“That's a hell of an R&D spend for the next 10 years of a company whose lifetime revenue (not profit) is reportedly $4.6e10.”
Towards a future space-based, highly scalable AI infrastructure system design
设想用搭载 TPU 的太阳能卫星集群在太空做机器学习算力
论文探讨在太空建设机器学习算力:卫星配备太阳能板、自由空间光学星间链路和 Google TPU,并近距离编队飞行,以 81 颗卫星、半径 1 km 的集群说明编队方案。Trillium TPU 经辐射测试可承受相当于 5 年任务寿命的总电离剂量而无永久故障;学习曲线分析认为,到 2030 年代中期发射到低轨的成本可能降到约 200 美元/公斤以下。
HN 上讨论“太空数据中心”时,评论者把它当作这一想法的来源,并引用其 200 美元/公斤的成本门槛和约 37 万吨发射量的假设;也有人认为这些假设过于乐观。
HN 书评
推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及“I think this is overoptimistic, as the Alphabet paper assumed 370,000 tons[1] getting launched as the process for reducing costs”
“The Google paper didn’t seem too concerned with radiator mass/size when I skimmed it, but maybe I just missed it.”
“that estimates that launch costs, just to roughly match data center energy costs on earth, would need to reach 200USD/kg, which requires a 10 fold cost reduction relative to the current launch costs of Falcon 9.”
“According to Google, the price threshold to make space make more sensible than building on the ground is $200/kg:”
“Also, from other statements they're clearly copying Alphabet's study which said cost parity in 2035, if they can actually launch 370,000 tons and maintain their learning rate.”
看全部 9 条 →
“The idea was from Google in a November 2025 paper.”
“However, they recon it will take SpaceX launching 370,000 tons to LEO to make the costs come down enough to be worth it: Towards a future space-based, highly scalable AI infrastructure system design”
“recon it will take SpaceX launching 370,000 tons to LEO to make the costs come down enough to be worth it: Towards a future space-based, highly scalable AI infrastructure system design”
- Towards a future space-based, highly scalable AI infrastructure system design ↗andsoitis · 2026-09-24 · 1 分