The most important unstated question is whether VMware’s AI stack changes your infrastructure roadmap or simply makes an existing VCF estate harder to unwind. NVMe memory tiering is the clearest example. It improves the economics of inference hosts only when memory pressure is driven by relatively cold model state rather than latency-sensitive active working sets. That means platform teams need benchmarking beyond generic GPU utilization: page temperature behavior, tail-latency impact, NVMe endurance consumption, and failure-domain design all become part of AI capacity planning. In practice, this turns storage selection into an inference-performance decision, not just a server BOM decision.
Architecturally, the bigger shift is the consolidation of VMs, containers, model serving, and agent execution under one governed platform. That can simplify operations for VMware-heavy shops, but it also creates tighter coupling between AI runtime choices and the virtualization layer. Teams evaluating this approach should press on integration boundaries: how portable are model-serving pipelines, observability data, policy definitions, and GPU scheduling assumptions if some workloads later move to Kubernetes distributions or public-cloud inference endpoints? A unified platform reduces assembly effort now, but can increase migration friction later.
Security teams should also read past the product framing. Treating agents as enterprise identities is directionally strong, but effective control depends on disciplined tool registration, short-lived credentials, network segmentation, and auditable policy lifecycles. The operational burden shifts from “can we run agents?” to “who approves agent scope changes, and how are exceptions detected?” For many enterprises, that governance model—not the model catalog or hardware automation—will determine whether private AI becomes manageable at scale.
VMware Explore 2026 opened Monday at The Venetian Convention and Expo Center in Las Vegas with a declaration that changes what enterprise private cloud is actually for: Broadcom officially merged VMware’s private cloud platform with enterprise AI infrastructure, launching VMware Private AI Cloud and announcing that VMware Cloud Foundation is no longer primarily a virtualization successor — it is now the single governed platform through which organizations are expected to run production inference workloads, autonomous AI agents, containers, and traditional VMs. The conference, which runs through Thursday, September 3, has drawn an estimated 10,000-plus IT engineers, cloud architects, and platform teams for four days of sessions, hands-on labs, and certifications centered on that expanded vision.
The announcement lands at a moment when the economics of enterprise AI infrastructure are forcing a structural choice. According to Broadcom Private Cloud Outlook 2026 — a global survey of 1,800 senior IT decision-makers — 56% of enterprises are already running or planning to run production AI inference on private cloud, while public cloud’s share of the same workloads dropped 15 percentage points in a single year, from 56% to 41%. The reason, as survey respondents described it: public cloud pricing was built for variable, CPU-bound workloads, while AI inference at scale runs GPUs at near-constant high utilization — a mismatch that makes on-premises ownership structurally cheaper once production volumes are reached.
VMware AI Factory: What "Metal to Model in Hours" Actually Requires
The technical centerpiece of Monday’s announcements is VMware AI Factory, the software-defined infrastructure layer that Broadcom says can compress the time from an unboxed server to a running AI model from weeks to hours. VMware AI Factory is the software-defined foundation of VMware Private AI Cloud, providing customers a simplified path to production with new automation innovations for deploying AI-ready infrastructure and Day 2 operations. Understanding what makes that claim credible — and where its limits are — requires understanding the hardware problem it is solving.
Server DRAM prices spiked dramatically in 2026. Standard 32 GB DDR5 modules that traded at roughly $95 in mid-2025 climbed past $500, and fulfillment times stretched significantly for high-density enterprise memory modules from the standard two to three weeks to twelve to sixteen weeks. For AI inference workloads, which require large memory footprints to load model weights into active memory, this created a genuine infrastructure bottleneck: the cost of equipping GPU servers with sufficient DRAM was threatening the on-premises economics that made private cloud inference appealing in the first place.
NVMe memory tiering is Broadcom’s architectural answer. Introduced in VCF 9.0 and production-ready, memory tiering creates a two-tier hierarchy inside the ESXi hypervisor: Tier 0 is conventional DRAM, handling hot pages — data actively needed by running workloads; Tier 1 is high-performance enterprise NVMe solid-state storage, holding cold pages that are accessed infrequently. The ESXi memory manager monitors page access patterns in real time, promoting cold pages to DRAM when they become hot and demoting hot pages when access slows. The default configuration effectively doubles a host’s addressable memory — for every 1 terabyte of DRAM installed, another terabyte of NVMe capacity is presented as contiguous memory — at a fraction of the cost of purchasing additional DIMMs.
