7 min read

VMware Cloud Foundation 9.1 is here, and it has one clear goal: to run AI workloads in your own data centre rather than in the public cloud. With support for the latest NVIDIA and AMD accelerators, Kubernetes control for up to 500 clusters, and built-in data-sovereignty features, the release directly addresses the DACH region’s concern over who actually retains control of the data-and the uncomfortable fine print sits in the licence agreement.

Related:800-volt DC for AI racks  /  Data-centre energy balance and the EnEfG

What VCF 9.1 delivers technically

What is VMware Cloud Foundation? VCF is an integrated platform that bundles compute, storage, networking and management into a private-cloud stack. It runs in your own data centre and manages virtual machines, containers and Kubernetes through a single control plane. Version 9.1 has been available since May.

The most compelling innovations sit where AI meets infrastructure. ESXi 9.1 now supports NVIDIA ConnectX-7 and BlueField-3 with Enhanced DirectPath, plus AMD’s MI350 accelerator. In plain terms, the hardware you need for AI training can now be passed through with near-native performance-a long-standing weak spot in virtualised environments. VMware has finally closed that gap.

On the Kubernetes side, the platform can govern up to 500 clusters through a single control plane, provisioning workloads up to 70 % faster thanks to linked-clone technology. Virtual machines, containers and Kubernetes coexist under the same management umbrella. Memory Tiering adds NVMe SSDs as a second storage tier, boosting VM density and cutting total cost of ownership. These aren’t marketing talking points; they’re the dials an ops team actually turns in day-to-day operations.

Why this is a sovereignty issue in the DACH region

The real lever for German, Austrian and Swiss operators isn’t GPU support. It’s the sovereignty features. VCF 9.1 introduces data-residency controls and TPM-backed security. For regulated sectors-from banking and insurance to healthcare-this is the prerequisite for running AI on sensitive data at all.

The public cloud remains a gray area for many of these workloads. When you send patient records or credit files through an AI model, you need to know where the data resides and who has legal access. A private platform with built-in residency controls shifts that question from an act of trust to a technical specification. That’s exactly what NIS2, DORA and the EU AI Act demand in practice.

With this, VCF positions itself as the answer to a question preoccupying many DACH companies right now: how to run AI without surrendering control of your data. The platform delivers the technical foundation. It doesn’t absolve you of the strategic choice over what belongs in-house and what can safely stay in the public cloud.

The trade-off nobody talks about

Technically, VCF 9.1 is a strong release. The honest conversation starts with the price tag. Since the Broadcom acquisition, VMware’s licensing model has shifted noticeably-bundling, subscription mandates and higher costs have left many existing customers weighing stay-or-migrate decisions. If you’re planning VCF 9.1 as your private-AI platform, you’re budgeting for that licensing model, not just the technology.

A platform is chosen not by the spec sheet but by the five-year cost. VCF 9.1 wins the spec sheet. The contract, everyone must calculate for themselves.

The math isn’t the same for everyone. A company with a large VMware footprint and regulatory pressure for data sovereignty finds VCF 9.1 the obvious path because migration plus sovereignty would cost more than the subscription. A mid-market firm without heavy legacy investment runs the numbers differently and may eye Kubernetes-native alternatives without the VMware stack underneath. Both routes are valid. The mistake is buying the technology and being blindsided by licensing costs.

What remains is a sober assessment. VCF 9.1 makes private AI infrastructure more mature and operable than it was a year ago. It convincingly solves the sovereignty challenge at the technical layer. Whether it’s the right pick isn’t decided by the features, but by what your organization will pay for this stack in five years-and whether you’re willing to pay it.

Frequently Asked Questions

Is VCF 9.1 worth it for AI workloads in your own data center?

Technically, yes. With support for NVIDIA ConnectX-7, BlueField-3, and AMD MI350, as well as Enhanced DirectPath, virtualized AI training with nearly native performance becomes feasible. However, the decision heavily depends on the licensing model.

What do the sovereignty features actually deliver?

Data-residency controls and TPM-backed security enable running AI on sensitive data without ceding storage location and access to an external provider. This directly addresses requirements from NIS2, DORA, and the EU AI Act.

How many Kubernetes clusters can VCF 9.1 manage?

Up to 500 clusters via a single control plane, with up to 70 percent faster provisioning compared to the previous generation. Virtual machines, containers, and Kubernetes are all managed under one roof.

What’s the biggest catch?

The Broadcom licensing model. Bundling and mandatory subscriptions have pushed costs up for many existing customers. If you’re planning VCF 9.1, run a five-year total-cost-of-ownership calculation-not just the technical upsides.

Are there alternatives without the VMware stack?

Yes. Kubernetes-native private-cloud approaches that skip the underlying VMware stack are especially attractive for companies without a large legacy footprint. The effort then shifts from licensing to building your own platform expertise.

Image source: AI-generated (May 2026), C2PA certificate embedded in image