{
  "schemaVersion": "0.11.0",
  "canonical": "https://www.pystone.net/notes/cloud-provider-gpu-nvidia-analysis/",
  "atlas": "https://www.pystone.net/?node=cloud-provider-gpu-nvidia-analysis#knowledge-atlas",
  "markdown": "https://www.pystone.net/notes/cloud-provider-gpu-nvidia-analysis.md",
  "context": "https://www.pystone.net/notes/cloud-provider-gpu-nvidia-analysis.context.json",
  "knowledgeVersion": "224c990773de.5fa8af6e39fa",
  "build": {
    "siteCommit": "224c990773de166d23a886306577dd90379529ce",
    "notesCommit": "5fa8af6e39fa3891d1b9b4832bfa6c4e0ecaaf0a",
    "builtAt": "1970-01-01T00:00:00.000Z",
    "version": "224c990773de.5fa8af6e39fa"
  },
  "id": "note:cloud-provider-gpu-nvidia-analysis",
  "slug": "cloud-provider-gpu-nvidia-analysis",
  "title": "关于云服务供应商的GPU是否是NVIDIA",
  "type": "note",
  "visibility": "public",
  "idStability": "rename-stable",
  "author": {
    "name": "Perrin Yong",
    "profile": "https://www.pystone.net/profile/"
  },
  "publisher": {
    "name": "Perrin Yong",
    "profile": "https://www.pystone.net/profile/"
  },
  "aliases": [],
  "summary": "关于云服务供应商的GPU是否是NVIDIA",
  "contentRole": "unspecified",
  "isMoc": false,
  "mocRecognition": "none",
  "generated": false,
  "attribution": "unspecified",
  "domain": "40-社会科学、制度与经济",
  "tags": [],
  "mocs": [],
  "contentHash": "f89a1a833652889b782b92c314ef1161b16c455878ea661831d6b296d7df9c4b",
  "assets": [],
  "headings": [
    {
      "depth": 1,
      "text": "关于云服务供应商的GPU是否是NVIDIA",
      "anchor": "关于云服务供应商的gpu是否是nvidia",
      "citation": "https://www.pystone.net/notes/cloud-provider-gpu-nvidia-analysis/#%E5%85%B3%E4%BA%8E%E4%BA%91%E6%9C%8D%E5%8A%A1%E4%BE%9B%E5%BA%94%E5%95%86%E7%9A%84gpu%E6%98%AF%E5%90%A6%E6%98%AFnvidia"
    },
    {
      "depth": 3,
      "text": "小结",
      "anchor": "小结",
      "citation": "https://www.pystone.net/notes/cloud-provider-gpu-nvidia-analysis/#%E5%B0%8F%E7%BB%93"
    }
  ],
  "claims": [],
  "outgoing": [],
  "incoming": [
    {
      "id": "note:society-institutions-economics",
      "title": "社会科学、制度与经济",
      "url": "https://www.pystone.net/notes/society-institutions-economics/",
      "atlas": "https://www.pystone.net/?node=society-institutions-economics#knowledge-atlas",
      "label": "社会科学、制度与经济",
      "origin": "explicit",
      "humanReviewed": true,
      "context": "互联网社会观察中的“关于云服务供应商的GPU是否是NVIDIA”导航项",
      "citation": "https://www.pystone.net/notes/society-institutions-economics/#%E4%BA%92%E8%81%94%E7%BD%91%E7%A4%BE%E4%BC%9A%E8%A7%82%E5%AF%9F"
    }
  ],
