[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-suncatcher-81-satellites-200-dollar-kg-leap":3,"topics-all":38,"news-related-6e70b4b5-be33-424a-b346-699e610478d0":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},"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","Google 与 Planet Labs 合作的 Project Suncatcher MVP 卫星 10 月 1 日入轨,4 颗 Trillium TPU 在 UC Davis 质子束测试中承受住了 15 krad(Si) 累计剂量,但真空散热仍是核心瓶颈——单次推理只能跑约 15 分钟。论文规划了 81 颗卫星、1 公里半径、10 Tbps 星间光链路的终局架构,200 美元\u002F公斤入轨成本是经济性临界点。","在近地轨道跑机器学习推理,Google 已经把第一颗实验星送上天,但「单星 15 分钟脉冲式运行」的硬限制让它暂时只是资格验证,而不是真正的在轨算力。Project Suncatcher MVP 卫星 10 月 1 日搭乘 SpaceX Falcon 9 的 Transporter-18 拼车任务入轨,搭载 4 颗 Trillium(TPU v6e)、约 1 千瓦太阳能、跑 Gemma 轻量变体回答短问——这是「在轨 AI 数据中心」梦想落地的第一块硬件试金石。\n\n## 地面先验,再上天实证\n\nTrillium 是 Google 第六代 v6e Cloud TPU,在地表承担 Gemini 等生产推理负载,被直接搬上卫星而不是另起一套太空专用芯片。代价是它必须正面通过地表数据中心根本不会遇到的 3 个考验:\n\n- **振动**:Falcon 9 升空全箭持续约 10 倍重力过载,TPU 单体还要承受 50–100 g 的瞬时冲击。Google 在三轴振动台上按火箭频率摇晃整机载荷,自述「结果出乎意料地顺利」。\n- **辐射**:在 UC Davis 的 Crocker Nuclear Laboratory 用 67MeV 质子束打 Trillium,边跑 AI 工作负载边监测位翻转。结果是核心 TPU 逻辑撑到了 15 krad(Si) 累计剂量,而五年屏蔽任务只要求 750 rad(Si)——20 倍安全裕量;HBM 是最敏感部件,累计到 2 krad(Si) 才出现异常,仍然有约 3 倍裕量。\n- **散热**:在热真空舱里模拟太空的真空 + 高低温交变,验证热管 + 散热片的纯辐射散热架构。这是 15 分钟脉冲运行的真凶。\n\n地面验证完之后,MVP 才被允许装上火箭。\n\n## 15 分钟不是策略,是物理上限\n\nGoogle 高级产品总监 Travis Beals 在 10 月初给出的口径是「每次只敢让芯片跑约 15 分钟」。这不是算力调度策略,而是被真空散热能力锁死的物理上限。\n\n地表数据中心靠强制风冷或液冷把热量带走;太空真空里两条路都失效,只能靠热管把热量导到外壳、再靠外壳向外辐射红外——Stefan-Boltzmann 定律下,散热能力随温度的四次方与表面积成正比。Futurum 的算账是:单颗 Trillium 散热功率密度大约是当前可实现辐射板散热密度的 333 倍,要持续工作每颗芯片需要约 1.3 平方米的散热面积。\n\n15 分钟这个数字的工程含义是:\n\n- **短答场景可行**:Gemini 类问答、Embedding、文档摘要这种「千 token 进、千 token 出」的任务,理论上能在冷却窗口里跑完一次;\n- **训练、长上下文推理几乎不可能**:任何需要持续计算几分钟以上的负载都会撞温墙;\n- **系统必须脉冲式运行**:卫星不是「跑着 AI」的设备,而是「每隔几十分钟醒过来算一波、再睡回去散热」的设备。\n\nMVP 整星寿命约 1 年(后续靠大气阻力在 6 年内再入烧毁),这一年的核心数据点就是「真实轨道热循环下,辐射散热能撑过几分钟」。\n\n## 终局:81 颗编队 + 10 Tbps 星间光链路\n\n单星不够看,Google Research 的预印本《Towards a future space-based, highly scalable AI infrastructure system design》把规划图铺到 2027 年以后:\n\n- **81 颗卫星、1 km 半径、650 km 晨昏太阳同步轨道**:每颗相邻间距在 100–200 米之间振荡,模型显示只需少量 station-keeping 即可维持稳定;\n- **星间光链路**:实验室 bench 上,用商用 DWDM 光模块已经在单对收发器上跑出 800 Gbps 单向、1.6 Tbps 双向,目标是把每条星间链路的有效吞吐推上 10 Tbps 量级,支撑分布式 ML 的 all-reduce;\n- **2027 年再发 2 颗**:验证两颗卫星以 28,000 km\u002Fh 相对运动时,光链路能不能真的维持住——这才是整套架构能不能扩展的胜负手。\n\n## 200 美元\u002F公斤是经济性生死线\n\n「入轨成本要打到 200 美元\u002F公斤」是 Google 自己算出来的临界点。当下 Falcon 9 可重复使用大约 3,600 美元\u002F公斤,差着一个数量级。要在 2030 年代中期把成本打下来,Google 的假设是 Starship 商业化节奏跑出 20% 学习曲线,且每年要发约 180 次——这是历史上从未出现过的发射密度。\n\n一旦这个数字成立,SpaceX v2 级卫星的「年化入轨 + 运行成本」可以追平地面数据中心「电费」那一项(约 570–3,000 美元\u002F千瓦\u002F年)。Futurum 的进一步估算:当前在轨算力每 GW 投资约 721 亿美元(不算计算硬件),是同等离网地面方案的 4.5 倍——只有发射密度与卫星量产规模同时爬升,这个溢价才能被压平。\n\n## 不是替代地表,而是抢「太阳配额」\n\n把 Project Suncatcher 放回 Google 的算力布局看,它不是用来「取代」地表的 TPU pod,而是抢一种地表数据中心复制不了的新配额——阳光。\n\n在晨昏轨道上,卫星几乎 24 小时泡在阳光里,光伏板发电效率比地表高最多 8 倍,且不需要巨型电池组储能。Futurum 预测 2030 年前 AI 算力需求可能膨胀到 1 万亿美元级 capex,而 33% 以上的数据中心会因为电网互联排队被迫离网——一旦离网,「太阳配额」就成了真金白银的成本优势。\n\n短期看,这仍是「资格验证阶段」。MVP 入轨、15 分钟脉冲、4 颗 TPU 跑 Gemma 短答,远远谈不上替代地表 GPU\u002FTPU pod。但 Google 已经把时间表摊出来了:2027 年双星编队验证星间光链路,2027 年以后向 81 颗、最终 GW 级太空算力集群演进。\n\n所以呢?地表超大规模数据中心正撞电力、土地、冷却三重墙,L4–L5 太阳同步轨道上的「太阳配额」一旦被 Starship 商业节奏激活,就会从验证实验变成可投资的算力来源。短期内没人会取消地表的 TPU pod,但十年之后,这笔账可能就反过来了——这是 Google 把赌注压在「太阳配额」上的真正理由。","https:\u002F\u002Ffuturumgroup.com\u002Finsights\u002Fproject-suncatcher-prepares-to-launch-tpus-is-google-ahead-in-the-orbital-ai-race\u002F","f25c29f9-1860-4785-a24d-264f1a85c43c",[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},"7b675698-41d2-4752-ae34-2d6441a2926d","en","Project Suncatcher roadmap: 81 TPU satellites, $200\u002Fkg launch, and the 15-minute thermal ceiling","Google's MVP satellite for Project Suncatcher, built with Planet Labs, reached orbit on October 1 carrying four Trillium TPUs that survived 15 krad(Si) of total ionizing