文 | 斯凯有话说
很多人还没看懂，朱雀二号新发射的这一组卫星，真正的变革不在上天，而在算力模式的切换。从过去的天感地算，数据传回地面再处理，到如今天数天算，在太空就地完成计算再下传，国星宇航星算计划正在搭建属于中国的太空算力枢纽，这到底会改写哪些产业规则？
2026年10月14日晚上7点40分，东风商业航天创新试验区，朱雀二号改进型遥四火箭点火升空，将10颗卫星送入预定轨道。
这10颗卫星里，有8颗属于国星宇航的“星算”计划，另外2颗是蓝塔智能的光传试验星。
新闻稿里写的是“发射圆满成功”。但真正值得说的，不是发射本身。是这10颗卫星要去做的事。
一个被忽略的事实：95%的太空数据，被扔了
先看一组数字。
据央视新闻报道，我国在轨民商用遥感卫星超过640颗，规模世界第二。每天，这些卫星在太空中拍摄海量图像——农田、城市、海洋、灾害现场。
但受星地下传带宽限制，真正传回地面的数据不足5%。
剩下的95%，要么被压缩到失真，要么直接被丢弃。
我们花了那么多钱发射卫星、维护星座、建设地面站，结果大部分数据从来没被看过一眼。
这是一个极其荒诞的现实：太空中有几百双“眼睛”在盯着地球，但我们只能看到它们看到的5%。
传统卫星的工作方式是“天感地算”——卫星负责拍，拍完把数据传回地面，地面再算。但传输带宽是有限的，卫星过顶的时间窗口只有几分钟到十几分钟。数据传不完，就只能扔掉。
“天数天算”，就是在解决这个问题。 让卫星获取数据后，在轨直接完成AI处理。只把“结果”传回来，不传“原始素材”。
农田监测，卫星自己判断哪里缺水、哪里有病虫害，只传回结论。灾害救援，卫星自己识别受灾区域、评估受损程度，只传回坐标和图像。远洋监测，卫星自己锁定可疑船只，只传回位置和航向。
数据不用全回来，有用的才回来。
这次发射的10颗卫星，在验证什么
“星算”计划02组卫星搭载了多项关键技术载荷。
高性能计算载荷，支持AI模型在轨部署与推理。高速星间/星地激光通信载荷，实现卫星间高速数据交换。微光相机，用于遥感成像。
具体验证的技术方向包括：万瓦级液冷流体回路，解决太空环境下高性能计算的散热问题；超大尺寸展开式辐照冷翼，利用太空低温环境被动散热；全域激光通信网络，让卫星之间可以高速交换数据；基础大模型在轨训练，探索在太空环境中进行AI模型训练。
琶洲太空智算星组首批两颗算力卫星同步入轨。该星组规划共24颗卫星，将在3年内投运，培育百亿级太空计算产业链。
据央视网报道，琶洲实验室技术负责人说了一句话，把这件事讲透了：“我们的目标是用天上的数据在太空训练大模型，同时推理的输入也来自天上的数据，让整个闭环在天上完成，成为原生的‘天基大模型’。”
翻译一下：过去是卫星拍完照传回来，地面用AI分析。未来是卫星自己拍、自己算、自己判断，只把结论传回来。卫星从一个“传感器”，变成了一个“智能体”。
这不是一家公司在做的事
“星算”计划由国星宇航联合108家高校、科研院所、上市公司及投资机构共同建设。
规模：由2800颗计算卫星组成，其中2400颗为推理计算卫星，400颗为训练计算卫星。
算力目标：实现十万P级的推理算力和百万P级的训练算力，组网形成覆盖全球陆海空天的训推算力网。
时间表：2030年前完成千星规模组网和商用，超95%为推理计算卫星；2035年前完成全部组网，实现服务数以亿计硅基智能体的计算能力。
随着02组8颗卫星入轨，整个“星算”星座的在轨总算力达到50P，相当于每秒可在太空中完成5亿亿次运算。
这不是孤军奋战。
之江实验室发起的“三体计算星座”，是千星规模的太空计算基础设施，建成后总算力可达1000POPS。2025年5月14日，首发12颗计算卫星成功入轨，经过9个月的在轨测试，已形成组网、计算、模型部署、科学载荷在轨验证等四大核心能力，完成10个人工智能模型与应用的部署与验证。
北京邮电大学天算星座团队自主研发的太空服务器，搭载“北邮二号”“北邮三号”卫星在轨稳定运行近一年，实现图像解析信息获取效率提升50倍，并验证了6G星载核心网在严苛太空环境下的稳定运行能力。
全球首个太空算力云已正式常态化提供在轨算力服务。自投入服务以来，累计完成数百次太空算力调用，服务用户超过百家，大模型推理能效比达到10词元/焦耳，太空算力服务覆盖率超过10%。
天基光计算也在启动。上海东方天算科技与光本位智能科技宣布联合建设天基光计算创新中心，启动全球首颗天基光计算卫星的联合研制。以光子为计算和传输载体，可支撑在轨AI推理、星上大模型运行、万亿级硅基智能体协同服务。
国家层面，据科技日报报道，2026年4月，国防科工局商业航天司副司长于国斌在太空算力产业大会上透露，国家航天局牵头举办了太空智能算力星座专项论证启动会和专家组会议，此项工作正在有序开展。
为什么这件事重要
太空算力的核心价值，不在“算力”本身，在响应速度。
传统卫星的工作链路是这样的：卫星拍照→数据传回地面→地面AI分析→生成决策建议→上传回卫星或下发给用户。这个链路的时间，以小时甚至天计算。
但在应急救灾场景里，需要的是分钟级响应。
地震发生后，黄金救援时间是72小时，但真正能救人的是前几个小时。洪水来临时，几小时内就要决定往哪里派救援队。森林火灾，风向变了就要立刻调整扑救方案。