The mechanism is not magic: it requires that a workload’s active working set fits within the DRAM tier. For AI inference, where model weights are loaded once and then repeatedly queried, a significant portion of memory is cold at any given moment — making memory tiering a practical fit for exactly the workload category VMware Private AI Cloud is targeting. The NVMe drives must be enterprise-class (Class D endurance or higher, Class F or G write performance), and the feature ships with host-level and VM-level AES-XTS encryption for cold pages so that cold pages written to the NVMe tier remain compliant with enterprise security requirements.
Beyond memory tiering, VMware AI Factory automates the full infrastructure path: bare-metal server provisioning, VCF deployment, GPU pool configuration, network and storage integration, and Day 2 operational tooling. It supports heterogeneous GPU hardware — AMD Instinct GPUs with ROCm, NVIDIA GPUs, and server platforms from Dell, Cisco, Lenovo, Supermicro, and others — through a partnership with MetalSoft for bare-metal provisioning automation.
Model-as-a-Service on Premises: 150 Models, Five Validated
Once the infrastructure is running, VMware AI Factory delivers what Broadcom calls "Validated AI Models for VCF" — enabling enterprises to run AI models as governed, on-premises services through VCF’s built-in vLLM-based runtime. vLLM (an open-source large language model serving framework) provides efficient inference through continuous batching and GPU memory management, allowing multiple model tenants to share the same hardware without requiring separate model copies per team. VCF supports more than 150 models, including Nemotron 3, Gemma 4, cotomi, Qwen 3.7-Max, and GLM 5.2. Five models have been validated by their providers specifically for VCF deployment: models from NVIDIA, Google DeepMind, NEC, Alibaba, and Z.ai. IT administrators gain a built-in observability dashboard tracking token throughput, latency, GPU utilization, and memory utilization — the metrics that matter for managing AI tokenomics at scale.
An AI Gateway is listed as "coming soon" in Broadcom’s VMware Explore announcements. It will add application authorization, token rate limiting, and intelligent prompt routing to the model-serving layer once available.
A note on model providers: Alibaba, a Chinese company, is among the five validated commercial model providers. Organizations deploying Alibaba model weights on VCF should review their data governance policies and confirm that model telemetry and update mechanisms do not route data outside their private network perimeter. This applies to model weights from any provider whose infrastructure is subject to non-US data access laws.
AgentMinder: Governing Autonomous AI Agents as Enterprise Identities
The most architecturally novel announcement is AgentMinder by Broadcom, a control plane for autonomous AI agents that Broadcom released generally on Monday.
The challenge AgentMinder addresses is one the security industry has been grappling with since enterprise AI agent deployments began scaling in 2025: unlike conventional software, autonomous AI agents can initiate actions without explicit human instruction, exceed their intended scope, and accumulate access to APIs, external networks, credentials, and data sources over time. A 2026 BeyondTrust analysis of enterprise environments found that AI agents inside organizations increased by 466.7% year-over-year, while a separate survey found 92% of security leaders consider agent governance critical but only 44% have implemented any policies.
AgentMinder’s approach treats each AI agent as an enterprise identity rather than a software process — the same conceptual shift that Zero Trust applied to human users two decades ago, now applied to non-human actors. Each agent is assigned a unique identity, bound to a specific mission, a declared set of approved tools, and a defined set of authorized resources. At runtime, every tool invocation by the agent triggers a least-privilege access check: the agent cannot call a tool, access a data source, or reach an API that is not explicitly authorized for its current task, regardless of what the agent’s LLM "wants" to do. Compliance-grade audit logs record every action.
"We’ve had 50 years to figure out how to control and manage employees," Clayton Donley, Broadcom’s VP and general manager of Identity Management Security, said in a media briefing. "The question today is how to put the same kinds of controls around AI agents — and how do we control it when it’s actually happening, not just at the governance level after the fact."
