Google 要把 TPU 送上太空跑 AI
推荐指数 41.0 NO. 015 · 2026.09.25
发布2026/09/24Score55Comments100
为什么值得看
Google 的 Project Suncatcher 即将发射首颗原型卫星，测试 TPU 在低轨太空环境下的实际运行表现。近地轨道太阳能效率可达地面 8 倍，若验证成功，太空或成为未来大规模 AI 算力的新部署场景。
编辑判断
太空算力的核心瓶颈不是能源而是散热和辐射。TPU 在真空环境中无法依赖空气对流散热，辐射硬化改造成本可能抵消能源收益，且低轨卫星每 90 分钟经历一次剧烈温差循环。Google 此前在海底数据中心 Project Natick 的经验部分可迁移，但太空环境严苛数个数量级。
更现实的短期路径可能是"边缘推理+地面训练"的混合架构，而非完整训练集群上星。关注 SpaceX 星链与 AWS Outposts 的合作动向，卫星间激光通信带宽才是决定星座能否协同处理大模型的真正瓶颈。
社区反馈
负面 83 条评论
核心争论：太空数据中心是工程突破还是营销噱头，散热与经济性成最大质疑点
And just like that Elon Musk shifted the overton window away from "unthinkable" on yet another topic.
It's physically impossible It's technically impossible It won't be profitable <= You are here It's bad for the environment It needs to be nationalized This is just a joke, obviously, but the debate over on ArsTechnica about Falcon 9 going away is all about how it needs to be nation
I think one of the reasons for putting these in space is to move them beyond the borders of any national control.