VMware Explore 2026 kicked off this week in Las Vegas, and Broadcom used the stage to announce VMware AI Factory, the software-defined foundation behind a bigger push called VMware Private AI Cloud. I’ve been buried in VCF 9.1 on my own R640s for the lab build, so this one caught my attention fast.
The pitch is simple. Enterprises want to run AI where their data already sits, but going from bare metal to a working model has been slow and expensive. Paul Turner, VMware Cloud Foundation’s chief product officer, put it plainly: the trip from metal to model is slow, complex, and expensive. AI Factory is Broadcom’s attempt to close that gap by automating infrastructure deployment and handling Day 2 operations, so teams get to a first deployed model faster and keep a handle on token costs as usage grows.
A few things stood out to me.
Bare-metal provisioning gets a serious speedup. Broadcom is partnering with MetalSoft to bring heterogeneous bare-metal automation into VCF. The claim is that provisioning drops from weeks down to minutes, with servers from multiple vendors handled through the VCF console instead of a pile of vendor-specific tools. Anyone who’s racked and re-imaged physical hosts by hand (guilty) knows how much time that eats.
Hardware choice stays open. AI Factory pairs VCF with certified AI ReadyNodes from Cisco, Dell, Lenovo, Supermicro and others, plus your accelerator of choice. Broadcom and AMD are also collaborating specifically on a stack pairing VCF with AMD Instinct GPUs and ROCm, which is a notable move away from the assumption that private AI infrastructure means Nvidia by default.
Model choice is wide too. VCF customers can reportedly run more than 150 open source and commercial models, including Nemotron 3, Gemma 4, Qwen 3.7-Max and GLM 5.2, delivered through VCF’s built-in services. That’s a lot of flexibility if you’re trying to avoid locking your whole AI strategy to one provider.
The cost story is the real target. Broadcom is framing VMware Private AI Cloud around three cost drivers: hardware CapEx, operational complexity, and tokenomics. VCF 9 leans on NVMe memory tiering and cluster-wide storage deduplication to bring hardware costs down, and adds token monitoring, multi-tenant model sharing, GPU/vGPU tracking, and an AI metrics observability dashboard to keep tokenomics in check once things are running.
There’s more around the edges too, a new agent governance layer, AI-ready data foundations in Tanzu, and agentic security work through vDefend and Avi Load Balancer. It’s clear Broadcom wants VCF to be the whole story for private AI, not just the virtualization layer underneath someone else’s AI stack.
For homelabbers, none of this is something you’re spinning up on a single R640 anytime soon, GPU-backed AI ReadyNodes and multi-tenant token tracking are squarely enterprise territory. But it’s worth watching, because the automation ideas (fast bare-metal provisioning, unified lifecycle management) tend to trickle down into the tooling the rest of us eventually get to use. I’ll be keeping an eye on how much of the AI Factory automation layer shows up in VCF releases outside the AI-specific SKUs.
Official announcement: Broadcom Introduces VMware Private AI Cloud