如果卫星“天数天算”，拍照、分析、决策建议的链路可以压缩到几分钟。卫星在天上就把活干了，地面只需要接收“往哪里走、救谁、怎么救”。
这就是太空算力最核心的商业逻辑：不是把算力送上天，是把决策送上前线。
但挑战依然存在
“天数天算”的技术路径清晰，但工程难度不小。
散热是第一道坎。 太空是真空环境，散热只能靠辐射。高性能计算芯片在太空中产生大量热量，如何高效散热，决定了算力能堆多高。这次发射验证的万瓦级液冷流体回路和超大尺寸展开式辐照冷翼，就是在解决这个问题。
能源是第二道坎。 太空算力需要稳定的电力供应。卫星的太阳能帆板面积有限，算力越大，功耗越高。如何在有限的帆板面积下支撑高性能计算，是工程上的硬骨头。
星间通信是第三道坎。 单颗卫星的算力有限，多颗卫星协同才能完成复杂任务。但卫星之间高速交换数据，需要激光通信网络。这次发射的蓝塔智能光传卫星，就是在验证这个能力。
商业模式是第四道坎。 太空算力的成本远高于地面算力。什么场景愿意为“太空算力”支付溢价？目前来看，应急救灾、远洋监测、国防安全是优先方向。但这些场景的共同特点是：对响应速度的敏感度远高于对成本的敏感度。 一次地震的黄金救援窗口只有几小时，远洋上的可疑船只不会等你把数据传回地面再分析——在这类场景里，太空算力的溢价是有支付方愿意承接的。但大规模商业化，还需要找到更多愿意付费的日常场景。
写在最后
回到那个95%。
我们每天扔掉95%的太空数据，不是因为那些数据没有价值，是因为我们没有能力把它们“用起来”。太空算力要解决的，不是“算得快不快”，是“数据能不能被用上”。
从“天感地算”到“天数天算”，本质上是从“收集数据”到“生产决策”的跃迁。
这次朱雀二号发射的10颗卫星，只是一个开始。“星算”计划2800颗卫星的组网目标，2035年完成。三体计算星座、太空算力云、天基光计算——多条技术路线正在并行推进。
太空算力的竞赛，比的不是谁先把算力送上天，是谁先让天上的算力“有用”。
95%的数据还在被扔掉。但至少，我们开始动手去捡了。
太空算力，不是简单的“算力上天”，是人类卫星从被动传感器到太空智能体的时代拐点。95%的太空数据从废弃闲置、到在轨生智、分钟级决策，中国商业航天的下半场，早已不是拼发射次数，而是拼天上的算力闭环。
我会持续追踪太空算力、天基AI、商业航天赛道的真实落地进度与产业红利，不吹概念、只讲硬核逻辑。关注我，看懂中国太空新基建的底层变局。
免责声明
本文所有内容仅为行业资讯与产业逻辑分析，不构成任何投资建议、商业决策建议或操作指导。
文中数据、案例及产业进展均来源于公开权威渠道，作者不对信息的实时性、完整性、适用性做绝对保证。
资本市场、科技产业存在不确定性与政策风险，读者据此做出的一切决策，风险自行承担。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Man, Yan, Skye has something to say.
Many people have yet to understand that the new group of satellites launched by the Chuan II is the real change not in the sky, but in the shape of the algorithm. From the past, when data were sent back to the ground for reprocessing, to today ' s days, when calculations were done on the ground in space and then down, the National Astronaut Project was setting up a space computing hub for China.
On the evening of 14 October 2026, at 7.40 p.m., the commercial East Wind commercial space innovation test area, the Peacock II modified remote four rocket, set fire to 10 satellites into their scheduled orbit.
Of these 10 satellites, eight are the "Star Count" programme of the National Planetary Space and two are the light-transmitting test stars of the Blue Tower.
The press release states that “the launch was successful”. But it's really worth saying, not the launch itself. It's what these 10 satellites are going to do.
