VMWARE EXPLORE 2026 // PRACTICAL BRIEFINGLAS VEGAS · 2026

VMware Explore 2026:
What the Announcements
Mean to You?

Six takeaways from VMware Explore 2026 for architects working on private AI, platform engineering, security and cloud costs.

The A2A Pod: Listen Along
VMware Explore 2026: What the Announcements Mean to You?

VMware Explore 2026 ran from August 31 to September 3 in Las Vegas. The official closing recap reports more than 4,700 attendees and 400 sessions. The announcements cover private AI infrastructure, agent security, application delivery and cloud costs.

VMware Cloud Foundation (VCF) 9.1 was announced in May. VCF 9.1.1 became generally available on September 3. The sessions cover both the existing platform and the additions announced around Explore.

The announcements position VCF as a shared platform for conventional applications, models and agents. Putting an AI workload into production brings specific decisions: what data it can use, how much capacity it needs, and how it will be supported when something fails. Those decisions require agreement between the infrastructure, application and data teams.

FROM REQUIREMENTS TO OPERATIONSFIG. 01
REQUIREMENTS
Workload needs
  • Data access and location
  • Capacity and budget
  • Agent permissions
PLATFORM
Shared infrastructure
VMsKubernetesDataAI

VMware Cloud Foundation

OPERATIONS
Production checks
  • Named service owners
  • Measured performance and cost
  • Tested recovery and controls
Use the pilot to check ownership, performance, cost and recovery.

Start with an AI operating model

The problem

A production AI workload needs accelerators, a model endpoint and access to data. It also needs owners for patching, identity, model access, recovery and cost. These responsibilities often span infrastructure, application and data teams. Agree on them before deployment so support does not depend on finding the right person during an incident.

At Explore

The VMware Private AI Cloud announcement describes VCF as the infrastructure for virtual machines, Kubernetes, data services and model serving. The aim is to manage lifecycle, policy, monitoring and costs across those workloads. VCF 9.1.1 adds some of the capabilities discussed at Explore; other parts of the Private AI Cloud announcement remain planned. Each deployment still needs a check against the relevant product release.

Sources: VMware Private AI Cloud · VCF 9.1.1 GA

What to do next

Choose an AI workload scheduled for production. List its compute, data, model serving, identity, network, recovery and monitoring requirements. Assign an owner and a service target to each, including cost. Resolve gaps in ownership before the pilot becomes a production service.

Check first. Check the availability and licensing of each required capability. The Private AI Cloud announcement covers multiple products and release schedules.

Related sessions

Plan capacity beyond GPUs

The problem

GPU capacity is only part of an AI workflow. Agents also retrieve context, apply policy, call tools, wait for APIs and evaluate results. CPU, memory, network or data service limits can delay those steps even when GPU capacity is available. Sizing needs to account for the complete workflow.

At Explore

The VMware AI Factory announcement covers model serving and infrastructure operations. The session descriptions discuss pooled accelerators, mixed compute, time to first token, throughput and utilization. The Private AI Services architecture session compares managed model services, VMware vSphere Kubernetes Service (VKS) and Deep Learning VMs. The validated-model program adds supported models, each of which still needs testing against workload requirements.

Sources: VMware AI Factory · validated models · AI innovations in VCF

WHERE THE TIME GOESFIG. 02
  1. Retrieve context

    Data services

  2. Apply policy

    Control plane

  3. Model inference

    Accelerators

  4. Tools + evaluate

    CPU, network, APIs

Measure separately: time to first token, tokens/sec, GPU utilization, orchestration CPU, retrieval latency, cost per workflow
Accelerators cover one stage. Size the platform on the complete workflow.

What to do next

Measure inference and the surrounding agent workflow separately in a pilot. Record time to first token, tokens per second, GPU utilization, CPU time spent in orchestration, retrieval latency and cost per completed workflow. Use those measurements to size the next deployment.

Check first. Test the model, quantization, context length, data path and concurrency pattern you plan to use. Model validation alone does not establish performance on your data.

Related sessions

Give agents identities and boundaries

The problem

An agent can choose which tool to call and what data to pass to it. A service account identifies the caller, but the system also needs to check whether each action is permitted. Model Context Protocol (MCP) connects agents to tools and data. Those connections need owners, access controls and records of what the agent did.

At Explore

The AgentMinder announcement describes controls for agent identity and actions. It evaluates actions against mission, intent, tools, resources and risk, with traces for later review. The MCP administration session covers the infrastructure that connects agents to enterprise systems. The announcements also describe agent-aware vDefend, Avi, gateway and framework controls. Several of those capabilities are planned for future releases.

Sources: AgentMinder · agentic security

WHERE THE BOUNDARIES BELONGFIG. 03
  1. Agent / model

    Identity + mission

  2. Action decision

    Allow · deny · escalate

  3. Tools / MCP

    Explicit permissions

  4. Enterprise data

    Scoped access

End-to-end evidence: actions, denials, latency and cost
Check each control against the availability of the product that provides it.

What to do next

Choose a low-risk internal agent and document its permissions before granting tool access. Define its purpose, allowed tools, data access, behavior after a denial, escalation contact and audit fields. Test permitted and prohibited tasks. Check that denials work and that the records explain each tool call.

