# 专题日跟踪 · 推理 · 2026-09-03

> 日期：2026-09-03（Asia/Shanghai）。原文摘录保持源语言。不改写数字。

## 摘要

**本日无公开战略级新稿。** 最强近期公开页仍是 2026-08 的推理 SKU 墙与引擎分层，规划按这些页锁口径，不必等「今天」的新闻。

- **Nutanix NAI 2.8（2026-08-26 GA）**：Private Inference 上 **<8B LoRA 微调 GA**；**100B+ 多节点流水线并行 = Tech Preview，官方禁止生产**；KV cache 卸主机内存同为预览；NIM 气隙导入 GA；可指定 vLLM 版本/参数。投机解码最多 **2.5× 为厂商自称**。
- **VMware Explore / AI Factory（2026-08-31）**：GPU 池化 + 模型廊 **厂商自称 150+**；文档（2026-07-15 更新）写死引擎：GPU 补全 **vLLM**、Embedding **Infinity**、CPU **llama.cpp**；NIM 走 NVAIE 另购。能力锁 **VCF Private AI**。
- **Red Hat llm-d（开发者文 2026-08-05；AI Inference 3.5 手册）**：KServe `LLMInferenceService` + 路由器；vLLM/SGLang 后端；**prefill/decode 分离** 走 NIXL。这是 OCP 发行版能力，不是 HCI 盒。
- **国产 MaaS 已产品化：** SmartX 榫卯 AI 平台 1.1（2026-07-22）vLLM/SGLang + 网关配额；ZStack AIOS 模板含 vLLM/SGLang/MindIE + Zentrix 网关（2026-08-19，明文不训练）；深信服 HCI+AICP（加 GPU 节点，对照 Ollama 的倍数为 **厂商自称**）。
- **品牌陷阱：** FusionCube 1000H 页无 MindIE SKU（本日官网 Akamai Access Denied）；FusionOne HCI 页无 vLLM，FusionOne AI/TokenBox（2026-05-20）才是推理软件。HPE VME ≠ Private Cloud AI。
- **无原厂 serving SKU（可写 BYO）：** RiverMeadow、Proxmox、Vates、Harvester、Platform9、Scale HyperCore、VergeIO（仅 GPU 抽象）。k0rdent/Rafay/OpenNebula 卖的是 GPU 云/NIM appliance，不是 ESXi 替身。

规划即时含义：对标表按 A 只给卡 / B 引擎模板 / C 目录+网关+Token 拆行；小模型 GA 与大模型预览必须分列；网关与引擎不要做成同一不可拆 SKU，但必须同入口。

## 原文摘录

### [Charting the Path to Governed Agentic AI with Nutanix Enterprise AI 2.8](https://www.nutanix.com/blog/charting-the-path-to-governed-agentic-ai-with-nutanix-enterprise-ai-2-8)

- 来源：Nutanix Blog / 2026-08-26

> Fine-Tuning: tune smaller models (<8B parameters) with organization and domain-specific data to match the accuracy of large LLMs for targeted use cases, maximizing resource efficiency and controlling GPU spend.
>
> Air-gapped NVIDIA NIM support: enable highly regulated customers with NVIDIA AI Enterprise (NVAIE) licenses to deploy the latest NVIDIA models via NVIDIA NIM microservices in air-gapped environments with operational simplicity.
>
> Multi-Node & Multi-GPU Inference (Tech Preview): serve frontier-class models (100B+ parameters) with pipeline parallelism to run inference across multiple GPU nodes when single-node memory isn’t enough.
>
> kvCache Offloading (Tech Preview): maximize GPU efficiency by moving KV cache context from GPU memory to CPU host memory when capacity is reached…
>
> *Tech Preview indicates these features should not be used in production environments.

### [Nutanix Enterprise AI (NAI) 2.8 is Now Available!](https://next.nutanix.com/nutanix-agentic-ai-180/nutanix-enterprise-ai-nai-2-8-is-now-available-45852)

- 来源：Nutanix Community / 2026-08-26

> Fine-tuning for models under 8B parameters is now Generally Available…
>
> Nutanix Enterprise AI enables users to leverage their preferred version of vLLM and custom parameters…
>
> Multi-node inference utilizes pipeline parallelism to serve frontier-class models with 100B+ parameters that exceed the physical VRAM limitations of a single node.
>
> Note: Tech Preview features should not be deployed in a production environment.

