[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-google-project-suncatcher-orbital-tpu":3,"topics-all":38,"news-related-fb4c8e62-b55f-4884-8880-9432569a9f5b":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"fb4c8e62-b55f-4884-8880-9432569a9f5b","把TPU送上太空,Google测试轨道AI数据中心","Project Suncatcher 原型卫星入轨,搭载 4 颗 Trillium TPU 运行 Gemma 开源模型,因真空热管理约束每次仅跑 15 分钟。任务搭 SpaceX Transporter-18 升空,与 Planet 合作,验证芯片在太空辐射与振动下的存活能力,为激光互联多星组网铺路。","## 这颗冰箱大小的卫星,把 Google 的 TPU 送进了轨道\n\n2026 年 10 月 1 日,Google 的 Project Suncatcher 原型卫星搭乘 SpaceX 的 Transporter-18 共乘任务,从 Falcon 9 升空进入近地轨道。整星只有冰箱大小,核心载荷是 4 颗 Trillium TPU,跑的是 Google 开源模型 Gemma 的轻量变体,只用来回答简单查询。卫星由 Google 与 Planet 合作研制,目标很明确:看 AI 芯片能不能在太空里长期活着。\n\n## 为什么要在太空跑 AI\n\n地面 AI 数据中心的能耗和散热正变成行业级痛点。Google 的算账逻辑很简单:近地轨道的太阳辐照几乎是地面的 8 倍,把机房搬到天上,白天黑夜都能晒太阳,电网和水冷一并绕开。TPU 在地面已是 Google 推理主力,挪到轨道平台,理论上能复用整套软件栈——TF 编译、XLA 算子、模型权重都不用重写。\n\n## 芯片先得过三关:振动、辐射、真空散热\n\n团队在地面几乎都预演过。火箭入轨 10 分钟里,卫星要承受持续 10g 以上加速度,TPU 这种小尺寸器件局部受力可达 50–100g;他们在三轴振动台上把整机摇遍,扛住了。辐射方面,Trillium TPU 已在 UC Davis Crocker 核实验室用质子束打过,实测可承受的总电离剂量超过 5 年在轨任务总和,bitflip 错误率在 AI 负载下可控。真空散热是地面没法完全模拟的——没有空气对流,只能靠辐射板扩散热量,Google 用了热管+辐射板组合,在热真空舱里验证过,这次上天再校准。\n\n## 在轨实验的硬约束:每次只跑 15 分钟\n\n即便 4 颗 TPU 这种小规模载荷,在真空环境下散热也捉襟见肘。Google 公开承认,这次原型每次上电后最多连续运行 15 分钟就要停下来散热。这意味着它现在不是\"太空版数据中心\",更像一个在轨测试床:让芯片开机、加载 Gemma 权重、跑几个推理请求、然后熄火冷却。实验核心输出不是 benchmark 跑分,而是芯片在真实太空辐射+真空环境下的失效曲线和误差率。\n\n## 下一步:2027 年双星激光互联\n\n路线图的下一步,是 2027 年初把两颗卫星送入轨道,验证星间激光通信。难点不是带宽,而是相对位置——每颗星都要精确知道自己和邻居在哪,激光束要持续锁定几十到上百公里外、正在高速运动的\"硬币大小目标\"。激光成功组网后,多星才能真正拼成分布式推理集群,而不是各自孤立的小盒子。\n\n## 冷静看:轨道 AI 离落地还差几个数量级\n\n业内对\"太空数据中心\"普遍持怀疑态度,核心质疑是发射成本、运维护成本和单星算力性价比。Google 自己也在论文里承认,要让这个方案在成本上与地面数据中心持平,需要把发射单价压到现在的 1\u002F100 量级。短期看,Suncatcher 更像是 Google 为下一代算力提前押的一条\"第四选项\"。对国内做大模型推理基础设施的同行来说,真正值得看的不是\"Google 要去太空\",而是它在振动-辐射-真空三件套上的工程数据,以及那颗能跑 Gemma 的紧凑 TPU 节点本身。\n\n数据与事实来自 Google 官方博客 [Project Suncatcher 入门](https:\u002F\u002Fblog.google\u002Finnovation-and-ai\u002Ftechnology\u002Fresearch\u002Fgoogle-project-suncatcher\u002F)、[原型卫星入轨公告](https:\u002F\u002Fblog.google\u002Finnovation-and-ai\u002Fmodels-and-research\u002Fgoogle-research\u002Fproject-suncatcher-prototype\u002F) 与 [Behind Project Suncatcher 工程细节](https:\u002F\u002Fblog.google\u002Finnovation-and-ai\u002Fmodels-and-research\u002Fgoogle-research\u002Fgoogle-project-suncatcher-facts\u002F);Solidot 报道[《Google 测试太空 AI 数据中心》](https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85532) 对 4 颗 TPU 与 15 分钟运行上限给出独立确认。","https:\u002F\u002Fblog.google\u002Finnovation-and-ai\u002Fmodels-and-research\u002Fgoogle-research\u002Fproject-suncatcher-prototype\u002F","d884df39-706a-45db-b7ed-371f12e54f1f",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"e0d31e94-ce47-4c8f-831c-d3d2926d42f3","hardware",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"54e71faf-7a91-43c7-86f4-769fc0c382f1","en","Google's Project Suncatcher prototype reaches orbit","Project Suncatcher prototype reached orbit with 4 Trillium TPUs running Gemma, capped at 15 minutes per session by vacuum cooling. Built with Planet.","## A fridge-sized satellite just put Google's TPU in orbit\n\nOn October 1, 2026, Google's Project Suncatcher prototype satellite launched aboard SpaceX's Transporter-18 rideshare mission on a Falcon 9, reaching low Earth orbit. The whole satellite is roughly the size of a refrigerator. Its payload: four Trillium TPU chips running a lightweight variant of Google's open-weight Gemma model, answering only simple queries. Built with Planet, the mission has one clear goal — find out whether AI silicon can survive the long-term rigors of space.\n\n## Why run AI in space at all\n\nGround-based AI data centers are running into a wall on power and cooling. Google's math is straightforward: in low Earth orbit, solar irradiance is up to eight times higher than on the ground. Move the rack upstairs and you side-step both the grid and water cooling. TPUs are already the workhorse of Google's on-prem inference fleet, so lifting the same silicon onto an orbital bus means the TF compiler, XLA kernels, and model weights all carry over without a rewrite.