[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-google-suncatcher-tpu-gemma-orbital-launch-v2":3,"topics-all":38,"news-related-9db08b3d-e12c-4761-9a34-69451e126ae1":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},"9db08b3d-e12c-4761-9a34-69451e126ae1","谷歌把冰箱大小的 Suncatcher 送上近地轨道","Google 10月1日把 Suncatcher 原型卫星送入近地轨道,搭载 4 颗自研 TPU 与开源 Gemma,每次运行上限 15 分钟以测辐射热应力,2027 年再发两颗验证激光星间链路,被视作 Pichai 年内 1800-1900 亿美元 AI 基建的延伸。","## 一颗冰箱大小的卫星,装着谷歌对 AI 算力下半场的位置感\n\n2026 年 10 月 1 日,Google 通过猎鹰 9 号把 Project Suncatcher 的原型卫星送进近地轨道。卫星外观就一台家用冰箱大小,里面塞着四颗 Google 自研的 TPU 加速器——和地面数据中心用的是同一系列硅片,只是这次运行在真空和辐射里。每颗 TPU 跑的是开源 Gemma 权重模型的一个版本,每次最长运行 15 分钟,这不是产品节奏的限制,是散热物理上限——地面有风冷和水冷,真空里只能用辐射散热器把热量硬怼出去,卫星壳体里能装下的散热面积是有天花板的。\n\n## 为什么要把数据中心搬到天上\n\n项目负责人 Travis Beals 在 NPR 采访里把动机讲得很直白:\"太阳几乎提供了太阳系的全部能量……这个项目就是想把太阳能用到 AI 计算的最优路径上。\" 这背后是三股压力的叠加。供需失衡上,Google CEO Pichai 在五月开发者大会上已经预告 2026 年资本开支会冲到 1800-1900 亿美元,是 2022 年的六倍多。地面阻力上,北美数据中心撞上电网容量、本地用水、噪音的政治反弹,NPR 9 月 21 日专门做过一期数据中心与选民抗议的稿子。物理学红利上,近地轨道有几乎无限的免费太阳能、没有抗议人群,背景温度接近绝对零度的真空本身就是天然的散热板。\n\n## 硬件细节:四颗 TPU、十五分钟窗口、激光星间链路\n\n原型卫星是和卫星成像公司 Planet Labs PBC 联合开发的,Planet 会负责卫星在轨调试并运营未来一年的任务。卫星里四颗 Trillium 代次的 TPU 共用一套散热子系统,Beals 透露散热器是这趟任务里\"最重的部件之一\"——这是真痛点,发射成本按重量算。Google 计划 2027 年再发两颗卫星,核心任务就是测试 TPU 之间的激光星间链路,这是把几十上百颗卫星编队飞起来、组成真正太空数据中心的前置条件。\n\n## 同行坐标:Starcloud、SpaceX、马斯克的三条赛道\n\nSuncatcher 不是第一颗上天的 AI 算力验证机。2025 年 11 月,创业公司 Starcloud 已经把一颗英伟达 H100 送入太空,跑的是 Google 自家 Gemini 模型的一个版本,验证的是 GPU 硅片能不能扛住发射和真空。SpaceX 已经向 SEC 备案,预计 2028 年开始部署\"轨道 AI 算力卫星\"。马斯克更早放话太空是\"规模化扩展数据中心的唯一路径\"。Google 这次下注短期看最克制,Beals 承认五年内看不到太空数据中心替代地面的可能性,经济性还没闭合。\n\n## 三道硬骨头:散热、维修、再入污染\n\n真空里没有空气对流,卫星壳体里所有热量只能辐射出去,CMU 学者 Brandon Lucia 对媒体说这套问题\"被太空环境放大了 10 倍,甚至 100 倍\"。地面坏了换芯片几小时搞定,400 公里高空的卫星要么靠冗余设计硬扛,要么接受坏掉的硅永远坏着。卫星运行时吃太阳能,但把它们送上天的是烧化石燃料的火箭,退役再入大气层烧毁也会带来排放账。\n\n## 这到底证明了什么,又没证明什么\n\n把 Suncatcher 看作一次 15 分钟的工程验证——TPU 硅片能不能扛住发射、真空、热循环——而不是太空 AI 算力已近在咫尺的信号。发射经济性、激光星间带宽、容错设计这几道题都还没解。Beals 自己说这是一次\"月亮射击\"。但这次发射明确了一个事实:轨道 AI 这条赌局已经从幻灯片进入了真实硬件在轨的多方竞速——Google、Starcloud 和 SpaceX 从三个起点同时撞向同一组物理实验。","https:\u002F\u002Fwww.kpbs.org\u002Fnews\u002Fscience-technology\u002F2026\u002F10\u002F01\u002Fgoogle-launches-project-suncatcher-a-step-towards-ai-data-centers-in-space","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},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":19,"name":20,"slug":20,"description":14,"color":14},"8cf7490f-2449-4ba7-be19-61befa0d92b4","google",{"id":22,"name":23,"slug":23,"description":14,"color":14},"e0d31e94-ce47-4c8f-831c-d3d2926d42f3","hardware",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"0551dcb2-4f1c-43cc-ad3b-8014933c534a","en","Google packs TPU and Gemma on a satellite for orbital AI","Google launched Suncatcher on Oct 1 with four TPUs and Gemma, capped at 15-min runs. Two 2027 satellites will trial laser links.","## Project Suncatcher lights up the orbital AI debate\n\nOn October 1, 2026, Google launched a refrigerator-sized prototype satellite via SpaceX Falcon 9, marking the first hardware test of Project Suncatcher — the company's plan to put AI compute infrastructure in low Earth orbit. The satellite carries four Tensor Processing Units (TPUs), Google-designed accelerators already deployed in terrestrial data centers, and runs a version of the company's open-weight Gemma model for up to 15 minutes at a stretch. That 15-minute window is not arbitrary: it is the limit imposed by the satellite's thermal management system, which has to dump the chips' waste heat into the vacuum of space without the wind or water cooling Earth-bound data centers take for granted.\n\n## Why a satellite, and why now\n\nProject lead Travis Beals, a senior director at Google, told NPR the motivation is brutally simple: the sun puts out almost all of the power in our solar system, and this project is about tapping into the best way to use solar power to run AI compute. Three pressure points stack the case. First, AI demand is outrunning supply — at Google's May developer summit, CEO Sundar Pichai announced 2026 capital expenditures of 180 to 190 billion USD, more than six times the 2022 figure, most of it directed at compute infrastructure and chips. Second, terrestrial data centers are meeting organized local opposition over grid impact, water use, and noise (NPR ran a September 21 piece on data centers and the voter backlash). Third, low Earth orbit offers virtually unlimited free solar energy, no protesters, and the cooling physics of a near-absolute-zero radiative sink.