VMware Explore · VCF 9.1 · Private AI · Event Recap
VMware Explore 2026 ran in Las Vegas from 31 August to 3 September. This recap covers what Broadcom announced across VCF, Private AI, security and Tanzu, which parts you can install today, what is still on the roadmap, and what it means for VCF customers in India. Details come from Broadcom’s press releases that week, the VCF 9.1.1 release notes published on day four, and keynote coverage.
VCF 9.1 had already shipped in May, so there was no new platform release on stage. Broadcom instead added new layers on top of it: an AI assistant inside VCF Operations, a packaged AI Factory, governance for AI agents, and agent tooling in Tanzu. Running AI on your own infrastructure was the theme that tied all of it together.
Section 1 · EventExplore 2026 at a glance
| Item | Detail |
|---|---|
| Dates | 31 August to 3 September 2026 |
| Venue | Venetian Convention and Expo Center, Las Vegas |
| Attendance | About 4,700 people, as per Broadcom’s day 4 recap |
| Sessions | 400+ breakouts, plus hands-on labs and an on-site certification centre |
| Tracks | Cloud Infrastructure, App Modernization, Security, Innovation |
| Theme | Private AI Cloud: scale AI at lower cost, run it securely, build agents quickly |
| Keynote speakers | Ram Velaga (President, Infrastructure Software Group), Paul Turner (Chief Product Officer, VCF), Purnima Padmanabhan (GM, Tanzu) and other Broadcom leaders |
| Next US event | Resorts World Las Vegas, 3 to 6 May 2027 |
| Regional events | Explore on Tour: Mumbai, Singapore, Frankfurt, Tokyo, London, Washington D.C., Sydney |
Two things in that table tell you where Explore is heading. Attendance is a fraction of what VMworld used to draw, and Broadcom is clearly fine with that, because the event now targets people who run VCF rather than a wide partner and prospect crowd. And moving to May puts Explore closer to the spring release cycle. VCF 9.0 and 9.1 both arrived in May, so a May Explore will probably launch the next platform release on stage rather than recap it three months later.
Section 2 · OverviewEverything announced, by product area
Before going product by product, here is the full list on one screen. Some of these were announced a few weeks before Explore and then shown on stage, and that is noted where it applies.
| Product | What is new | Who should care |
|---|---|---|
| VCF 9.1 and 9.1.1 | Reboot-free NVMe memory tiering, global deduplication, live patching on TPM hosts, AI Assistant for VCF (9.1.1), multi-tenant model sharing (9.1.1) | Every VCF admin and architect |
| VMware AI Factory | Software-defined base of VMware Private AI Cloud: automated path from bare metal to model serving, model gallery, GPU pooling, AI gateway | Teams building an internal AI platform |
| MetalSoft partnership | Provision and repave servers from several vendors through the VCF console | Operations teams with mixed hardware |
| vDefend and Avi | AI Assistant in Distributed Firewall and Avi Load Balancer, vACT 3.0 for migrating off agent-based firewalls (announced 6 August) | Network and security teams |
| AgentMinder | Identity, policy enforcement and audit trail for AI agents; generally available | Security teams, platform owners letting agents call tools |
| Tanzu Platform 10.4 | Agent buildpack (tech preview), MCP gateway, AI-ready data foundations | Developers and platform engineers |
| TrueSource | Verified builds of open-source libraries for Java, Python and Node.js, plus supported data services | Application security and procurement |
| VMUG lab licences | Home-lab keys tied to VCP certification, kept alive by a yearly refresher exam | Anyone who runs a home lab |
Section 3 · PlatformWhat VCF 9.1 and 9.1.1 bring to your clusters
VCF 9.1 was the product on almost every slide, even though it was four months old. That looks deliberate: Broadcom wants existing customers to move off 5.x and 9.0, and Explore was the push. Here are the features that matter for day-to-day operations.
NVMe memory tiering, now without a reboot
ESX uses a local NVMe device as a second tier of memory. Hot pages stay in DRAM. Cold pages, the ones a VM has allocated but rarely touches, move to NVMe. From the guest’s point of view nothing changes; it simply sees its configured memory.
Memory tiering itself arrived with VCF 9.0. In 9.1, Broadcom added reboot-free enablement, so you can switch it on without taking the host down, and a software-based tiering mode that no longer needs a specific rate controller. Visibility of how much memory sits on each tier also improved.