### [Nutanix Gives Enterprises the Freedom to Run Production Agentic AI Their Way](https://www.nutanix.com/press-releases/2026/nutanix-gives-enterprises-the-freedom-to-run-production-agentic-ai-their-way)

- 来源：Nutanix Press / 2026-08-26

> Speculative decoding which accelerates LLM inference token generation by up to 2.5x using lightweight draft models, cutting output latency without sacrificing model accuracy.
>
> Parameter-Efficient Fine-Tuning which supports Low-Rank Adaptation (LoRA) fine-tuning for smaller models (<8B parameters)… using single-GPU compute

### [Running Completion or Embedding Models by Using Model Endpoints](https://techdocs.broadcom.com/us/en/vmware-cis/private-ai/foundation-with-nvidia/9-0/private-ai-foundation-9-x/what-is-private-ai-services/deploying-model-endpoints.html)

- 来源：Broadcom TechDocs / Last Updated July 15, 2026

> You run models on GPU by using the open-source vLLM inference engine for completion models.
>
> You run embedding models by using the open-source Infinity inference engine on GPU or CPU.
>
> You run embedding or completion models on CPU by using the open-source llama.cpp inference engine. Models must be available in the GGUF file format.

### [Broadcom Announces VMware AI Factory…](https://www.globenewswire.com/news-release/2026/08/31/3353363/0/en/broadcom-announces-vmware-ai-factory-enabling-faster-time-to-production-ai-and-greater-control-over-ai-tokenomics.html)

- 来源：GlobeNewswire / LAS VEGAS, Aug. 31, 2026（本日对该 URL 的 WebFetch 超时，摘录依据检索与次级报道核对）

> VCF pools and shares GPU resources across the organization so teams can run multiple models on shared hardware instead of dedicating infrastructure to each workload. A unified model gallery gives IT and data science teams a single interface for deploying and managing model inference and RAG workflows… with built-in observability into token throughput, latency, and compute and memory utilization.
>
> VCF customers can run more than 150 open source and commercial models, including Nemotron 3, Gemma 4, cotomi, Qwen 3.7-Max, and GLM 5.2.

### [榫卯AI平台1.1](https://www.smartx.com/blog/2026/07/ai-platform-1-1/)

- 来源：SmartX / 2026-07-22

> 榫卯 AI 平台 1.1 在已有的 GPU 直通和逻辑隔离的能力基础上，引入隔离性更强的原生 GPU 虚拟化（vGPU）支持能力。
>
> 榫卯 AI 平台是榫卯企业云基础设施面向 AI 推理场景的重要组件，聚焦算力模型管理、模型治理和高性能推理

### [ZStack Zentrix 首发](https://www.zstack.io/brand-news/news/product-news/6446.html)

- 来源：ZStack / 2026-08-19

> ZStack Zentrix 是一款面向企业的 AI 网关，为模型、MCP 工具、Agent 和 Skill 提供统一接入与治理能力。
>
> Zentrix 不负责训练模型，也不替代业务系统的 Agent 编排

### [ZStack AIOS 产品功能 · 推理模板](https://docs.zstack-cloud.com/en/docs/ai/aios/cloud-user-guide/topics/%E4%BA%A7%E5%93%81%E5%8A%9F%E8%83%BD)

- 来源：ZStack Cloud 手册

> 系统推理模板由系统预置，对应多种推理框架，包括：Transformers、llama、vLLM、Diffuser、Sentence Transformer、MindIE、SGLang等
>
> 支持单机多卡、多机多卡等多种推理模式。实际场景下可用的推理模式和部署模式 (基于云主机部署或基于容器部署) 及GPU型号相关

## 对专题档案的影响

- 对照表必须保留 Nutanix **<8B GA / 100B+ 预览** 原文墙，不得写成「已支持大模型」。
- VMware 引擎三件套（vLLM / Infinity / llama.cpp）写入验收，NIM 另购。
- FusionCube / FusionOne / VME 的 HCI 行不得填入各自 AI 品牌的引擎。
- 本日 globenewswire AI Factory 抓取超时、华为 1000H 页 Akamai 拒绝，引用时标注，不发明规格。