Original post · AI translation
彭博社：微软在决定不续签一份围绕必应签订的 2023 年合同后，正关闭其从甲骨文租用的 AI 服务器。
据知情人士称，由于甲骨文的硬件运行的是较早一代的英伟达芯片，已不再符合微软的规格要求，这些算力的价值随时间推移而下降。
简单来说，微软按 GPU 小时付费。Blackwell GPU 在功耗相近的情况下，推理输出量是 H100 的数倍，因此现在每个 Hopper 或 Ampere GPU 小时能完成的工作少得多。
目前的前沿模型通常采用大型混合专家架构，需要大量内存和 GPU 之间的高速连接。H100 拥有 80GB 内存，在其服务器内只能与 8 个 GPU 高速通信。Blackwell 机架将 72 个 GPU 连接为一个整体，每个 GPU 拥有 192 至 288GB 内存。大型模型要么无法装入旧服务器，要么在其较慢的网络上运行缓慢。
甲骨文表示，其 AI 基础设施的需求依然强劲；今年 9 月，该公司以比原价高 20% 的价格续签了使用已有 4 年的 GPU 的合同。
Original
Bloomberg: Microsoft is switching off AI servers it rented from Oracle after declining to renew a 2023 contract built around Bing. The capacity lost value over time because Oracle’s hardware ran older-generation Nvidia chips that no longer matched Microsoft’s specifications, according to people familiar with the matter. Basically, Microsoft pays per GPU-hour. A Blackwell GPU produces several times more inference output than an H100 for similar power, so each Hopper or Ampere hour now buys far less work. Current frontier models, often large mixture-of-experts designs, need lots of memory and fast links between GPUs. An H100 has 80GB of memory and talks quickly to only 8 GPUs in its server. Blackwell racks link 72 GPUs as one unit, with 192 to 288GB each. Big models either don’t fit on the old boxes or run slowly across their slower network. Oracle says demand for its AI infrastructure remains strong, and in September it renewed contracts for 4-year-old GPUs at a 20% premium to the original price.