1 SEP 2026 — Broadcom's VMware division has launched VMware AI Factory, including a version built for AMD Instinct GPUs and the ROCm software stack. It is a reference architecture, not a product you buy, and it requires VMware Cloud Foundation underneath.
What was announced
The announcement came at VMware Explore on 31 August. An AI factory, in this usage, is a reference architecture — a vendor-validated description of the hardware and software needed to run inference workloads.
VMware's version automates hardware provisioning, software stack enablement and lifecycle management. It deploys a model once and shares it across users, tenants or business units, spans on-premises and cloud infrastructure while deciding where jobs should run, and adds sandboxes and governance controls to limit what agents can reach. It requires VMware Cloud Foundation and runs on servers from Cisco, Dell, Lenovo and Supermicro.
Borrowing a competitor's vocabulary
AI factory is Nvidia's phrase, used since 2023, and VMware has applied it to an AMD-based offering. That says something about how vendor language works.
A term coined to describe one company's stack becomes the generic name for the category, and competitors then have to use it or invent something nobody recognises. Nvidia gets the branding win and loses the exclusivity, which is the ordinary fate of a successful coinage.
The precedent here is VMware's own: in 2023 it announced VMware Private AI Foundation with Nvidia. The AMD version is the same shape with a different accelerator, and the naming makes the substitution explicit rather than hiding it.
A reference architecture is a document
The distinction matters for anyone budgeting from this announcement.
A reference architecture tells you which servers, GPUs, drivers, network configuration and software versions have been tested together, and it removes a large amount of integration risk. That is valuable, because the failure mode it prevents — a stack that works in isolation but not in combination — costs weeks of integration time.
It does not arrive as a single product you can install. Buying into this means acquiring the hardware from one of the named server vendors, the AMD accelerators, and the VMware licences, then following the design. The automation described sits inside the VMware layer you were already licensing.
ROCm support is bounded by what it runs on
Supporting AMD's open ROCm ecosystem is presented as openness, and it is, at one layer.
The requirement is VMware Cloud Foundation. The choice, then, is an alternative accelerator inside a proprietary platform, not an alternative to the platform itself. An organisation that wanted to avoid vendor concentration by moving off Nvidia arrives at a design where the concentration has shifted to Broadcom instead.
That is a legitimate offering for an existing VCF customer, which is who it is for. It is not a route out for anyone else, and the openness in the announcement describes the GPU layer only. We reported last week that Broadcom's TrueSource announcement covered open source it mostly already maintains, without stating a price. Both announcements share a shape, offering an open-sounding commitment whose value depends on unpublished commercial terms.
The governance features are the underrated part
Buried in the feature list are sandboxes and access controls for AI agents. These address a problem most organisations have not yet formalised.
An agent with tool access is a program acting with a user's permissions, often across systems that were never designed to distinguish a person from a process holding that person's credentials. The controls that exist for human accounts — approval steps, session limits, separation of duties — mostly do not apply, because the agent authenticates as the human.
Putting that in the virtualisation layer is a defensible place for it. The hypervisor already mediates what a workload can reach, already enforces tenancy, and is already audited. Extending those mechanisms to agent workloads is more plausible than each application inventing its own limits.
Whether it works depends on details the announcement omits, such as what the sandbox isolates, whether governance is enforced or advisory, and how it handles an agent's existing credentials. Those are the questions to ask a sales engineer, and they are more consequential than the accelerator choice.
The missing number is cost
Prashanth Shenoy, vice president of product marketing at Broadcom's VCF division, describes the offering as a full-stack automated operational system that treats AI token generation as a continuous production pipeline.
The pitch is about cost: deciding where jobs run and which models minimise expense. No figures accompany it — no benchmark, no cost-per-token comparison, no example workload. For a product positioned on economics, that is the missing number a buyer would want.
The claim is also structurally hard to verify, because savings from routing depend entirely on the mix of workloads and the price of the alternatives at a given moment. A vendor cannot honestly quote one figure for that. It could publish a methodology and a worked example, and has not.
Why AMD in a rack matters here
For data centre operators across this region the interesting part is not the VMware layer but the signal about accelerator supply.
Nvidia allocation has been the binding constraint on regional AI buildouts, with delivery times rather than prices determining what gets built. A validated design for AMD Instinct in an enterprise virtualisation platform is one more route to capacity that does not queue behind the same allocation.
The caveat is the software. ROCm has improved substantially and the ecosystem assumption is still CUDA: most published training recipes, most inference optimisations and most of the tooling. An operator choosing AMD accepts a smaller pool of engineers who have run it in production, and that constraint is felt more sharply in a regional market than in one where those engineers are abundant.
View author profile → · Editorial policy