Clusters that run out of memory long before they run out of CPU. That describes most general-purpose VCF clusters: plenty of idle cores, memory at 80% consumed, and a hardware refresh justified almost entirely by RAM. With DRAM prices up sharply in 2026, this is now a budget conversation.
What Broadcom and Lenovo claim
Vendor figures, not independent test results. Your own ratio of cold to hot memory decides what you actually get.
Hardware, storage and patching
Three quieter changes that matter more in daily operations than they did on stage.
Blackwell, Grace, ConnectX-7
Keynote coverage mentioned support for NVIDIA B200 GPUs, Grace CPUs and ConnectX-7 SmartNICs, plus topology-aware CPU scheduling that Broadcom says gives up to 30% higher SQL Server transaction rates. That 30% does not appear in the written release notes, so treat it as a keynote claim for now. Keeping a VM’s vCPUs and memory close together on many-core sockets does make sense.
Global deduplication is GA
Global deduplication is generally available in 9.1, with an improved compression algorithm and data-at-rest encryption working alongside both. For vSAN ESA clusters full of similar VMs (VDI pools, dev and test clones, templated app servers) this is real capacity back. Test on a copy of production data first; dedup ratios on slides rarely match your own.
Live patching on TPM hosts
Zero-downtime live patching now covers hosts with TPM, a common blocker in regulated environments. VCF 9.1 also added on-premises automated ransomware recovery, continuous compliance enforcement and zero-trust controls for Kubernetes AI workloads. For most VCF teams, patching without evacuating hosts matters more day to day than anything with “agent” in its name.
AI Assistant for VCF (arrived in 9.1.1)
A chat panel inside VCF Operations where you ask questions in plain English, such as “why is this host showing memory contention” or “why did this vMotion fail”. It reads alerts, configuration and logs, sends that context to a large language model, and returns a diagnosis and a suggested fix. It can also help build management packs for third-party systems without API code.
First, triage of a security alert across hundreds of ESX hosts, with compatibility checks and live patching pre-checks. Second, it spotted that an administrator had disabled DRS during maintenance and never turned it back on, fixed the imbalance and raised a compliance alert. Anyone who has run operations has lived through that one.
According to William Lam’s walkthrough, you enable Log Management, turn on the “VMware Intelligent Assist” extension, and register an LLM provider. Two are supported today: Google Gemini (cloud) or VCF Private AI Services (PAIS) running on your own GPUs.
Figure 1. How a troubleshooting request moves through AI Assistant for VCF in 9.1.1.
Known limits
- It is available only in full VCF deployments, not in VMware vSphere Foundation (VVF).
- Chat history is not kept. Move to another page and your conversation is gone.
- If you pick Gemini, VCF management components need internet access to reach it, and your operational data leaves your data centre. For banks and government customers that alone decides it: PAIS or nothing.
- When SSO is enabled, administrators need the “Extensions” permission to use it.
Other items in 9.1.1
GitOps service
Native Argo CD inside VCF Automation, with OIDC sign-in for pipelines. Lab use only for now.
NSX EVPN improvements
Direct VXLAN tunnels for east-west traffic, plus support for NAT, load balancing and DHCP relay.
VKS 3.7 add-on management
A proper framework for cluster add-ons in vSphere Kubernetes Service (VKS), which clears up who supports each one.
VCF 9.0 versus 9.1 for these features
| Feature | VCF 9.0 | VCF 9.1 / 9.1.1 |
|---|---|---|
| NVMe memory tiering | Generally available; enabling it needed a host reboot | Reboot-free enablement, software-based tiering, better per-tier visibility |
| Global deduplication | Not generally available | Generally available, with improved compression |
| Live patching | Available, with limits on host types | Extended to TPM-equipped hosts |
| AI Assistant for VCF | Not available | Generally available in 9.1.1 (VCF only) |
| Multi-tenant model sharing | Not available | Added in 9.1.1 |
| GitOps with Argo CD | Not available | Tech preview in 9.1.1 |
| Containers in vSphere Client | Managed separately from VMs | Containers and VMs managed side by side |
| Ransomware recovery | Depended on add-on services | On-premises automated recovery built in |
For step-by-step upgrade guidance, see the VCF 9.1 Upgrade Series. If you are still learning how 9.x fits together, start with the VCF 9 Complete Guide.
