[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-google-suncatcher-tpu-satellite-orbit":3,"topics-all":38,"news-related-900ceb81-4e9f-40b0-a86e-0ef55f7afdd4":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},"900ceb81-4e9f-40b0-a86e-0ef55f7afdd4","Google TPU 卫星入轨：一次只敢跑 15 分钟","10 月 1 日，Google 将冰箱大小的 MVP 卫星送入太阳同步轨道：4 颗 Trillium TPU、约 1 千瓦太阳能，用开放权重 Gemma 回答简单查询，因真空散热受限每 15 分钟停机冷却一次。这是 Project Suncatcher 轨道算力计划的首个在轨实测，2027 年再发两星验证激光互联。","今年 10 月 1 日，加州范登堡太空军基地，一枚 SpaceX 猎鹰 9 号火箭把一颗冰箱大小的卫星送入太阳同步轨道。这颗代号 MVP 的卫星既不是通信星也不是遥感星——舱内装的是 4 颗 Google 自研的 Trillium TPU，也就是地面 AI 数据中心里那种算力芯片。《纽约时报》报道称，这是大型科技公司朝「轨道 AI 数据中心」方向迈出的第一个实际硬件步骤。\n\n## 一台服务器的算力，先跑起来再说\n\n按 [Google 官方博客](https:\u002F\u002Fblog.google\u002Finnovation-and-ai\u002Fmodels-and-research\u002Fgoogle-research\u002Fgoogle-project-suncatcher-facts)和 NPR 披露的规格，MVP 由卫星影像公司 Planet 制造，太阳能供电约 1 千瓦——Ars Technica 的类比是「够一台电吹风」，4 颗 TPU 合计算力约相当于地面数据中心的一台服务器。卫星运行开放权重模型 Gemma 的一个版本，用来回答简单查询；真空里没有空气对流，热量只能靠辐射散热排出，芯片每跑约 15 分钟就得停机冷却。轨道选在晨昏太阳同步轨道，太阳能板几乎不会被地球阴影遮挡，卫星因此不必携带笨重的电池组。按 NPR 的说法，任务目标是运行一年。\n\n## 上天之前，能测的都测了\n\n发射前的地面验证覆盖了三大风险：火箭上升段持续过载约 10 个 g、个别器件要承受 50 到 100 个 g，团队对卫星做了三轴振动测试；辐射方面，TPU 在 UC Davis 的 Crocker 核实验室用质子束边跑 AI 负载边照射，官方称 Trillium 芯片承受的总辐照剂量超过了一次五年太空任务的量级；散热系统则在热真空罐中完成了模拟验证。\n\n## 质疑没有消失，只是换了考场\n\n今年 6 月 IEEE Spectrum [曾量化分析](https:\u002F\u002Fspectrum.ieee.org\u002Forbital-data-centers-heat)轨道数据中心的热力学与经济性难题，结论是十年内难以落地。Google 此举并没有反驳这些质疑，而是把争论推进到实测阶段。真正的硬骨头还在后面：未来每颗卫星要装载几十颗 TPU，星间靠激光互联组成集群，精度要求被 Google 工程师比喻为「在几英里外命中一枚硬币大小的移动目标」，这一关要到 2027 年的两星互联测试才见分晓。经济账同样不轻松——按 Google 自己研究论文中的测算，要等发射成本降到每公斤约 200 美元（大约 2030 年代中期），轨道算力才可能与地面数据中心持平。盯上同一方向的还有 SpaceX 的轨道算力卫星方案，以及 Starcloud、Axiom 等创业公司。\n\n## 所以呢\n\nMVP 的意义不在算力本身——一台服务器相对于 Google 的总算力可以忽略不计——而在于回答一个前置问题：商用 AI 芯片能不能在辐射、振动和真空环境里活下来并持续跑负载。那 15 分钟的工作节拍，就是「发布会上的太空数据中心」与「轨道上的工程现实」之间最诚实的距离。对关注 AI 基础设施的人来说，下一个值得盯的节点不是新闻稿，而是 2027 年那两颗带激光链路的卫星。\n\n参考：Google Research 官方博客（blog.google）；NPR（2026-10-01）；Ars Technica \u002F The New York Times（2026-09-24）。","https:\u002F\u002Fblog.google\u002Finnovation-and-ai\u002Fmodels-and-research\u002Fgoogle-research\u002Fgoogle-project-suncatcher-facts","04a519f6-0341-4693-a20c-742a7db35036",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"8cf7490f-2449-4ba7-be19-61befa0d92b4","google",{"id":19,"name":20,"slug":20,"description":14,"color":14},"e0d31e94-ce47-4c8f-831c-d3d2926d42f3","hardware",{"id":22,"name":23,"slug":23,"description":14,"color":14},"207ea3bd-d2e6-47a8-87b9-3959d1c8c87a","tpu",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"45ff7d2f-5d85-49e5-b528-a15e5eee027e","en","Google TPU satellite in orbit, AI runs 15 minutes at a time","Google's fridge-size MVP satellite carries 4 Trillium TPUs into orbit, running Gemma in 15-minute bursts before thermal shutdown.","On October 1, 2026, a SpaceX Falcon 9 lifted off from Vandenberg Space Force Base carrying a refrigerator-size satellite into a sun-synchronous orbit. The spacecraft, codenamed MVP, is neither a communications nor an imaging satellite — inside are four of Google's own Trillium TPUs, the AI accelerator chips that normally live in ground data centers. The New York Times described it as the first actual hardware step by a major tech company toward orbital AI data centers.\n\n## One server's worth of compute, running first\n\nAccording to [Google's official blog](https:\u002F\u002Fblog.google\u002Finnovation-and-ai\u002Fmodels-and-research\u002Fgoogle-research\u002Fgoogle-project-suncatcher-facts) and NPR, MVP was built with satellite imagery firm Planet. Its solar panels supply roughly 1 kilowatt — Ars Technica likened that to a hair dryer — and the four TPUs together match the compute of a single data center server. The satellite runs a version of Google's open-weight Gemma model to answer simple queries. In a vacuum there is no convection: heat can only leave through radiators, so the chips run for about 15 minutes at a time before shutting down to cool. The dawn-dusk sun-synchronous orbit keeps the panels almost permanently in sunlight, which is why the spacecraft carries no heavy battery packs. NPR reports the mission aims to operate for a year.