{
 "source": "https://thenewstack.io/ai-cloud-taxonomy-2026/",
 "source_date": "2026-03-14",
 "about": "骨架：The New Stack《A practical guide to the 6 categories of AI cloud infrastructure in 2026》（Janakiram MSV，2026-03-14）的六类：Traditional hyperscalers / Neoclouds / Developer-oriented clouds / Inference-optimized platforms / GPU marketplaces / Orchestration & portability layers。其中 Inference-optimized 拆成软件推理平台（b1）和芯片推理云（b2）。tns_category = 原文那一类和它点名的公司，ai_era = 老做法 → AI 时代变成什么；new_layer = 原文没有、我们补的层（去中心化算力、Serverless GPU、微调与后训练、模型聚合与路由、推理引擎商业公司、沙箱与 Agent 运行环境）。box = 我们加的四个大框（租算力 / 买推理 / 调度与引擎 / 跑 Agent）。who = 第二套拆法（按开发者阶段），layers 指向第一套的 code。",
 "layers": [
  {
   "code": "a1",
   "name": "超大云的 AI 算力",
   "ai_era": "通用云加几种 GPU 规格 → 自研芯片（TPU、Trainium）加万卡集群，和模型公司签多年大单",
   "tns_category": "Traditional hyperscalers · AWS、Azure、GCP、Oracle",
   "new_layer": false,
   "box": "租算力",
   "box_zh": "按卡、按小时：自己拿 GPU 跑训练和推理"
  },
  {
   "code": "a2",
   "name": "Neocloud · 大型 GPU 云",
   "ai_era": "托管机房 → 专为 GPU 建的数据中心，先签包量合同再建，卖裸金属集群",
   "tns_category": "Neoclouds · CoreWeave、Lambda、Crusoe、Nebius",
   "new_layer": false,
   "box": "租算力",
   "box_zh": "按卡、按小时：自己拿 GPU 跑训练和推理"
  },
  {
   "code": "a3",
   "name": "开发者云",
   "ai_era": "VPS 云 → 加单卡和小集群 GPU，价格透明、刷卡就能开",
   "tns_category": "Developer-oriented clouds · DigitalOcean、Vultr、Hyperstack、Latitude.sh",
   "new_layer": false,
   "box": "租算力",
   "box_zh": "按卡、按小时：自己拿 GPU 跑训练和推理"
  },
  {
   "code": "a4",
   "name": "GPU 市场与算力交易",
   "ai_era": "二手主机出租 → 按小时撮合各家闲置卡，算力像现货、期货一样买卖",
   "tns_category": "GPU marketplaces · Vast.ai、Runpod、TensorDock",
   "new_layer": false,
   "box": "租算力",
   "box_zh": "按卡、按小时：自己拿 GPU 跑训练和推理"
  },
  {
   "code": "a5",
   "name": "去中心化算力",
   "ai_era": "挖矿网络 → 用代币激励把全世界的闲置 GPU 聚成一朵云",
   "tns_category": null,
   "new_layer": true,
   "box": "租算力",
   "box_zh": "按卡、按小时：自己拿 GPU 跑训练和推理"
  },
  {
   "code": "b1",
   "name": "推理平台 · 开源模型 API",
   "ai_era": "自己部署模型 → 一个 API 调几百个开源模型，按 token 计费，拼速度和价格",
   "tns_category": "Inference-optimized platforms（软件侧）· Fireworks AI",
   "new_layer": false,
   "box": "买推理",
   "box_zh": "按 token、按秒：不碰显卡，直接调模型或跑自己的代码"
  },
  {
   "code": "b2",
   "name": "芯片推理云",
   "ai_era": "卖芯片 → 自研芯片配自己的 token API，用速度换客户",
   "tns_category": "Inference-optimized platforms（芯片侧）· Groq、SambaNova、Cerebras",
   "new_layer": false,
   "box": "买推理",
   "box_zh": "按 token、按秒：不碰显卡，直接调模型或跑自己的代码"
  },
  {
   "code": "b3",
   "name": "Serverless GPU",
   "ai_era": "常驻实例按小时付 → 写个函数就能上 GPU，按秒计费、冷启动秒级",
   "tns_category": null,
   "new_layer": true,
   "box": "买推理",
   "box_zh": "按 token、按秒：不碰显卡，直接调模型或跑自己的代码"
  },
  {
   "code": "b4",
   "name": "微调与后训练",
   "ai_era": "自己搭训练栈 → 上传数据、调 API 就拿到微调模型，强化学习也托管",
   "tns_category": null,
   "new_layer": true,
   "box": "买推理",
   "box_zh": "按 token、按秒：不碰显卡，直接调模型或跑自己的代码"
  },
  {
   "code": "b5",
   "name": "模型聚合与路由",
   "ai_era": "每家推理商签一个 key → 一个入口按价格、速度路由到多家推理商",
   "tns_category": null,
   "new_layer": true,
   "box": "买推理",
   "box_zh": "按 token、按秒：不碰显卡，直接调模型或跑自己的代码"
  },
  {
   "code": "c1",
   "name": "编排与多云调度",
   "ai_era": "绑定一朵云 → 一套配置跨云找最便宜的卡，训练、推理统一调度",
   "tns_category": "Orchestration & portability layers · BentoML、SkyPilot、Anyscale",
   "new_layer": false,
   "box": "调度与引擎",
   "box_zh": "让同一套模型在任何云上跑得更快、更省"
  },
  {
   "code": "c2",
   "name": "推理引擎商业公司",
   "ai_era": "自己调 CUDA → vLLM、SGLang 这些开源引擎背后的公司，卖企业版和托管",
   "tns_category": null,
   "new_layer": true,
   "box": "调度与引擎",
   "box_zh": "让同一套模型在任何云上跑得更快、更省"
  },
  {
   "code": "d1",
   "name": "沙箱与 Agent 运行环境",
   "ai_era": "给人用的云主机 → 给 agent 秒开、用完即弃的 microVM，按秒计费",
   "tns_category": null,
   "new_layer": true,
   "box": "跑 Agent",
   "box_zh": "给 agent 写的代码和动作一个隔离环境"
  }
 ],
 "who": [
  {
   "code": "W-big",
   "name": "万卡集群",
   "desc": "超大云和 neocloud",
   "layers": [
    "a1",
    "a2"
   ],
   "section": "训练模型"
  },
  {
   "code": "W-tune",
   "name": "微调与后训练",
   "desc": "不想自己搭训练栈",
   "layers": [
    "b4"
   ],
   "section": "训练模型"
  },
  {
   "code": "W-cheap",
   "name": "开卡就用",
   "desc": "开发者云、GPU 市场、去中心化",
   "layers": [
    "a3",
    "a4",
    "a5"
   ],
   "section": "开发和试验"
  },
  {
   "code": "W-orch",
   "name": "跨云调度",
   "desc": "一套配置到处跑",
   "layers": [
    "c1"
   ],
   "section": "开发和试验"
  },
  {
   "code": "W-api",
   "name": "直接调模型",
   "desc": "推理平台、芯片推理云、路由",
   "layers": [
    "b1",
    "b2",
    "b5"
   ],
   "section": "上线推理"
  },
  {
   "code": "W-own",
   "name": "跑自己的模型",
   "desc": "Serverless GPU、推理引擎",
   "layers": [
    "b3",
    "c2"
   ],
   "section": "上线推理"
  },
  {
   "code": "W-sbx",
   "name": "沙箱",
   "desc": "隔离、秒开、用完即弃",
   "layers": [
    "d1"
   ],
   "section": "跑 Agent"
  }
 ]
}