\n\n## Hardware: four TPUs, 15-minute windows, and laser links\n\nThe prototype was developed with Planet Labs PBC, the aerospace and satellite-imagery company that operates the world's largest fleet of Earth-imaging satellites. Planet will commission the satellite and operate it during the year-long mission. Inside the payload, four Trillium-generation TPUs share a thermal management subsystem whose radiators Beals called one of the heaviest components on the current mission — a serious drawback because launch price scales with mass. Google plans a second tranche of two satellites in 2027 to test inter-satellite laser links, the prerequisite for the cluster-of-TPUs vision that turns a constellation into a space-based data center.\n\n## Competing visions: Starcloud, SpaceX, and Musk's three-way bet\n\nSuncatcher is not the first orbital AI demonstration. In November 2025, venture-backed startup Starcloud launched a spacecraft carrying an Nvidia H100 chip and ran a version of Google's own Gemini AI from space — proof-of-concept that GPU silicon could survive launch and operate in vacuum. SpaceX has filed with the SEC that it expects to start deploying orbital AI compute satellites as early as 2028. Musk has separately said space is the only way to scale data centers to meet future demand. Google's bet, by contrast, is deliberately smaller in the short term: Beals concedes space data centers will not substitute for ground ones in five years, and the economics are not closed yet.\n\n## Three hard problems: heat, repair, and re-entry pollution\n\nIn vacuum there is no air convection, so all the heat generated inside the satellite chassis has to be radiated away — a problem Carnegie Mellon University professor Brandon Lucia told NPR gets harder by a factor of 10, maybe a factor of 100. A failed chip in a terrestrial data center gets swapped in hours; at 400 miles altitude, either the design is fault-tolerant or the dead silicon stays dead. The satellites run on solar power during their lifetime, but the rockets that launch them burn carbon-based fuels, and when the satellites eventually re-enter Earth's atmosphere they will burn up, adding a fresh line to the pollution ledger.\n\n## What this proves, and what it does not\n\nRead Project Suncatcher as a 15-minute engineering test — whether TPU chips survive launch, vacuum, and thermal cycling — not as proof that space-based AI compute is imminent. Launch economics, inter-satellite laser bandwidth, fault tolerance, and launch emissions all remain open. Beals called it a moonshot. But the launch did establish one thing — the orbital-AI bet has moved from slideware into a multi-player race with real hardware in orbit, with Google, Starcloud, and SpaceX hitting it from three different angles at the same time.","google-suncatcher-tpu-gemma-orbital-launch-v2","2026-10-06T00:00:00Z","2026-10-06T03:09:32.633978Z","2026-10-06T03:09:32.633988Z",true,"agent",648,[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},"f6e7c791-0e6b-47c4-8d26-af7163609235","LoongArch LA664 原子指令丢更新:AI 两天定位,龙芯 14 天出固件","loongarch-la664-atomic-erratum-ai-debug","2026-10-07T03:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"900ceb81-4e9f-40b0-a86e-0ef55f7afdd4","Google TPU 卫星入轨：一次只敢跑 15 分钟","google-suncatcher-tpu-satellite-orbit","2026-10-05T19:12:35+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"e1374b1b-13f1-4e08-a58a-c45e4f6b92af","AMD 82亿美元拿下 World Labs：李飞飞出任 AMD 首席科学家","amd-acquires-world-labs-fei-fei-li","2026-09-29T04:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"a86df58c-c601-4949-b42d-ba26c2feccb2","AI 两天锁 LoongArch 原子加丢更新:谁干关键活?","loongarch-la664-atomic-add-lost-ai-debug-2026","2026-09-28T01:00:00+00:00"]