VMware Intros Private AI Cloud VMware推出私有AI云

TL;DR

• Enterprises are shifting AI workloads from public cloud to on-premises private clouds and "AI factories" due to cost, sovereignty, and security concerns
• Omdia predicts global datacenter investments will reach ~$1.6 trillion by 2030, with $600+ billion spent on AI infrastructure this year
• VMware and Broadcom are launching VMware Private AI Cloud and VMware AI Factory model-as-a-service to address the shift, built on VCF 9
• VMware's June survey of 1,800 IT decision-makers found 56% running or planning production inference in private cloud, while public cloud usage for AI workloads fell 15% year-over-year
• New security features include AI Gateway for unified governance, secure AI sandboxes for agentic workloads, and TrueSource for verifying open-source AI models

Why It Matters

The enterprise AI infrastructure landscape is undergoing a fundamental shift from public cloud dependency toward private, on-premises deployments driven by real cost and security pressures. For AI practitioners and infrastructure leaders, this signals that production-grade AI is no longer a cloud-only proposition, and organizations must plan for hybrid or private AI deployment strategies. The emergence of "AI factories" as a distinct datacenter model also redefines how enterprises think about scaling AI inference and training workloads.

Technical Details

• AI Factory Concept: Coined by Nvidia, AI factories are infrastructure designed for token production at their core, characterized by ultra-high capital intensity, geopolitical attributes, and complex engineering barriers. They are essentially supercomputers running GenAI inference to produce tokens as a revenue-generating or value-generating output.
• VMware Cloud Foundation 9 (VCF 9): Supports multi-vendor CPUs, GPUs, and accelerators including AMD Instinct MI350 Series GPUs with open ROCm platform. Features NVMe memory tiering (claimed 42% per-host cost reduction) and VMware AI Assistant for resolving complex infrastructure issues across CPUs, memory, and hypervisors.
• VMware AI Factory: Delivers an integrated package from metal to model, orchestrated via MetalSoft's platform. Includes VCF AI ReadyNodes from Dell, Cisco, Lenovo, and Supermicro. Supports over 150 open and commercial AI models including Nvidia Nemotron 3, Google DeepMind Gemma 4, Alibaba Qwen 3.7-Max, NEC cotomi, and Z.ai GLM 5.2.
• Security Architecture: Introduces AI Gateway for unified model governance across cloud and on-premises, secure AI sandboxes (virtualized container spaces for isolating agentic code execution), TrueSource by Broadcom for open-source AI verification, and vDefend enhancements for zero-trust security in agentic AI workloads. Model sharing between tenants uses isolated nameplates to eliminate redundant deployments.

Industry Insight

• The 15% year-over-year decline in public cloud usage for AI workloads suggests a structural shift, not a temporary correction. Organizations should evaluate private cloud and on-premises AI strategies now rather than reacting to cost overruns after deployment.
• Security remains the top concern for 51% of enterprises repatriating AI workloads. The sandbox breakout incidents involving major AI providers (OpenAI, Anthropic) make secure agentic AI execution environments a critical differentiator—invest in platforms with proven isolation guarantees.
• The "bring data to the model, not model to the data" principle is becoming the dominant paradigm for enterprise AI. Infrastructure vendors that simplify the full stack from hardware to model deployment (like VMware's AI Factory approach) will capture significant enterprise adoption as organizations seek to reduce the manual complexity of building AI infrastructure from scratch.

TL;DR

• 企业AI生产部署正从公有云向私有云和本地“AI工厂”迁移，核心驱动力为成本压力、数据主权与安全合规。
• Omdia预测2030年全球数据中心投资将达近1.6万亿美元，今年科技企业在AI基础设施上的支出将超6000亿美元。
• VMware联合Broadcom推出Private AI Cloud与AI Factory模型即服务，提供从底层硬件到AI软件栈的一体化集成方案。
• 内部调查显示56%企业计划或已在私有云运行生产级推理，公有云AI负载占比同比下降15%至41%。
• 新增安全能力聚焦Agent隔离沙箱、跨环境统一模型治理网关及开源模型溯源验证，应对Agentic AI运行时风险。

为什么值得看

本文清晰刻画了企业级AI从“公有云实验”向“私有化生产”转型的关键拐点，为云厂商、硬件OEM与IT决策者提供了明确的市场风向标。随着AI基础设施资本密集度飙升与安全合规要求收紧，掌握私有云AI工厂的构建路径与治理工具，将成为企业规模化落地生产级AI的核心竞争力。

技术解析

“AI工厂”由Nvidia首创，指以Token生成为核心、具备超高资本密集度与复杂工程壁垒的专用基础设施。VMware将其封装为模型即服务（MaaS），依托MetalSoft平台实现从物理服务器、网络存储到Kubernetes与AI软件栈的统一编排，大幅降低自建门槛。

VCF 9及AI ReadyNodes兼容多厂商CPU/GPU与服务器硬件，已集成Dell、Cisco、Lenovo、Supermicro等OEM方案，并深度支持AMD Instinct MI350系列GPU及开放ROCm生态，有效避免单一硬件绑定，提升供应链弹性。

成本优化方面，VCF 9引入NVMe内存分层技术，官方称可降低高达42%的单节点成本；模型库扩展至超150种开源与商业模型（含Nemotron 3、Gemma 4、Qwen 3.7-Max等），支持SLM与LLM混合部署，适配不同推理场景的算力需求。

安全架构针对Agentic AI的运行时风险进行专项设计：通过隔离命名空间实现多租户模型共享，AI Gateway统一治理云边模型调用，安全沙箱隔离Agent代码执行，并集成TrueSource开源溯源与vDefend零信任防护，防范模型逃逸与供应链攻击。

行业启示

公有云AI红利期正在消退，成本与合规压力将加速企业构建“数据不动模型动”的本地化AI基础设施，私有云将成为生产级推理与微调