Broadcom announced VMware Cloud Foundation (VCF) 9.1, a private cloud platform rebuilt for production AI workloads, offering up to 40% lower server costs, tighter zero-trust security controls, and support for agentic AI alongside traditional enterprise applications on a single infrastructure layer.

The release arrives as private cloud gains ground for AI deployments. Broadcom’s Private Cloud Outlook 2026 survey found that 56% of organisations are running or planning to run production AI inference in private environments, while public cloud use for production inference dropped 15% year-on-year to 41%.

Cost and performance at the infrastructure layer

VCF 9.1 targets three cost pressure points that have slowed AI production rollouts:

• Up to 40% reduction in server costs through intelligent memory tiering for mixed AI and non-AI workloads
• Up to 39% lower storage total cost of ownership through enhanced compression and deduplication for AI data pipelines
• Up to 46% reduction in Kubernetes operational costs for AI workloads at scale

Cluster upgrade speed improves fourfold and fleet management capacity doubles to 5,000 hosts, addressing the operational bottleneck of scaling GPU-heavy infrastructure across distributed enterprise environments.

VCF 9.1 is a single unified platform that addresses data and IP privacy concerns, surging infrastructure costs, and readiness for agentic AI. We enable zero-trust security for AI, reduce costs through intelligent infrastructure optimization and hardware choice, and provide the flexibility to run both agentic workflows and accelerated inferencing on the same platform. – Krish Prasad, Senior Vice President and General Manager, VMware Cloud Foundation Division, Broadcom

Zero-trust extended to AI and Kubernetes workloads

VCF 9.1 extends zero-trust segmentation to Kubernetes and AI workloads for the first time, adding distributed IDS/IPS protection capable of 9 Tbps threat inspection performance. The platform also integrates with CrowdStrike Falcon for isolated ransomware recovery and environment validation, a notable pairing given the sensitivity of AI model weights and training data as enterprise intellectual property.

On-premises ransomware recovery allows organisations to restore AI models and training data without cross-border data movement or the bandwidth costs associated with cloud-based recovery. Zero-downtime live patching covers up to 80% of use cases without host evacuation, maintaining availability for inference services and agentic applications.

Open hardware ecosystem across AMD, Intel and NVIDIA

VCF 9.1 supports mixed CPU and GPU workloads across AMD, Intel Xeon 6, and NVIDIA Blackwell hardware, including NVIDIA ConnectX-7 NICs and BlueField-3 with Enhanced DirectPath I/O for high-speed AI model training and data transfer. The open hardware stance responds to the reality that agentic AI workloads are compute-intensive on the CPU side, while inference remains GPU-bound, meaning enterprises need a platform that can handle both on the same infrastructure without partitioning.