THE SHORT VERSION
- Google's Project Suncatcher prototype space data centre is online and communicating with the ground after launching earlier this month on a Falcon 9 alongside more than 100 other spacecraft.谷歌 Project Suncatcher 太空数据中心原型已上线并与地面通信，本月早些时候与 100 多艘航天器一同由猎鹰 9 号发射。
- The satellite carries four Google TPUs and generates up to 1 kW, but the power demand means it can switch on only one TPU for about 20 minutes at a time.该卫星搭载四颗谷歌 TPU，发电功率最高 1 kW；受功耗限制，一次只能让一颗 TPU 运行约 20 分钟。
- After roughly three weeks of commissioning, engineers sent the first prompt to a pre-loaded Gemma 4 model, which produced a haiku; Google said the spacecraft was operating nominally.约三周的调试后，工程师首次向预装的 Gemma 4 模型发出提示并得到一首俳句；谷歌表示航天器运行正常。
A prototype data centre in orbit在轨的原型数据中心
Google's Project Suncatcher prototype space data centre is online and communicating with the ground, after launching earlier this month alongside more than 100 other spacecraft on a SpaceX Falcon 9 rocket. The satellite is intended to pave the way for future networks of close-flying satellites that perform large inference workloads in orbit.谷歌 Project Suncatcher 太空数据中心原型已上线并与地面通信，本月早些时候它与 100 多艘航天器一同由 SpaceX 猎鹰 9 号火箭发射。该卫星旨在为未来近距离编队、在轨执行大规模推理的卫星网络铺路。
The hardware is modest by data-centre standards: four Google Tensor Processing Units and up to 1 kW of generating capacity. That combination is the experiment. A ground data centre treats power as a contract with a utility; a satellite treats it as a design constraint that also governs how long a model can run.按数据中心的标准，其硬件很朴素：四颗谷歌张量处理单元（TPU），发电功率最高 1 kW。这一组合本身就是实验：地面数据中心把电力视为与电网的合约，卫星则把电力视为一项设计约束，而它同时决定模型能运行多久。
One TPU at a time, for about 20 minutes一次一颗 TPU，约 20 分钟
The power demand of the TPU means the satellite can turn on only one of the four for about 20 minutes at a time, according to TechCrunch. That single limit complicates the already difficult business of space-to-ground communications, because a model's answer has to be produced and transmitted inside a window that closes.据 TechCrunch 报道，TPU 的功耗需求意味着卫星一次只能开启四颗中的一颗，持续时间约 20 分钟。这一限制让本就困难的星地通信更难：模型的回答必须在窗口关闭前完成并传回。
The constraint also reframes what the mission is testing. It is not a demonstration that orbital inference can match terrestrial throughput; it is a test of whether useful inference can be scheduled around an intermittent power and link budget. That is a different engineering problem, and the answers will determine whether the concept moves past a prototype.这一限制也重新界定了这次任务在测试什么。它不是在证明轨道推理能比肩地面吞吐，而是在测试「能否围绕间歇性的电力与链路预算安排有用推理」。这是另一个工程问题，答案将决定该概念能否走出原型阶段。
A haiku and an orbital report一首俳句与一份轨道报告
After a commissioning process that tested the satellite's subsystems over about three weeks, engineers were ready to work. Their first prompt this week asked the model pre-loaded onto the satellite, Google DeepMind's Gemma 4, to write a haiku about being in space. The result, as TechCrunch reported it: "Blue marble below / Data streams in silent grace / Starlight feeds my core."在约三周、针对各子系统的调试之后，工程师开始正式使用。本周的第一次提示，是让预装在卫星上的谷歌 DeepMind Gemma 4 模型写一首关于身处太空的俳句。据 TechCrunch 报道，结果为：「Blue marble below / Data streams in silent grace / Starlight feeds my core.」
More consequential was the second task. Google asked the spacecraft to report on the orbital conditions around it, and a spokesperson told TechCrunch that the satellite was operating nominally. A poetry request tests that the link and the model work; a conditions report tests that the satellite can produce useful output from its own environment.更有意义的是第二项任务。谷歌让航天器报告其周围的轨道条件，一位发言人向 TechCrunch 表示卫星运行正常。写诗测试的是链路与模型能否工作；轨道条件报告测试的则是卫星能否从自身环境产出有用结果。
Why put inference in orbit at all为什么要把推理放到轨道上
The stated purpose is a network of satellites flying close together and performing large inference workloads. The appeal is proximity: a satellite can process data where it is collected instead of downlinking raw volume to a ground station, and sunlight in orbit is continuous in a way that terrestrial power contracts are not.公开目标是让近距离编队的卫星网络执行大规模推理。吸引力在就近：卫星可以在数据产生处处理数据，而不必把原始数据量传回地面站；在轨阳光的连续性也不同于地面的电力合约。