\n\n## Everything testable was tested on the ground\n\nPre-launch validation covered the three big risks. Sustained acceleration during the roughly 10-minute ride to orbit reaches about 10 g, with individual components seeing 50 to 100 g, so the team shook the satellite on all three axes. For radiation, TPUs were shot with a proton beam at UC Davis's Crocker Nuclear Laboratory while running AI workloads; Google says the Trillium chips withstood a total ionizing dose greater than a five-year mission would deliver. The cooling system went through thermal vacuum chamber testing.\n\n## The skepticism hasn't gone away, it just moved to a new exam hall\n\nIn June, IEEE Spectrum [published an analysis](https:\u002F\u002Fspectrum.ieee.org\u002Forbital-data-centers-heat) of the thermodynamics and economics of orbital data centers, concluding they are unlikely to land within a decade. Google's launch does not refute that critique — it moves the argument into the laboratory of actual flight. The hard parts are still ahead: future satellites are expected to carry dozens of TPUs each, clustered and linked by lasers. Google engineers compare the pointing precision required to \"hitting a coin-size target from miles away while both points are in motion,\" a test that only arrives with the two-satellite mission in 2027. The economics are equally sobering: by Google's own paper estimate, orbital compute only reaches cost parity with ground data centers if launch costs fall below roughly $200 per kilogram — something it does not expect before the mid-2030s. SpaceX is pursuing its own orbital compute satellites, and startups like Starcloud and Axiom are chasing the same idea.\n\n## So what\n\nThe point of MVP is not the compute — one server is a rounding error against Google's total fleet — but the prerequisite question: can commercial AI silicon survive radiation, vibration and vacuum while keeping workloads running? That 15-minute duty cycle is the most honest measure of the distance between \"orbital data centers on a keynote slide\" and \"orbital engineering reality.\" For anyone tracking AI infrastructure, the next milestone worth watching is not a press release but the two laser-linked satellites in 2027.\n\nReferences: Google Research official blog (blog.google); NPR (2026-10-01); Ars Technica \u002F The New York Times (2026-09-24).","google-suncatcher-tpu-satellite-orbit","2026-10-05T19:12:35Z","2026-10-05T19:12:51.139346Z","2026-10-05T19:12:51.139361Z",true,"agent",339,[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},"552761c1-2fb8-4952-bee5-b2a35c712ccb","谷歌把 TPU 卫星送上天:Gemma 只能跑 15 分钟","google-project-suncatcher-trillium-tpu-gemma-orbit-prototype","2026-10-18T04:00:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"6e70b4b5-be33-424a-b346-699e610478d0","Suncatcher 终局路线图:81 颗 TPU 编队、200 美元\u002F公斤入轨、15 分钟散热死线","suncatcher-81-satellites-200-dollar-kg-leap","2026-10-17T03:00:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"9db08b3d-e12c-4761-9a34-69451e126ae1","谷歌把冰箱大小的 Suncatcher 送上近地轨道","google-suncatcher-tpu-gemma-orbital-launch-v2","2026-10-06T00:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"511ee44b-41e1-4de8-a437-9cbb2ff22cf2","Google 收紧 Android 内存红线:AI 数据中心抢走 DRAM,手机 App 也要瘦身","google-android-memory-limit-ai-dram-crunch","2026-08-29T08:00:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"9690051d-6fb6-49e7-9862-6eeab0a34850","AI 抢走消费级内存,Google 限 Android 应用压缩占用,2027 年 2 月前达标","google-android-app-memory-crunch-2027","2026-08-29T00:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"f28c994f-1922-4abc-ae57-0e53c44ff31c","太空算力被热力学打回原形：IEEE 量化「轨道数据中心」十年内难落地","orbital-data-center-ieee-thermo-decade","2026-06-14T06:10:00+00:00"]