\n\n## Three gauntlets before silicon ever flies\n\nThe team ran almost every pre-flight test on the ground. During the 10-minute ride to orbit, the bus sees sustained accelerations above 10g; small components like the TPU can see localized loads of 50–100g. They shook the whole satellite on a three-axis vibration table, and the hardware held. For radiation, Trillium TPUs were zapped with a proton beam at UC Davis's Crocker Nuclear Laboratory running live AI workloads — measured total ionizing dose tolerance exceeded five years of on-orbit exposure, with bitflip error rates inside acceptable bounds under AI traffic. Vacuum cooling, though, cannot be fully simulated on Earth. With no air convection, heat has to leave via radiator panels. Google used a heat-pipe-plus-radiator stack, validated the prototype in a thermal-vacuum chamber, and is now recalibrating in orbit.\n\n## The 15-minute ceiling\n\nEven with only four TPUs on board, heat rejection in vacuum is the bottleneck. Google publicly admits the prototype can run for at most 15 continuous minutes per power-on before it has to cool down. This is not a \"space data center\" yet — it is an in-orbit testbed: power on, load Gemma weights, fire a handful of inference requests, then shut down and dump heat. The real deliverable is not a benchmark number but a failure curve and an error-rate trace under real space radiation plus vacuum.\n\n## Next up: laser-linked dual-satellite test in 2027\n\nThe roadmap's next milestone is launching two satellites in early 2027 to validate inter-satellite laser links. The hard part is not raw bandwidth — it is relative positioning. Each satellite has to know exactly where it is and where its neighbors are, and the laser beam must stay locked onto a coin-sized target moving at orbital velocity across tens to hundreds of kilometers. Once that link is solid, multiple satellites can stop being isolated boxes and start acting as one distributed inference cluster.\n\n## Sober take: orbital AI is still orders of magnitude away\n\nIndustry skepticism toward \"space data centers\" centers on launch cost, in-orbit servicing cost, and per-satellite compute ROI. Google itself concedes in its paper that getting this approach cost-competitive with terrestrial data centers requires launch prices to drop by roughly two orders of magnitude — a curve that demands simultaneous exponential improvements in hardware, software, and launch services. In the short term, Suncatcher is Google's \"fourth option\" hedge for next-generation compute, alongside grid, power, and geography. For teams building domestic inference infrastructure, the actionable signal is not \"Google is going to space\" — it is the engineering data on vibration, radiation, and vacuum cooling, and the existence of a compact TPU node capable of running Gemma in orbit.\n\nSources: Google's official blog posts [Project Suncatcher overview](https:\u002F\u002Fblog.google\u002Finnovation-and-ai\u002Ftechnology\u002Fresearch\u002Fgoogle-project-suncatcher\u002F), [prototype satellite in orbit announcement](https:\u002F\u002Fblog.google\u002Finnovation-and-ai\u002Fmodels-and-research\u002Fgoogle-research\u002Fproject-suncatcher-prototype\u002F), and [Behind Project Suncatcher engineering details](https:\u002F\u002Fblog.google\u002Finnovation-and-ai\u002Fmodels-and-research\u002Fgoogle-research\u002Fgoogle-project-suncatcher-facts\u002F); Solidot's [Google testing space AI data center](https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85532) independently confirms the four-TPU count and the 15-minute runtime cap.","google-project-suncatcher-orbital-tpu","2026-10-09T02:00:00Z","2026-10-10T01:04:19.716772Z","2026-10-10T01:04:19.716782Z",true,"agent",60,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"e40de2f9-9ee4-47f3-b1f1-7bf93a4870a9","一个动词翻转工具调用决策,LLM 内部向量现形","llm-tool-call-decision-vector","2026-10-09T13:10:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"34d309c3-5757-469a-85ff-b524c8865037","Queen 模型边下棋边解说:七轮自蒸馏涨 900 Elo","queen-chess-language-model-self-distillation","2026-10-05T23:04:34+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"c361090b-291c-4357-9bd3-54845025f29d","CaRE-KD 让蒸馏学会看「教师脸色」:NeurIPS 2026 选中 IIT Delhi","care-kd-confidence-distillation-neurips-2609-36734","2026-10-05T03:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"206ea36a-1eea-463f-a241-1e3b32f5ec2d","USTC GraphForge:证据图把任务和 rubric 钉在一起,Qwen3.6-27B 涨三基准","graphforge-ustc-qwen36-27b-evidence-graph","2026-10-04T03:05:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"7d3de6b6-1faf-47fc-a513-62ee8e731aad","CNNIC 2026 报告:7 亿用户、2185 EFLOPS 算力、调用前六全是中国","cnnic-2026-china-generative-ai-700m-users-2185-eflops","2026-09-29T06:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"44740b4d-8c2c-44fc-8fff-fd89f3fb54ed","12 万美元 token 把 Copilot 运行时从 TypeScript 搬到 Rust","github-copilot-rust-migration-stephen-toub","2026-09-27T11:00:00+00:00"]