VMware Private AI Foundation with NVIDIA is a joint platform from Broadcom and NVIDIA that helps enterprises run generative AI securely in their own data centers, instead of relying only on public cloud services.

Built on VMware Cloud Foundation (VCF), it combines:

• NVIDIA AI Enterprise – a cloud-native software platform for building and deploying production-grade AI (including generative AI, computer vision, and speech AI).
• NVIDIA NIM – microservices for high-performance AI inference across data centers, clouds, and workstations.
• NVIDIA LLMs and community models – including models from ecosystems such as Hugging Face.
• Private AI Package – with vector databases, deep learning VMs, data indexing and retrieval, and an AI Agent Builder service.

With this stack, organizations can:

• Run RAG (retrieval-augmented generation) workflows using vector databases powered by pgvector on PostgreSQL.
• Fine-tune and customize LLMs for specific industries and use cases.
• Run inference workloads on-premises while maintaining control over data, cost, and performance.

Because it runs on VCF, the platform delivers public cloud–like scale and agility with private cloud–level security, resilience, and performance, helping teams reimagine how they deploy AI while keeping a lower overall total cost of ownership (TCO).

The platform is designed specifically to keep AI workloads close to your data while enforcing strong security and governance controls.

Key privacy and security capabilities include:

• On-premises deployment: You run training, fine-tuning, and inference in your own data center, which helps keep proprietary and regulated data under your control.
• Model Store with RBAC: ML Ops teams and data scientists can curate and publish LLMs in a Model Store with integrated role-based access control (RBAC), improving governance and protecting enterprise IP.
• Air-gapped support: Through VCF automation, the platform can be deployed in air-gapped environments, providing data isolation for highly sensitive workloads.
• AI software security patching: NVIDIA AI Enterprise customers receive regular security updates—monthly patches for critical and high CVEs on production branches and quarterly patches for long-term support branches, while maintaining API compatibility.

VCF adds additional security and compliance controls:

• Workload security: Features such as Secure Boot, Virtual TPM, vSphere Trust Authority, and VM Encryption help protect AI VMs and data.
• Identity and access management: Integration with VMware Identity Manager and third-party identity providers ensures only authorized users and applications can access models and datasets.
• Network security: Micro-segmentation, full-stack network security, and software-based firewalls help protect AI applications and their data at the network level.
• Multi-tenancy: With VCF 9.0, you can create secure, private environments for multiple tenants on the same infrastructure, which is useful for service providers or large enterprises with multiple business units.

Together, these capabilities help organizations meet privacy, security, and compliance requirements while still moving forward with AI initiatives.

The platform is built to make AI infrastructure easier to operate while keeping performance high and costs under control.

Infrastructure simplification and cost optimization:

• Unified private cloud platform: VMware Cloud Foundation provides a full-stack, software-defined platform for AI and non-AI workloads, so IT teams manage a single, unified environment instead of multiple siloed stacks.
• AI Blueprints Quick Start: Line-of-business admins can quickly design and publish infrastructure catalog items via VCF’s self-service portal, simplifying Day 0 and Day 1 deployment of AI workloads.
• vGPU profile visibility: Admins can see all vGPUs across the GPU footprint in a single vCenter UI, reducing manual tracking and saving admin time.
• GPU and vGPU monitoring: Host, cluster, and VM-level GPU monitoring helps identify over-provisioning or under-utilization, which supports better capacity planning and TCO optimization.
• Distributed Resource Scheduler (DRS): Automatically places workloads on the right hosts to balance performance and cost across clusters.

Performance and model lifecycle efficiency:

• Near bare-metal performance: A benchmark using MLPerf Inference v5.0 showed performance similar to bare metal, so you gain virtualization benefits without a major performance trade-off.
• Vector databases for RAG: Managed vector databases (via pgvector on PostgreSQL and Data Services Manager) make it easier to deploy retrieval-augmented generation applications without building the data layer from scratch.
• Model Runtime service: Data scientists can create and manage model endpoints for applications, simplifying scaling and operationalizing LLMs.
• Agent Builder Service: GenAI developers can build AI agents that use the Model Store, Model Runtime, and Data Indexing and Retrieval Service, speeding up application development.
• NVIDIA NIM microservices: Prebuilt containers support a wide range of models—from open-source community models to NVIDIA AI Foundation and custom models—streamlining deployment across clouds, data centers, and workstations.

By combining these capabilities, VMware Private AI Foundation with NVIDIA helps organizations rethink how they deploy AI: they can keep workloads on-premises, maintain strong control over resources, and manage TCO while still giving teams the flexibility and performance they need.