Broadcom is already running AgentMinder in production internally. The system handles peak loads of approximately 43 million API calls per day, according to Broadcom’s internal AgentMinder deployment, with zero downtime during maintenance and upgrades.
Justin Giardina, CTO at 11:11 Systems — a managed services provider and VMware partner — called the agent gateway capabilities "a solution to a real problem today," noting their ability to monitor token usage and actual AI consumption in production environments.
AgentMinder runs alongside two additional agentic security layers announced Monday: Agentic Zero Trust with vDefend, which extends hypervisor-level lateral security to agentic workloads through continuous traffic monitoring, shadow AI detection, and AI-generated Intrusion Detection and Prevention (IDPS) signatures; and Agentic Threat Defense with Avi Load Balancer, which restricts agents from accessing unauthorized tools, isolates anomalous agent behavior in real time, and enforces data protection guardrails against unauthorized exfiltration. All of this is detailed in Broadcom’s agentic security announcement.
The Tanzu Platform’s deny-by-default agent containment architecture underpins all of this at the application layer. Agents deployed through Tanzu are granted zero access to APIs, external networks, MCP servers, or the internet unless explicitly authorized on a per-task basis. A new isolated credential store ensures that credentials are shielded from agent processes entirely — eliminating the attack surface for credential theft and prompt injection.
Tanzu Platform’s full agentic AI capabilities, including out-of-the-box agentic harnesses and a curated marketplace of AI models and data products, will be generally available in Fall 2026.
TrueSource: Open-Source Security at AI Speed
The third major announcement cluster addresses a category that has received less industry attention than inference infrastructure or agent governance: supply-chain security for the open-source software on which enterprise AI applications depend.
TrueSource by Broadcom consolidates Broadcom’s open-source security effort under a single brand. It operates by continuously scanning the Spring framework and its full dependency tree using frontier AI models, with human-verified patches delivered simultaneously across every active release line. In the five months prior to the conference, Broadcom engineers used more than 12 billion tokens scanning open-source dependencies — finding and fixing vulnerabilities before attackers could exploit them.
Two new services launched Monday alongside the existing Spring Enterprise offering: TrueSource Trusted Artifacts extends clean-room builds to the broader Java ecosystem, Python, and Node.js, as well as the Bitnami Secure Images catalog. TrueSource Data Services applies the same standard to the data tier — PostgreSQL, RabbitMQ, MySQL, and Valkey. Both are available now and are sold separately from VCF under tiered site licenses.
The timing matters: several critical VMware vulnerabilities disclosed earlier in 2026 — including CVE-2026-59310, a CVSS 9.8 directory-traversal flaw in vCenter that a China-nexus APT actor actively exploited across hundreds of systems within days of disclosure — have highlighted the urgency of both software-supply-chain hygiene and rapid patching for the VCF install base. TrueSource’s AI-generated IDPS signatures in vDefend represent Broadcom’s response to the speed at which AI-assisted attackers can now operationalize newly disclosed vulnerabilities.
What the AI Push Actually Means for VCF Customers
Broadcom’s decision to rebrand and unify its private AI offering under the "VMware Private AI Cloud" umbrella at this year’s Explore is not primarily a product story. It is a strategic retention story.
Broadcom completed its acquisition of VMware in November 2023 for approximately $61 billion in equity, assuming additional VMware debt. The integration that followed — collapsing VMware’s portfolio into VCF, eliminating perpetual licenses, shifting to subscription-only billing, and significantly raising prices — produced the most significant customer-satisfaction crisis in VMware’s history. A Gartner survey released in April 2026 found 76% of IT leaders hold a negative outlook toward Broadcom’s ownership of VMware — up from 64% in 2025 and 33% in 2024. Sixty-seven percent are actively seeking VMware alternatives. Gartner projects that by 2029, 55% of enterprises will have migrated 100% of their VMware workloads away from the platform.
Nutanix is winning hundreds of VMware customers quarterly — 500 to 1,000 new per quarter — with its CEO describing the majority as migrating directly from VMware. The competitive threat has a finite window — as VMware contracts expire and customers face renewal pricing — but the window is open now.
