To create a safe haven for businesses that combines cost-effectiveness and data sovereignty, Broadcom announced the launch of products specifically designed for production-grade AI workloads.VMware Cloud Foundation (VCF) 9.1This newly upgraded private cloud platform not only deepens the integration of Kubernetes containers and traditional virtual machines, but also enables enterprises to deploy inference and agentic AI applications at a lower total cost of ownership by connecting with the open hardware ecosystems of AMD, Intel and NVIDIA.
The Rise of "Private AI": An Inevitable Backlash from Uncontrolled Public Cloud Costs
According to a preview of Broadcom's latest "Private Cloud Outlook 2026" report, market trends are undergoing a subtle reversal.
The report indicates that as many as 56% of the surveyed companies are currently running or planning to run production-grade AI inference workloads on "private clouds"; in contrast, only 41% are using public clouds for production-grade inference, a significant drop of 15% compared to the same period last year.
The core reason driving this wave of "AI returning home" is very practical: 62% of IT executives are "extremely concerned" about the infrastructure costs of generative AI on the public cloud, while 36% of executives are facing severe pressure regarding data privacy and regulatory compliance.
Krish Prasad, Senior Vice President and General Manager of the VCF division at Broadcom, stated that enterprises face three major challenges: data privacy, soaring costs, and insufficient preparedness for "proxy AI." VCF 9.1 is a single, unified platform designed to address these pain points.
VCF 9.1's three major upgrade highlights: maximizing hardware performance, zero-trust security, and an open ecosystem.
To address the challenges of large-scale enterprise deployment, VCF 9.1 features deep optimizations to its underlying architecture, focusing on the following three aspects:
• Ultimate optimization of resources and costs:
This is not simply a software upgrade, but a cost reduction that can be directly reflected in financial statements. Through advanced intelligent memory tiering technology, server costs can be reduced by up to 40% when running a mix of AI and non-AI workloads; while next-generation AI data pipeline compression and deduplication technologies can reduce total cost of ownership for storage by up to 39%.
Furthermore, through automated cluster operations (supporting up to 5000 hosts and up to 4 times faster cluster upgrade speeds), the maintenance costs of Kubernetes when running AI at scale can be reduced by as much as 46%.
• An "open ecosystem" that refuses to be tied to a single hardware component:
In an era where GPUs are scarce, VCF 9.1 offers a multi-accelerator option across NVIDIA and AMD, and fully supports mainstream Intel and AMD CPU platforms. Addressing the extreme network bandwidth demands of AI training, VCF 9.1 also integrates an NVIDIA ConnectX-7 network card and a BlueField-3 DPU, and supports the Arista open networking architecture, ensuring enterprises can freely choose the best hardware combination in its class.
• Zero-trust security from hypervisor to application:
Targeting the most sensitive models and training data in the AI era, VCF 9.1 introduces zero-trust lateral security protection (for the first time extending IDS/IPS to Kubernetes AI workloads) and has a built-in ransomware recovery mechanism (integrating with major cybersecurity companies such as CrowdStrike).
More importantly, it supports "zero-downtime instant patching" in up to 80% of scenarios, ensuring that production-grade AI services can meet enterprise-level SLA stability requirements.