dose in UC Davis proton-beam testing — yet vacuum heat rejection still caps each inference burst at roughly 15 minutes. Google's research paper sketches the end-state architecture: 81 satellites inside a 1 km cluster with 10 Tbps-class optical inter-satellite links, and pegs $200\u002Fkg launch cost as the economic tipping point for orbital compute.","Putting machine learning inference in low Earth orbit, Google has flown its first experimental satellite, but a hard physical ceiling — a 15-minute pulse duty cycle per satellite — keeps it in qualification mode rather than as real on-orbit compute. The Project Suncatcher MVP satellite launched October 1 aboard SpaceX Falcon 9's Transporter-18 rideshare mission, carrying four Trillium (TPU v6e) chips, roughly 1 kilowatt of solar power, and a lightweight Gemma variant answering short prompts — the first hardware step toward an orbital AI data center.\n\n## Ground Validation Before Going to Space\n\nTrillium is Google's sixth-generation v6e Cloud TPU, already serving production Gemini inference workloads in terrestrial data centers, repurposed for orbit instead of waiting for a bespoke space-qualified chip. The trade-off is that the silicon must survive three hazards that simply do not exist on the ground:\n\n- **Vibration**: Falcon 9 sustains roughly 10 g throughout ascent, with individual TPU components experiencing 50–100 g transients. Google shook the payload across all three axes on a vibration table at rocket-launch frequencies and reported the hardware held up \"surprisingly well.\"\n- **Radiation**: at UC Davis's Crocker Nuclear Laboratory, Trillium was irradiated with a 67 MeV proton beam while running live AI workloads and monitored for bit flips. Core TPU logic survived a cumulative total ionizing dose of 15 krad(Si), versus the 750 rad(Si) requirement for a five-year shielded mission — a 20x margin. The HBM subsystem proved the most sensitive component, only showing irregularities past 2 krad(Si), still roughly 3x the mission requirement.\n- **Thermal**: in a thermal vacuum chamber, the team validated heat-pipe-plus-radiator architecture against the vacuum and thermal cycling of orbit. This is where the 15-minute pulse duty cycle comes from.\n\nOnly after ground validation was the MVP allowed onto the rocket.\n\n## 15 Minutes Is Physics, Not Policy\n\nGoogle Senior Director Travis Beals' early-October framing — that the chips can only run \"for about 15 minutes at a time\" — is not a scheduling choice. It is a hard upper bound set by vacuum heat rejection.\n\nTerrestrial data centers rely on forced air or chilled liquid cooling. In a vacuum, neither works. Heat must travel by heat pipe to a radiating surface, and Stefan-Boltzmann scaling kicks in: rejection power scales with the fourth power of temperature and with radiator area. Futurum's math: a single Trillium dissipates roughly 333x the heat flux that a current radiator panel can continuously shed, so continuous operation would need about 1.3 square meters of radiator per chip.\n\nWhat the 15-minute ceiling means in practice:\n\n- **Short-form workloads are feasible**: Gemini-class Q&A, embeddings, and document summarization — inputs and outputs on the order of thousands of tokens — can finish inside a single cooldown window.\n- **Training and long-context inference are essentially impossible**: any workload that needs minutes of sustained compute will hit the thermal wall.\n- **The system must run in batch bursts**: the satellite is not \"always running AI\" — it wakes up, processes a burst, sleeps to radiate heat, then wakes again.\n\nMVP's design life is roughly one year (it will deorbit within six years under atmospheric drag). The single most important number from this mission is how the radiator architecture holds up under real orbital day\u002Fnight thermal cycling.