IDC analyst Stephen Elliot, who covers Broadcom VMware, put it directly in TechTarget’s coverage of VMware Explore: "It’s been a journey for customers, but they have seen the execution, engineering and evolution the acquisition has brought, while AI has put private cloud back in the spotlight."
That last clause is the key. AI inference has changed the calculus in Broadcom’s favor in a way that licensing disputes could not: if VCF becomes the infrastructure platform that makes enterprise AI inference operationally and economically viable — through NVMe memory tiering, MLPerf-validated performance, heterogeneous GPU support, and AgentMinder’s agent governance — then migrating away from VCF means rebuilding that AI infrastructure stack from scratch on a competitor platform. VMware Private AI Cloud is Broadcom’s attempt to make the transition cost of leaving higher than the transition cost of staying.
Ryan Sheehan, SVP of advanced solutions at SHI International, captured the architectural reality in Network World’s VMware Explore coverage: "They’re at the hypervisor. They are the hypervisor." That proximity to enterprise data — and now to enterprise AI workloads running on that data — is the competitive moat Broadcom is betting on.
Flexential, a colocation provider and VMware partner, is already building private AI offerings on VMware AI Factory for release early next year, according to Craig Cook, its SVP of cloud and data protection. "Basically, you can do everything there," Cook told Network World. "You can bring your own model. You can spin up your agents. You can orchestrate it. And they’ve got observability, which is critical."
IDC analyst Matthew Flug added nuance: Broadcom’s on-premises focus creates real opportunities for vendors like IBM/Red Hat and Nutanix in hybrid cloud orchestration — scenarios where organizations want to govern AI workloads that span both on-premises VCF and public cloud endpoints.
Conference Program: 400+ Sessions Across Four Tracks
Tuesday’s program at VMware Explore 2026 includes sessions across all four content tracks: Cloud Infrastructure (VCF platform architecture, vSphere, vSAN, NSX, memory tiering, Kubernetes), Security (Zero Trust, vDefend, AI-powered threat defense, NIST CSF 2.0 compliance), App Modernization (Tanzu Platform, containerization, vSphere Kubernetes Service, platform engineering), and Innovation (AI infrastructure, GPU management, agentic application development, model deployment).
Highlighted Tuesday sessions include a deep-dive walkthrough of vSphere innovations in VCF 9.1, featuring technology updates, performance improvements, and management simplifications with hands-on demonstrations in the VMware Explore 2026 session catalog; a customer fireside chat on lessons learned deploying agentic AI workloads in production, including failures and recovery; and "Platform Engineering 2.0: Building the AI-Native Foundation with VCF," addressing what platform engineering means when 80% of enterprises now maintain dedicated platform teams and 57% are actively adding AI capabilities to existing applications.
Wednesday and Thursday sessions cover VCF upgrade migration pathways, hardware optimization strategies for constrained server environments, and — to close the week — a technical deep dive on GPU resource management within VCF 9.1 as AI inference workloads scale, presented by Frank Denneman and Shawn Kelly.
Hands-on Labs and Certifications
The Hands-on Labs program this year runs in four formats: expert-led workshops with full administrative access to live VCF 9.1 environments; self-paced labs available on demand at the HOL facility with no advance registration; a Hub tours and Command Center Wall program that runs 30-to-45-minute behind-the-scenes tours of the conference’s own live infrastructure at 10:00 AM and 3:00 PM daily; and the Odyssey Tournament, a gamified lab competition with a live leaderboard. Three brand-new VCF 9.1 lab modules are live at the conference: the redesigned VCF Operations interface, memory tiering configuration and observability, and VKS Kubernetes updates, as detailed in the Hands-on Labs program guide.
Every Full Event Pass holder receives a complimentary VMware Certified Advanced Professional (VCAP) exam voucher for use at the on-site Certification Exam Center. Five VCAP-VCF tracks are currently available (VKS, Automation, Operations, Networking, Storage). New role-based VCAP-VCF Administrator credentials and the AI Resilient Infrastructure Expert (ARIE) certification track were announced at the conference — the latter targeting Zero Trust enforcement and autonomous threat response in VCF environments. Broadcom’s Private Cloud Outlook 2026 identified AI infrastructure and operations as the top skills gap in enterprise IT, cited by 40% of respondents, with Kubernetes operations second at 37%.
