths

Thirty-one hands-on labs

The lab kit follows the lecture flow with 31 exercises, a mix of live labs and guided simulations. You deploy VCF with the Quick Installer, manage the environment through the VCF Operations UI, deploy the VCF Operations for Logs appliance, configure Single Sign-On and assign permissions to SSO users, and manage passwords across the fleet.

The workload domain module is particularly complete: commissioning hosts, creating a workload domain — including one without a vSphere cluster and one with the Supervisor — adding ESX hosts, assigning licenses, and importing an existing vCenter as a workload domain using VCF Operations.

A few more highlights:

• Modern VCF networking — the VPC consumption model with hands-on labs for configuring a Distributed Transit Gateway, creating distributed and centralized Virtual Private Clouds, deploying a VNA cluster with connectivity profiles, and building an Edge cluster with a Tier-0 gateway from vCenter.
• vSAN storage — reviewing the workload domain vSAN cluster, analyzing VM object and component placement, running vSAN health tests, and creating and applying storage policies.
• Platform operations — implementing certificates, configuring backups, and optional labs on the VCF APIs and SoS commands.
• Security and compliance — the integrated security features across all layers of VCF, and security management from VCF Operations.
• Upgrade paths — every supported route to 9.1, including vSphere 8.0 to VCF 9.1, vSphere with multiple Aria components, and VCF 5.x with multiple Aria components.

Who should attend

This course is aimed at administrators, engineers, and architects who are new to VMware Cloud Foundation or moving up to 9.1, with working experience of vSphere as the recommended starting point. It is also the prerequisite for the rest of the 9.1 curriculum — the What's New, Automate and Operate, and Troubleshooting courses all build on it — and it aligns with the upcoming VCP-VCF 9.1 Administrator certification.

Final thoughts

If you are building your VCF 9.1 skills from the ground up, this is the course to start with. It covers the full journey from bare hosts to a secured, workload-ready private cloud, and the 31 labs make sure you have actually done everything you have learned. The course will be released shortly through Learning@Broadcom — keep an eye on this blog for scheduling details.

Broadcom is about to release a brand-new edition of the VMware Cloud Foundation: Troubleshooting course, fully updated for VCF 9.1. I recently got a first look at the course material, and this five-day, hands-on training is shaping up to be an essential deep dive for anyone who operates and supports a VCF private cloud.

Why this course matters

VMware Cloud Foundation brings compute, storage, networking, automation, and operations together in a single platform — which is great, until something breaks. Misconfigurations and system failures can surface anywhere in the stack, from the initial bring-up to day-2 lifecycle operations. This course teaches administrators a structured approach to isolating and resolving problems across the entire VCF fleet, using the tools built into the platform itself.

What's new in the 9.1 edition

Compared to the 9.0 course, two things immediately stand out. The course now opens with a dedicated module on Monitoring and Troubleshooting with VCF Operations, including the Troubleshooting Workbench with its real-time analysis tools and log management. And at the other end of the week, there is a completely new module on Troubleshooting vSphere Kubernetes Service (VKS), covering the vSphere Supervisor architecture and troubleshooting VKS cluster deployments with logs and health checks — a welcome addition now that Kubernetes is a first-class citizen in VCF.

The upgrade module has also been reworked around real-world 9.1 scenarios, such as upgrading from vSphere 8 to VCF 9.1 and moving a VCF environment with multiple Aria components to 9.1.

Course outline

The course spans twelve modules over five days:

1. Course Introduction
2. Monitoring and Troubleshooting with VCF Operations
3. Troubleshooting New Deployment
4. Troubleshooting Upgrades
5. Troubleshooting License Management
6. Workload Domain Troubleshooting
7. Troubleshooting VCF Compute
8. Troubleshooting VCF Networking
9. Troubleshooting VCF Storage
10. Troubleshooting Using VCF Operations Fleet Management
11. Troubleshooting VCF Automation
12. Troubleshooting vSphere Kubernetes Service

Hands-on from day one

This is not a slideware-only course. Thirteen lab exercises run throughout the week, and they mirror what you actually face in production: analyzing SDDC Manager pre-check errors during a management domain upgrade, reviewing post-deployment logs, working with the VCF APIs, troubleshooting compute, networking, and vSAN issues, running vSAN health tests, fixing certificate and password management problems through Fleet Management, and diagnosing VCF Automation platform failures.