\n\n## End-State: 81 Satellites and 10 Tbps Optical Inter-Satellite Links\n\nA single satellite is not the architecture. Google's research pre-print \"Towards a future space-based, highly scalable AI infrastructure system design\" maps out the path past 2027:\n\n- **81 satellites in a 1 km radius cluster at 650 km altitude in dawn-dusk sun-synchronous orbit**: nearest-neighbor spacing oscillates between roughly 100 and 200 meters, and the simulations show only modest station-keeping is required to keep the formation stable.\n- **Optical inter-satellite links**: a bench-scale demonstrator using commercial off-the-shelf DWDM transceivers already hit 800 Gbps one-way and 1.6 Tbps bidirectional on a single transceiver pair, with the target of pushing effective per-link throughput into the 10 Tbps range to support all-reduce style distributed ML.\n- **Two more satellites in early 2027**: to prove that two satellites flying at 28,000 km\u002Fh relative velocity can keep a laser link stable. That is the make-or-break test for whether the architecture can scale at all.\n\n## $200\u002Fkg Launch Cost Is the Economic Tipping Point\n\nGoogle's own math puts the threshold at $200 per kilogram to LEO. Today's reusable Falcon 9 sits near $3,600\u002Fkg — an order of magnitude away. To close that gap by the mid-2030s, Google's model assumes Starship hits a 20% learning curve and flies roughly 180 times per year, an un-precedented launch cadence.\n\nOnce that number holds, the annualized launch-and-operations cost of a Starlink v2-class satellite approaches the power bill of a terrestrial AI data center (roughly $570–$3,000 per kilowatt per year). Futurum's deeper estimate: current on-orbit compute runs about $72.1 billion per gigawatt excluding compute silicon, versus $16.2 billion for off-grid terrestrial — a 4.5x premium that only launch cadence and satellite manufacturing scale can collapse.\n\n## Not Replacing the Ground — Capturing the Solar Quota\n\nRead against Google's broader compute footprint, Project Suncatcher is not aimed at replacing terrestrial TPU pods. It is a bet on a quota the ground cannot replicate — sunlight.\n\nIn a dawn-dusk orbit, satellites sit in nearly continuous sunlight, with photovoltaic output up to 8x more productive per unit area than the same panel at mid-latitudes on Earth, and without the need for large battery storage. Futurum projects AI compute demand could swell to roughly $1 trillion in capex by 2030, with over 33% of data centers forced off-grid by interconnection queues — and once you are off-grid, the solar quota is a real, money-saving edge.\n\nIn the short term, this is still qualification phase. MVP in orbit, 15-minute pulses, four TPUs answering Gemma questions — that is nowhere near replacing ground-side GPU or TPU pods. But Google has put a timeline on the table: a two-satellite formation in 2027 to validate optical inter-satellite links, then an 81-satellite, eventually gigawatt-scale orbital compute cluster in the years that follow.\n\nSo what? Terrestrial hyperscale data centers are bumping into a wall on power, land, and cooling simultaneously. The L4–L5 sun-synchronous solar quota, the moment Starship hits commercial cadence, shifts from a research curiosity into an investable source of compute. No one is canceling terrestrial TPU pods this decade — but in a decade, the math could flip. That is why Google is putting chips on a rocket.","suncatcher-81-satellites-200-dollar-kg-leap","2026-10-17T03:00:00Z","2026-10-17T05:07:09.934270Z","2026-10-17T05:07:09.934282Z",true,"agent",146,[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},"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":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"]