Explore on Tour 2026–2027
Following the Las Vegas event, Broadcom will take a curated version of VMware Explore on the road under the Explore on Tour banner. The confirmed Explore on Tour schedule includes Mumbai (September 29–30, 2026), Singapore (October 1–2), Frankfurt (October 13–14), Tokyo (October 20–21), London (November 18–19), Washington, D.C. (December 8), and Sydney (March 3–4, 2027). Singapore, Frankfurt, and London will include additional partner training days.
Frequently Asked Questions
What is VMware Private AI Cloud, and how does it differ from what Broadcom offered before?
VMware Private AI Cloud is Broadcom’s unified platform for running AI inference workloads, autonomous AI agents, containers, and traditional VMs on a single, on-premises or private-cloud infrastructure. It builds on VMware Cloud Foundation but extends it upward into production AI: the VMware AI Factory automates the infrastructure path from bare-metal servers to running models, NVMe memory tiering reduces the hardware cost of AI deployment during a period of acute DRAM price increases, and AgentMinder provides a governance control plane for autonomous agents. The key difference from prior offerings is consolidation — Broadcom is positioning this as one platform rather than requiring separate purchases for Private AI Foundation, model serving, and security tools. Read the official VMware Private AI Cloud announcement for the full product details.
How does NVMe memory tiering actually reduce costs for enterprise AI servers?
NVMe memory tiering works by using high-performance enterprise NVMe solid-state storage as a secondary memory tier below DRAM. The ESXi hypervisor identifies which memory pages are being actively accessed (hot pages, kept in DRAM) and which are rarely accessed (cold pages, stored on the NVMe tier). Since AI inference workloads load model weights into memory once and query them repeatedly, a large portion of server memory is cold at any given time — making this a practical fit. The default configuration doubles a host’s effective memory capacity without requiring additional DIMMs. In practical terms: a server that would otherwise need twice as much DRAM to handle a given AI inference workload can instead use a fraction of that DRAM paired with cheaper NVMe storage. With DDR5 prices spiking fivefold in the past year — from roughly $95 to over $500 per 32 GB module — this matters significantly to the on-premises cost case.
If I am evaluating whether to stay on VMware or migrate, what does this week’s announcement change?
VMware Explore 2026’s announcements are designed specifically to change this calculation. Broadcom is adding capabilities — NVMe memory tiering, integrated model serving, AgentMinder, vDefend agentic security — that are not available as a turnkey stack on Nutanix, Hyper-V, or most public cloud alternatives. If your organization is planning significant production AI inference investment over the next two to three years, the question is whether the cost of staying on VCF (including the pricing premium that followed the Broadcom acquisition) is lower than the cost of rebuilding comparable AI infrastructure governance on a different platform. A Gartner survey from April 2026 found 76% of VMware customers hold negative views of Broadcom ownership, and 67% are seeking alternatives — but 35% have already committed to migration. The decision is highly dependent on your current contract terms, renewal timing, and how deeply your AI strategy is tied to your existing VCF infrastructure.
Are AI agent governance tools like AgentMinder specific to VMware, or is this a broader industry direction?
This is a broader industry direction that Broadcom is trying to establish a leading position in. NIST launched an AI agent identity initiative in February 2026, focused on applying enterprise identity standards to autonomous agents, with a public comment period. Microsoft introduced Entra Agent ID and Conditional Access policies for AI agents in early 2026. BeyondTrust and Zscaler have published frameworks for least-privilege access and Zero Trust enforcement for non-human AI actors. The core principle — treat each AI agent as a unique enterprise identity with time-bounded, task-scoped permissions and full audit logging — is becoming a consensus framework. AgentMinder is Broadcom’s infrastructure-native implementation of that principle, embedded at the hypervisor and Tanzu Platform layer rather than as a separate identity management overlay.
Enjoyed this article? Sign up for our newsletter to receive regular insights and stay connected.