# 推理专题日跟踪 2026-09-07

> 日期：2026-09-07（Asia/Shanghai）。专题正文：`intel/topics/inference.md`。

## 摘要

- **VMware Private AI Cloud / AI Factory**：调研称 **56%** 计划私有云生产推理；模型目录扩至 150+（Nemotron 3、Gemma 4、Qwen 3.7-Max 等——**厂商自称**）。
- **Nutanix NAI 2.8**：<8B 微调、NIM 气隙；多节点多 GPU 推理与 KV cache offload **tech preview**。
- **Rafay**：NVIDIA 认证 VMaaS + Token Factory 叙事（归档仍在）。
- **HPE / Dell**：推理订单与「inference past training」叙事抬高传统服务器+AI 服务器需求。
- **Hugging Face 交易稿**（谨慎）：称 Agent 已成为 Hub 第一用户——影响模型分发与私有镜像策略，**交易本身未独立核实**。

## 原文摘录

### VMware Private AI Cloud

- 归档：`industry/vmware-intros-private-ai-cloud-ai-factory-as-workloads-shift-to-on-pre-24f6b7e2.txt`

> … 56 percent saying that their organizations are running or planning to run production inference in a private cloud. … expanding to more than 150 the open and commercial AI models that can run on AI Factory …

### Nutanix Private Inference

- 归档：`industry/nutanix-expands-cloud-platform-for-production-agentic-ai-660eff4d.txt`

> … fine-tuning for models with fewer than 8 billion parameters and support for deploying NVIDIA NIM microservices in air-gapped environments. Multi-node and multi-GPU inference … is in tech preview. KV cache offloading … is also in tech preview.

### Dell inference demand

> … Inference is past training, and it's pure demand in our industry. We think the tokens that inference drives is going to grow 87 times to 3,600 quadrillion tokens by 2030.（**厂商展望，非独立核实**）

## 对专题档案的影响

- `inference.md`：补 NAI tech preview 边界与 VMware 模型目录扩展；Dell token 展望标厂商自称。