# 专题日跟踪 · 端侧推理 · 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。** 端侧推理口径继续吃 2026-03 华为 A1000、2026-05 VCF Edge 9.1、2026-06 Scale Edge GenAI、以及 NVIDIA Orin Nano 2（H1 2027）这几份最强公开页。

- **VCF Edge 9.1（2026-05-05）**：边缘可跑 GPU 推理或 **llama.cpp CPU 推理**；不必每站一张卡。许可地板仍是 **≥10 站点 / ≤256 核每站**。气隙用 pull 内容库。
- **Scale Edge GenAI（2026-06-08）**：站点节点做推理+检索+动作；训练留中心。HyperCore 仍 **未点名 vLLM/Ollama**——有架构文、无 serving SKU。2 节点+本地 witness 是 HA，不是模型产品。
- **华为 FusionCube A1000（2026-03-17 数据存储春发）**：面向 **分支边缘推理** 的一体机，边云协同、模型下发、多模型切分。与 **1000H 鲲鹏通用 HCI** 分列。1000H 官网本日 WebFetch **Akamai Access Denied**。
- **Red Hat MicroShift 4.20**：OpenShift AI 的 ServingRuntime / InferenceService 可落到 Device Edge；Jetson 路径指向 NVIDIA 文档。
- **OpenNebula**：Ampere ARM64 2 节点 + Ray/Hugging Face appliance（Qwen 2.5 demo）——CPU 边缘小模型，不是 NVL72。
- **NVIDIA Jetson Orin Nano 2**：78 TOPS、约 2× 推理或 15 W 同性能 -40% 功耗，**厂商自称**；**H1 2027 上市**。跑的是设备内 SLM/VLM，不是 HCI。
- **无公开端侧推理 SKU：** RiverMeadow、SmartX、ZStack（仅智算 AIOS）、VergeIO、k0rdent、Rafay（无门店页）、Vates、Harvester、Platform9、HPE VME、Proxmox。深信服是「分支加 GPU 节点 + AICP」，不是模组。Qualcomm/Apple NPU 只作行业对照。

规划即时含义：边缘与核心推理必须两张 SKU；默认 7B–14B 量化 + CPU 或单卡；舰队下发/回滚先于自研引擎；禁止把 A1000/TokenBox/Jetson 写进 1000H/FusionOne HCI/VME 行。

## 原文摘录

### [Announcing VMware Cloud Foundation Edge 9.1](https://blogs.vmware.com/cloud-foundation/2026/05/05/announcing-vmware-cloud-foundation-edge-9-1-a-scalable-autonomous-edge-platform/)

- 来源：VMware Cloud Foundation Blog / 2026-05-05

> GPU availability, cost, and power constraints often limit where AI can be deployed. VCF Edge enables AI inference to run alongside existing workloads using GPUs or CPU-based inference (via llama.cpp).
>
> VCF Edge enables AI inference on standard CPU infrastructure using llama.cpp, reducing reliance on GPUs and enabling AI to run in constrained or remote edge environments where GPU deployment is not practical.
>
> This architecture is specifically engineered to enable reliable operations in low-connectivity, disconnected, or fully air-gapped environments by allowing each site to host and manage its own state locally.

### [Edge GenAI: Bringing Generative AI Closer to Data and Devices](https://www.scalecomputing.com/resources/edge-genai-bringing-generative-ai-closer-to-data-devices)

- 来源：Scale Computing / 2026-06-08

> Edge GenAI is the practice of running generative AI capabilities close to where data is created—at a site, device, or local edge node—so organizations can generate answers, summaries, and actions without always depending on a round trip to a centralized cloud.
>
> What Edge GenAI usually does not mean is training large models from scratch at every site. Training and heavy fine-tuning are generally centralized…
>
> On-site edge node: The most common pattern. A shared inference and retrieval service supports many devices at a site…
>
> For any “write action,” use allowlists and explicit approval steps. For example, drafting a ticket can be automatic, but submitting it to the system of record may require a human confirmation for the pilot phase.
>
> Version everything: models, prompts, retrieval indexes, and action connectors. Roll out in stages, monitor impact, and keep a rollback plan

### [华为发布面向中心推理与分支边缘推理场景的全新AI数据基础设施](https://www.doit.com.cn/p/554988.html)

- 来源：DOIT 转述华为数据存储新春发布会 / 2026-03-18（官方独立产品页本日未抓到，规划引用须再核华为原文）

> 3月17日，在2026华为数据存储新春发布会上，华为正式发布针对AI推理场景的全新AI数据基础设施，包含面向中心推理场景的AI数据平台，和面向分支边缘推理场景的FusionCube A1000 AI超融合一体机
>
> FusionCube A1000采用通算与智算融合一体化架构，实现AI部署“开箱即用”。
>
> 分支边缘侧产生的数据将回传至中心，在中心侧不断丰富语料、完成模型增训，模型更新后可一键下发至边缘侧
>
> 同时，FusionCube A1000支持算力的动态切分与调度，可同时运行多个不同的AI模型，并依据业务优先级智能分配资源，单卡算力利用率提升约30%。

### [NVIDIA Announces Jetson Orin Nano 2](https://nvidianews.nvidia.com/news/nvidia-announces-jetson-orin-nano-2-robotics-computer-to-redefine-entry-level-edge-ai)

- 来源：NVIDIA Newsroom

> Featuring 78 trillion operations per second of AI compute, 8GB of memory and an 8-core Arm CPU…
>
> Jetson Orin Nano 2 achieves 2x the inference performance of Jetson Orin Nano Super through improved Tensor Cores and higher memory bandwidth, while maintaining the same compact form factor. In 15-watt mode, Jetson Orin Nano 2 consumes 40% less power to deliver the same performance as its predecessor.
>
> The NVIDIA Jetson Orin Nano 2 module and developer kit are expected to be available in the first half of 2027.

### [Using Red Hat OpenShift AI with MicroShift](https://docs.redhat.com/en/documentation/red_hat_build_of_microshift/4.20/html/using_ai_models/microshift-rh-openshift-ai)

- 来源：Red Hat MicroShift 4.20 文档

> You can use Red Hat OpenShift AI (Red Hat OpenShift AI Self-Managed) to integrate a fleet of MicroShift-driven edge devices into the artificial intelligence and machine learning (AI/ML) operations cycle.
>
> To enable GPU or other hardware accelerators for MicroShift, follow the guidance specific for your edge device… For example, to use an NVIDIA GPU accelerator, begin by reading the following NVIDIA documentation: Running a GPU-Accelerated Workload on Red Hat Device Edge

### [AI Inference at the Edge with Ampere ARM64 and OpenNebula](https://opennebula.io/blog/product/ai-inference-with-ampere/)

- 来源：OpenNebula Blog / 2025-05-28

> OpenNebula enables deployment on a compact, 2-node ARM64-based edge cluster, supporting AI inference workloads with Hugging Face models and the Ray framework.
>
> The included chatbot interface enables real-time interactions with the Qwen 2.5 model…

## 对专题档案的影响

- VCF Edge 的 llama.cpp CPU 路径必须写入「无 GPU 站点也能推理」，同时保留 10 站点地板。
- Scale 有 GenAI 架构、无引擎 SKU，对照表不得填 vLLM。
- FusionCube A1000 与 1000H 分列；A1000 数字均为发布会/转述，标厂商自称。
- Orin Nano 2 是 2027 H1 模组，不能当 2026 出货 HCI 附件。