Check first. Use controls available in the pilot environment to enforce permissions. Planned agent security features cannot provide protection until they are deployed.

Related sessions

Make self-service governable

The problem

A self-service portal can provision a cluster while leaving developers to arrange policy checks, deployment and monitoring. Teams need a supported process that covers all of these steps. Agent-assisted development increases the need for automated checks because agents can submit changes faster than people can review them manually.

At Explore

The application modernization sessions cover VKS, GitOps and Tanzu in software delivery. The agentic SDLC tutorial follows work items through source changes, governed MCP services and deployment. The new Tanzu AI data and agent capabilities are planned for general availability in Fall 2026. TrueSource provides commercially supported, verifiably built open-source artifacts. These address two practical needs: controlling how code reaches production and checking the origin of its dependencies.

Sources: Tanzu AI-ready data foundations · TrueSource

ONE SUPPORTED PATHFIG. 04
  1. Approved work item

    Intent and owner

  2. Source and policy

    Checks before merge

  3. Deployment

    GitOps, supported runtime

  4. Runtime evidence

    Monitoring and rollback

The platform team owns versions, policy, upgrades, rollback and support boundaries
Measure delivery time, manual handoffs and policy failures caught before release.

What to do next

Build a supported delivery process for a small internal service, from an approved work item to deployment and monitoring. Measure delivery time, manual handoffs and policy failures caught before release. Test recovery after a failed change, then automate the steps causing the most delay or rework.

Check first. Assign owners for versions, policy, upgrades, rollback and support. Developers need to know which parts of the process the platform team maintains.

Related sessions

Move security inside the workload

The problem

Perimeter controls do not cover every connection between internal workloads. They also cannot determine whether an agent used an approved tool for an authorized purpose. Teams need to understand traffic between applications before restricting it. That applies to conventional applications as well as new AI services.

At Explore

The vDefend journey session describes four stages: protect shared services, segment broad environments, secure application tiers and restrict the remaining traffic. The security announcements add plans for discovery and protection of MCP, agents and AI data paths. Existing segmentation controls can support work today; the planned agent-aware features have separate availability requirements.

Sources: end-to-end agentic security

THE SEGMENTATION SEQUENCEFIG. 05
  1. Shared services

    DNS, identity, NTP

  2. Environment zones

    Production, test, dev

  3. Application tiers

    Web, app, data

  4. Remaining traffic

    Default deny

Observe traffic before enforcing. Record coverage, unexpected flows, change effort and rollback
A staged path. Each stage is enforceable with controls available today.

What to do next

Choose an application or shared service. Record its normal traffic, identify dependencies and test a limited set of segmentation rules. Measure policy coverage, unexpected connections, change effort and incident investigation time. Test rollback before expanding the policy.

Check first. Observe traffic before enforcing rules. Record planned agent security features as outstanding requirements until they are available and deployed.

Related sessions

Put cloud economics into the architecture review

The problem

Cloud cost comparisons often use different assumptions. A public cloud estimate may assume steady demand, while a private cloud estimate may count hardware but omit operating costs. An AI pilot may report GPU hours without including model usage, data movement or idle capacity. Those omissions can change the placement decision.

At Explore

The VCF FinOps sessions cover resource consumption, cost drivers, rate cards, showback and chargeback. ValueOps covers comparisons across private and public clouds, including AI token costs. The vSAN session covers compression, deduplication and automation, which also affect the cost of running workloads.

COMPARING LIKE WITH LIKEFIG. 06
  1. Same service outcome

    Service level, resilience

  2. Full cost inputs

    Utilization, egress, tokens

  3. Placement options

    Private, public, hybrid

  4. Revisit at 30 days

    Variance sets the rule

State the measurement window, utilization assumption, risk allowance and migration cost
A comparison is only useful when every option carries the same assumptions.

What to do next

Compare the cost of running a representative workload in each suitable location. Use the same service and resilience requirements, and include utilization, data egress, licensing, migration, operating labor and token consumption. After thirty days of representative use, compare actual costs with the estimate and revise the placement decision if needed.

Check first. Document the measurement period, utilization assumptions, risk allowance and migration costs. A comparison is only useful if these assumptions are consistent.

Related sessions

Availability and planned releases

Record the status of each dependency in the pilot design. A committed delivery date needs a fallback for anything still in preview or planned. These labels reflect the official announcements as of September 5, 2026.

✓ Available / GA

◇ Technical preview

→ Announced / future

Plan a pilot

Choose a workload with a named owner and agreed service requirements. Measure the complete workflow, test permissions and recovery, and track its operating costs. Record where the available products meet those requirements and where additional work is needed.

Use those results to decide what to deploy next. Any production plan that depends on a preview or future release needs an alternative if that capability is delayed.

Six sessions to open first.

These six sessions cover the main topics above. Durations are taken from the official video library.

Official announcements and further reading

The official library listed 189 recordings on September 5, 2026. This article draws on the published descriptions of eleven sessions and the official announcements linked below.

Back to all postsConnect on LinkedInSubscribe via RSS

Views expressed here are my own and do not represent the views of any current or former employer, client, or affiliated organization.