Section 4 · Private AIVMware AI Factory and Private AI Cloud
Broadcom announced VMware AI Factory on 31 August and described it as the software-defined foundation of VMware Private AI Cloud. In simple terms, Private AI Cloud is what you buy, and AI Factory is what runs inside it. Paul Turner framed it around “tokenomics”: knowing and controlling what each generated token costs you, in the same way you already track cost per VM.
Here is how its layers fit together, from hardware at the bottom to agents at the top.
Figure 2. VMware Private AI Cloud as a stack. Each layer runs on the one below it.
What is inside AI Factory
Models and hardware partners
150+ open and commercial models
Broadcom named these as validated:
Certified VCF AI ReadyNodes
Four server vendors, plus AMD Instinct GPUs on the open ROCm stack:
Having AMD as a real option matters. It gives buyers something to negotiate with when NVIDIA supply is tight or priced high.
MetalSoft: bare metal inside VCF
MetalSoft opened the keynote. Their integration lets you provision or repave physical servers from several vendors directly from the VCF management console, bringing firmware and hardware lifecycle into the same place as the software lifecycle. Broadcom says provisioning drops from weeks to minutes. If you have ever waited on a hardware team, a firmware baseline and a PXE server before VCF could even see a new host, you know where those weeks go. Ask your account team which server models are supported on day one and whether it needs a separate licence.
Section 5 · SecurityvDefend, Avi and TrueSource
vDefend and Avi Load Balancer
These updates were announced on 6 August, before Explore, and demoed in Las Vegas. Both vDefend Distributed Firewall (DFW) and Avi Load Balancer now include an AI Assistant for real-time insight, troubleshooting and faster rollouts. More practical is vACT 3.0 (vDefend and Avi Conversion Tool), which automates migration from legacy agent-based firewalls to DFW. Rule migration is where most micro-segmentation projects stall, so any tooling there saves real project weeks.
Versions to look for (all compatible with VCF 9.1)
TrueSource
TrueSource is Broadcom’s answer to compromised open-source dependencies. It groups three offerings:
Spring Enterprise
Supported Spring builds and patches for Java teams.
Trusted Artifacts
Clean-room builds of Java, Python and Node.js libraries at SLSA Build Level 3, plus the Bitnami Secure Images catalogue.
Data Services
Supported builds of PostgreSQL, RabbitMQ, MySQL and Valkey.
SLSA (Supply-chain Levels for Software Artifacts) is an industry framework for proving how a piece of software was built. Level 3 means the build ran on a hardened platform and produced signed records that cannot easily be faked. If your developers pull packages straight from public registries today, this closes a real gap. Pricing is by tiered site licence, so compare it with what you already pay for Bitnami or Spring support before signing anything new.
Section 6 · Apps and agentsTanzu 10.4, MCP gateway and AgentMinder
Much of Explore’s Innovation track was about AI agents: software that takes a goal, plans steps and calls tools on its own. Broadcom’s view is that agents will need the same controls already used around VMs and containers: an identity, network policy, an audit trail and a gate on what they can touch.
Tanzu Platform 10.4 agent foundations
AGENTS.md file describing what an agent should do, pushes it, and the platform turns it into a running agent. No application code needed for a simple case.AgentMinder
AgentMinder was announced on 31 August as generally available, which is unusual for a first release in a new category. Here is how it controls an agent, step by step:
It plugs into existing authorisation systems through the AuthZEN standard and runs on VKS, Google Cloud or any standards-based Kubernetes platform.
Customer stories from the keynote
Less data centre power after moving to a standardised private cloud of 200+ clusters at about 99% virtualisation. This is closer to what most VCF customers are working on right now.
Developers deploy onto VKS through automated Terraform pipelines without raising infrastructure tickets.
Section 7 · CommunityVMUG home-lab licences and certification
This was the most discussed change among practitioners, and it was not really a product announcement at all.
For years, a VMUG Advantage subscription gave you evaluation licences for most of the VMware portfolio. That ended in 2025, when Broadcom tied home-lab keys to certification. Here is where it stands today:
| Exam you pass | VMUG Advantage needed? | Licence you get | Term |
|---|---|---|---|
| VCP-VCF (Administrator or Architect) | Yes | VMware Cloud Foundation, 128 cores | 3 years |
| VCP-VVF | No | vSphere Standard, 32 cores | 1 year |
| Yearly refresher | As above | Keeps existing keys active (30 questions, open book) | Every 12 months |
Learners without a certificate
Students, freshers and people switching careers who want to learn VCF before they can afford the exam. In India that is a very large group, and it used to be one of VMware’s best talent pipelines. A lab is how you learn enough to pass the exam, and now you need the exam to get the lab.