The costs are equally structural. Thermal management in vacuum, radiation tolerance, limited mass and launch economics all constrain what can be flown, and the Suncatcher prototype is small enough that its power budget is measured in kilowatts rather than megawatts. The experiment is a first data point on whether those constraints bind at prototype scale or only at fleet scale.代价同样是结构性的。真空散热、抗辐射、有限质量与发射经济性都限制着可送入轨道的东西，而 Suncatcher 原型小到其功率预算以千瓦而非兆瓦计。这次实验给出的是第一个数据点：这些约束在原型尺度就生效，还是只在星座尺度才生效。
What the company said, and what it did not公司说了什么，没说什么
Google published the update on its own blog and confirmed the spacecraft's status through a spokesperson, which makes the basic facts first-party: the satellite is online, the model is pre-loaded, and the operations team is talking to it. Those are verifiable claims about a specific spacecraft.谷歌在自家博客发布更新，并通过发言人确认航天器状态，因此基本事实属第一方信息：卫星在线、模型已预装、运行团队能与其通信。这些都是关于具体航天器的可核实陈述。
What the material reviewed here does not include is any performance measurement: no tokens per second, no link throughput, no comparison with a ground-based TPU, and no schedule for a second satellite. Without those numbers, the prototype establishes that the concept works at all, not how well it works.已读材料中没有的是任何性能数据：没有每秒 token 数、没有链路吞吐、没有与地面 TPU 的对比、也没有第二颗卫星的时间表。缺少这些数字，原型只能证明概念可行，而不能说明可行到什么程度。
What a single satellite cannot settle单颗卫星无法确定的事
A constellation of close-flying satellites is a different problem from one spacecraft. Formation flying requires inter-satellite links, shared scheduling and a way to distribute a workload across nodes, none of which a single prototype can demonstrate. The prototype can only show that one node survives launch, powers up and answers a prompt.近距离编队星座与单颗航天器是不同问题。编队飞行需要星间链路、共享调度以及把负载分配到多个节点的方法，这些都不是单颗原型能演示的。原型只能说明一个节点能经受发射、通电并回应提示。
There is also the question of what workloads justify the trip. Inference that must happen in orbit is inference whose inputs are already there, such as Earth observation, and the prototype does not test an Earth-observation pipeline. It tests a language model writing about orbit, which is a demonstration rather than a workload.另一个问题是哪些工作负载值得上天。必须在轨推理的，是输入本就在轨的那些，例如对地观测；而这次原型并未测试对地观测流程。它测试的是一个写关于轨道的大语言模型，那是演示，不是工作负载。
What would show the concept scales什么能显示概念可扩展
Three things would move the picture: published measurements of inference throughput and link time, a second spacecraft that tests formation flying, and a workload that processes data collected in orbit rather than prompts sent from the ground. None of those is in this week's update.三件事会改变判断：公开的推理吞吐与链路时延数据、一颗用于测试编队飞行的第二颗卫星，以及一个处理在轨采集数据（而非地面发来的提示）的工作负载。本周更新里都没有这些。
Related coverage on this site includes Microsoft's Maia expansion as a 2027 infrastructure bet, Alphabet's AI capex sell-off and Google Cloud's Gemini agent work. All three describe the same question from the ground side: where compute sits, and who pays for it to be there.本站相关报道包括微软 Maia 扩张作为 2027 年的基础设施押注、Alphabet 的 AI 资本开支抛售与谷歌云的 Gemini 智能体工作。三篇都从地面一端描述同一问题：算力放在哪里，以及谁为它待在那里付费。
The link budget is the second constraintThe link budget is the second constraint
Power is only half of the operational problem. The other half is the link: a satellite has to send its answer to a ground station inside a pass, and TechCrunch notes that space-to-ground communications were already difficult before a language model was added to the payload.电力只是运行问题的一半，另一半是链路：卫星必须在一次过境窗口内把回答发回地面站，而 TechCrunch 指出，在载荷里加上语言模型之前，星地通信本就已经困难。
A twenty-minute window per TPU therefore sets the shape of any product. The satellite cannot hold a conversation in the way a chatbot does; it can accept a task, run it, and return a result before the window closes. That is closer to a batch job than to an assistant, and it is the constraint any future constellation would have to design around.因此，每颗 TPU 二十分钟的窗口决定了任何产品的形态。卫星无法像聊天机器人那样持续对话；它能接收任务、运行、并在窗口关闭前返回结果。这更接近批处理作业而非助手，也是任何未来星座都必须围绕设计的那条约束。
Sources & reporting limits来源与核查范围
Follow Industry to find related stories on your next visit.关注产业动态，下次访问时阅读相关文章。