  "contentMarkdown": "# 关于云服务供应商的GPU是否是NVIDIA\n\n# 关于云服务供应商的GPU是否是NVIDIA\n\n**Explaining NVIDIA GPU export controls**\n\nUser wants to understand why Alibaba Cloud can buy NVIDIA GPUs despite restrictions on China. The reason might involve special models like A800, H800, or domestic versions, possible license approvals, or purchasing from Hong Kong before the export ban. Small businesses may not easily import these, but they can rent cloud instances or purchase approved GPUs from domestic vendors, though at a higher cost. I’ll need to check the latest info and regulations around these export controls.\n\n\n**Explaining NVIDIA GPU availability**\n\nUser wants to know why Alibaba Cloud can still buy NVIDIA GPUs despite the block on China. The reason is likely due to special models like A800, H800 with export restrictions and possible licenses for domestic versions. Alibaba’s procurement scale might also help secure these. Small businesses can’t easily import, but they can rent cloud instances or buy from domestic vendors, which can be pricey. I’ll look into up-to-date info on the latest export controls to clarify.\n\n\nThought for 22 seconds\n\n\n**为什么阿里云还能拿到 NVIDIA GPU？**\n\n1. **“降配版”在政策窗口期合法进口**\n\n   * 2022 年 10 月起，美国对单卡、单节点和整机柜的计算与互联性能都设定了门槛（所谓 “性能密度” 阈值）。NVIDIA 随后把 A100/H100 做了削减带宽与互联速率的 **A800、H800、H20** 专门型号，官方获得了*一次性*出口许可证。阿里云、腾讯等头部厂商在 2023 年–2024 年间大量下单并清关，这批库存目前仍在其机房内运行。([Tom's Hardware][1])\n\n2. **库存、期货和整机渠道**\n\n   * 禁令针对 *芯片直接出口*；整机系统（GPU 板卡已焊接到服务器）在当时仍可获批。阿里云可通过 OEM 服务器厂商（Supermicro、浪潮、新华三等）采购整机，芯片早已焊好并一次性报关入境。\n   * 头部云商有规模优势、资金与海关资源，能提前锁定 6 – 12 个月的产能并囤货；中小企业难以做到。\n\n3. **云租赁不等于再出口**\n\n   * 美方出口管制限制*再销售硬件*，不限制*在中国境内出租算力*。阿里云提供的 GPU 实例物理上留在中国数据中心，仅向国内用户开放 API，因此不触碰“转移”条款。\n\n4. **2024 年 10 月后的收口**\n\n   * 2024/10 起，最新修订把 A800/H800/H20 也纳入许可证清单，并收紧“整机豁免” 规则。新单已经无法再进口，阿里云只能吃旧库存或转向国产 GPU（华为昇腾、海光鲲鹏 + 国产 AI 卡）或等待下一批“降配”方案获批。([The Verge][2], [Wall Street Journal][3])\n\n---\n\n**中小企业还能不能买到 NVIDIA GPU？**\n\n* **直接进口：几乎不可能**\n\n  * 任何面向中国的 A100/H100/H200/A800/H800/H20 级 GPU 都要商务部 (BIS) 许可证，小公司难以通过审查。\n  * 低端或旧款（A10、A30、T4、RTX 4090）不在禁令范围，仍可通过代理采购，但价格被市场哄抬。\n\n* **曲线渠道风险高**\n\n  * 通过港澳或第三国“二跳”再运大陆已经被美国列为重点执法对象；一旦查实会触及《反洗钱》《反逃避出口管制》处罚，企业与个人都可能被拉入 SDN 名单。\n\n* **现实可行的两条路**\n\n  1. **租云**：直接使用阿里、腾讯、华为云已落地的 A800/H800 及未来国内 H20 实例，省掉报关和设备维护，也不会触碰二次出口风险。\n  2. **国产/开源 GPU**：选华为 Ascend 910B、摩尔线程 MT‑T60 或联想 x86+寒武纪 MLU300 等方案，软件生态不如 CUDA 成熟，但足以跑常见推理任务。\n\n---\n\n### 小结\n\n* **阿里云能用 NVIDIA**：靠“降配版 + 窗口期一次性许可证 + 整机进口”提前囤货，并且云出租不违反出口再转移禁令。\n* **中小企业**：新一代 NVIDIA 高端芯片基本买不到；要么上云租赁，要么用仍允许进口的中端卡或转投国产 GPU 生态。\n\n[1]: https://www.tomshardware.com/news/us-prohibits-exports-of-nvidias-a800-and-h800-to-china-blacklists-chinese-gpu-developers?utm_source=chatgpt.com \"US Prohibits Exports of Nvidia's A800 and H800 to China, Blacklists ...\"\n[2]: https://www.theverge.com/2023/10/17/23921131/us-china-restrictions-ai-chip-sales-nvidia?utm_source=chatgpt.com \"Nvidia’s H800 AI chip for China is blocked by new export rules\"\n[3]: https://www.wsj.com/livecoverage/stock-market-trump-tariffs-trade-war-04-16-25/card/nvidia-records-5-5-billion-charge-on-new-h20-export-restrictions-rCwWLidXVAY85QV1DER3?utm_source=chatgpt.com \"Nvidia Warns of $5.5 Billion Charge on New China Export Curbs; Stock Drops\"\n"
}