Engineers already certified
A 128-core VCF licence for three years is generous for a home lab, and a short open-book refresher is much easier than retaking a full exam. It also gives hiring managers a slightly clearer signal.
For people preparing for their first VCP, VMware Hands-on Labs remain the free way to practise, although they do not replace building and breaking your own environment.
Section 8 · Status checkWhat you can use today and what is still roadmap
Statuses are based on release notes and press releases as of late September 2026. Please verify against your own build before planning around any of them.
| Announcement | Product area | Status | Suggested next step |
|---|---|---|---|
| VCF 9.1 | Platform | GA since 5 May 2026 | Plan your upgrade; land on 9.1.1 |
| NVMe memory tiering (reboot-free) | Platform | GA in 9.1 | Pilot on one memory-bound cluster with real metrics |
| AI Assistant for VCF | Operations | GA in 9.1.1 (3 Sep 2026) | Enable in a lab with PAIS; clear data policy before using Gemini |
| Multi-tenant model sharing | Private AI | GA in 9.1.1 | Use it when two or more teams want same model |
| GitOps service (Argo CD) | Automation | Preview | Lab only for now |
| VMware AI Factory (overall) | Private AI | Unclear mix of shipped and upcoming | Ask for a per-feature GA list and SKU |
| AI Gateway, secure AI sandboxes | Private AI | Roadmap (sandboxes reported for fall 2026) | Watch, do not plan around yet |
| MetalSoft integration | Hardware | Unclear partnership announced | Ask which servers are supported and how it is licensed |
| vDefend and Avi AI Assistant, vACT 3.0 | Security | GA in vDefend 9.1.1, Avi 32.1.4 | Run vACT against your current firewall rules to size a migration |
| AgentMinder | Agent governance | GA since 31 Aug 2026 | Pilot with one internal agent that calls real tools |
| Tanzu agent buildpack | Apps | Preview in Tanzu 10.4 | Try with a simple AGENTS.md in a dev space |
| Tanzu MCP gateway | Apps | Check part of 10.4 agent foundations | Confirm GA or preview for your edition |
| TrueSource | Supply chain | Available via site licences | Compare with existing Spring and Bitnami spend |
Section 9 · IndiaWhat this means for VCF customers in India
A few points apply to VCF customers in India more than the US coverage suggests.
Upgrade planning. Many Indian enterprises are still on VCF 5.x, often because of change freezes around financial year end and festive season traffic. If that is you, plan to land on 9.1.1 rather than 9.0. It carries a year of fixes on top of 9.0 and has features you will want anyway. If you are on 9.0 already, there is no rush unless memory cost or the AI Assistant is a real driver for you.
Figure 3. A simple decision path for moving to VCF 9.1.
Memory and GPU costs
Servers in India are priced off imported components, and DRAM has been one of the fastest-rising line items this year. Memory tiering is probably the feature with the clearest rupee value right now. B200-class GPUs are scarce and expensive here, so AMD Instinct support and multi-tenant model sharing (fewer GPUs serving more teams) matter more than the headline hardware.
Data residency
Under the Digital Personal Data Protection Act and RBI’s rules for regulated entities, sending a bank’s VCF Operations data to a cloud LLM will not pass most security reviews. For BFSI and government customers, AI Assistant for VCF is realistically a PAIS project, which means on-premises GPUs first.
Explore on Tour Mumbai
Mumbai is on this year’s Explore on Tour list. At the time of writing, the exact date was not yet published on the Explore site. If you could not travel to Las Vegas, it is the easiest way to see these demos live and ask Broadcom’s engineers about availability in person.
Sources: Broadcom press releases of 6 and 31 August 2026 (VMware AI Factory, AgentMinder, vDefend and Avi, TrueSource), VMware Explore blog day 1 and day 4 recaps, VCF blog posts on VCF 9.1 (5 May 2026) and VCF 9.1.1 GA (3 September 2026), William Lam’s AI Assistant for VCF setup guide, Virtualization Review and SiliconANGLE coverage of the keynote and the Lenovo memory tiering session. Keynote figures are Broadcom and partner claims, not independent test results.








DrJha