A few highlights from the content itself:

• VCF Operations as the starting point — fleet visibility monitoring, the Troubleshooting Workbench, and log analysis form the foundation for everything that follows.
• Deep networking coverage — from NSX management cluster status and Edge node validation to live traffic analysis, routing validation, and the new Virtual Network Appliance.
• Storage troubleshooting — monitoring vSAN through the UI and digging deeper with CLI tools.
• Fleet Management — certificates, passwords, identity and access management, tags, and configuration drift remediation using vCenter configuration profiles.
• VCF Automation — clustered deployment architecture and troubleshooting the automation platform, including microservice pod execution anomalies using container log architectures.

Who should attend

The course is aimed at administrators and support engineers responsible for a VCF 9.1 environment. You'll get the most out of it if you've already followed VCF: Build, Manage, Secure [v9.1] or have equivalent hands-on experience. It also aligns with the upcoming VCP-VCF 9.1 Support certification, making it a logical preparation step.

Final thoughts

Troubleshooting skills are what separate an administrator who can operate a platform from one who can keep it running when things go wrong. With the added coverage of VCF Operations, the Troubleshooting Workbench, and vSphere Kubernetes Service, the 9.1 edition reflects how troubleshooting in a modern VCF environment actually works today. Keep an eye on the course catalog — this one is worth booking as soon as it becomes available.

A new edition of one of the most hands-on courses in the VMware Cloud Foundation curriculum is about to be released: VMware Cloud Foundation: Automate and Operate [V9.1] (VCFAOP91). Built around the VCF 9.1 release, this course teaches you how to consume, automate, and operate a modern private cloud — from configuring identity providers and provider networking all the way to policy-driven operations, health diagnostics, and workload monitoring.

What the course is about

VCF 9.1 brings the automation and operations experience of the private cloud closer together, and this course follows that same structure. The first half focuses on VCF Automation: you learn how the platform is built from providers, Regions, and Organizations, how the self-service IaaS experience works for application teams, and how blueprints, catalogs, and the embedded orchestrator turn infrastructure into consumable services. The second half switches to VCF Operations: metrics, views, reports and dashboards, custom symptoms and alert definitions, policies, log management, and the Health, Findings, and Troubleshooting Workbench capabilities that keep your fleet healthy.

New in the 9.1 edition is coverage of the VCF Automation 9.1 network architecture, the Multi-Supervisor Region Quota (MSRQ) enhancements for content and capacity management, and automated monitoring of VKS (vSphere Kubernetes Service) clusters in VCF Operations.

What you will learn

By the end of the course you should be able to:

• Describe the VCF Automation components and supporting architecture.
• Perform infrastructure management as a cloud provider and set up Regions and Organizations.
• Configure networking at the provider and organization level using the VCF Automation 9.1 network architecture.
• Manage content libraries, blueprints, and catalogs — including the MSRQ enhancements introduced in VCF 9.1.
• Deploy VM workloads through the catalog and Kubernetes-based IaaS deployments, plus containers via VCF Automation.
• Use VCF Operations to build custom views, reports, dashboards, symptoms, alerts, and policies.
• Troubleshoot with VCF Health, Diagnostics, and the Troubleshooting Workbench, and monitor logs, storage, and VKS clusters.

Eighteen modules, thirty-three labs

The course is organized into 18 modules. Modules 2 through 10 cover the automation side: VCF Automation overview and architecture, provider infrastructure and identity management, Regions, Organizations, networking, organization management, content management, and deploying modern apps. Modules 11 through 18 cover operations: the VCF Operations platform, metrics, views and reports, dashboards, policies, symptoms and alerts, health and findings, logs and storage, and workload monitoring.

What really sets this course apart is the lab kit: 33 hands-on labs that mirror the lecture flow. You start by exploring the VCF Automation portals and creating your first Region, work your way through provider networking, projects, policies, content libraries, and blueprint deployments, and finish on the operations side building custom views, reports, dashboards, alert definitions, and troubleshooting workbench sessions. A few advanced scenarios — such as multi-Supervisor Region Quotas, the real-time metrics component, and VKS cluster monitoring — are delivered as guided simulations.

Who should attend

The course targets cloud administrators, operators, and architects who are responsible for delivering and running a private cloud on VMware Cloud Foundation 9.1. If you attended the v9.0 edition, the 9.1 course is the natural next step; combining it with the VCF: What's New [v9.1] course is a great way to get fully up to speed on the new release. Experience at the level of VCF: Build, Manage, Secure is the recommended starting point.

Availability

The course will be released shortly in both instructor-led and on-demand formats through Learning@Broadcom. Certified instructors are currently completing the train-the-trainer program and lab preparation, and the course also lines up with the upcoming VCP-VCF 9.1 Administrator certification.

Keep an eye on this blog — as soon as the course goes live I will share the scheduling details. If you want to prepare in the meantime, the VCF: What's New [v9.1] on-demand course is the perfect warm-up.