A neglected fact: 95 per cent of space data was thrown away.
Look at a set of numbers first.
According to press reports, there are more than 640 commercial remote sensing satellites in orbit, the second largest in the world. Every day, these satellites take large-volume images in space — agricultural land, cities, oceans, disaster sites.
However, less than 5 per cent of the data are actually sent back to the ground, limited by the bandwidth of the stars.
The remaining 95 per cent were either compressed to a distortion or simply discarded.
We spent so much money launching satellites, maintaining constellations and building ground stations that most of the data was never seen.
This is an extremely ridiculous reality: hundreds of eyes in space are staring at the Earth, but we can only see 5 per cent of what they see.
The traditional satellite works by “sky calculations” — satellites are responsible for filming, back-to-ground data and ground recalculation. The transmission bandwidth is limited, however, and the time window of the satellite is only a few minutes to ten minutes. You can't get the data out, you have to throw it away.
The “days of counting” is addressing this problem. Once the data are obtained by satellite, AI processing is completed directly in orbit. Only the results are sent back without the original asset.
Agricultural monitoring, where satellites determine for themselves where there is a shortage of water and where there are pests and pests, only conveys the conclusions. Disaster relief, where the satellite itself identifies the affected area, assesses the damage, and returns only coordinates and images. Oceanic monitoring, the satellite itself locks into suspicious vessels and returns only its position and course.
The data doesn't have to come back. It's useful.
Ten satellites launched this time.
Several key technology payloads were carried by the Star Count 2 satellite group.
High performance computing payloads to support AI model deployment and reasoning in orbit. High-speed interplanetary/terrestrial laser communication payload to achieve inter-satellite high-speed data exchange. Microlight cameras for remote sensing imaging.
Specific technical directions for validation include: the Vanwa class liquid cryopathic circuit, which addresses the problem of high performance-calculated dispersion heat in the space environment; ultra-large-scale extended irradiation cold wing, which uses passive dispersion in the low-temperature space environment; a global laser communication network, which allows for high-speed exchange of data between satellites; and basic LLM training in orbit to explore AI model training in the space environment.
The first two satellites of the Biwako Space Intellectuals are in orbit simultaneously. The group plans a total of 24 satellites that will be launched within three years to foster a billion-class space computing industrial chain.
According to OCNET, the technical director of the Biwako Laboratories said, “Our goal is to train LLM in space with data from the sky, and the reasoning input is also from the sky, so that the entire closed circle can be completed in the sky and become the original LLM.”
Translator: Used to be a satellite shot back, using an AI analysis on the ground. The future is for the satellite to film itself, to calculate itself, to judge for itself and to pass its conclusions back. The satellite evolved from a “sensor” to a “AI agent”.
It's not what a company is doing.
The “Star Count” project was built by the United National Astronautical Federation of 108 universities, research institutes, listed companies and investment agencies.
Size: Composed of 2,800 computing satellites, of which 2,400 are reasoned computing satellites and 400 are training computing satellites.
Objective: To achieve 100,000 P-level reasoning and million P-level training calculations, and to form a network of trainers covering global land, sea and air space.
Timeline: 1,000-star network and commercials by 2030, over 95 per cent of the reasoned calculation satellites; full network by 2035, realizing the computational capability of hundreds of millions of silicon-based AI agent services.
With the orbiting of a group of 8 satellites from 02, the overall “star-calculated” constellation had a total in orbit of 50P, equivalent to 500 million calculations per second in space.
It's not alone.
The “Triple Calculating Constellation” launched by the Noboa Laboratory is a thousands-star-scale space computing infrastructure with a total capacity of 1000 POPS. On 14 May 2025, 12 computing satellites were launched successfully and, after nine months of on-orbit testing, four core capabilities, such as a network, computing, model deployment, and in-orbit validation of scientific payloads, were formed, completing the deployment and validation of 10 individual industrial intelligence models and applications.
The autonomous space server developed by the Constellation Constellation team of Beijing Post and Telecommunications University, which carries the North Mail II satellite, North Mail III, in orbit for almost a year, has resulted in a 50-fold increase in the efficiency of image analysis information acquisition and has validated the stability of the 6G core network in a demanding space environment.
The world ' s first space computing cloud has been formalized to provide on-orbit computing services. Since the service was launched, hundreds of space computing calls have been cumulatively completed, with more than 100 service users, LLM reasoning efficiency ratio of 10 words/joules and space computing service coverage of over 10 per cent.
Space-based light computing is also being activated. In Shanghai, the East Sky Calculator Technology and Optical Intelligent Technology announced a joint construction of a space-based light computing innovation centre to launch the joint development of the first space-based optical computing satellite worldwide. Photons are used to calculate and transfer vehicles, which support on-orbit AI reasoning, the operation of LLM on stars, and the hundreds of billions of silica AI agent synergetic services.
At the national level, according to the daily Science and Technology newspaper, in April 2026, the Deputy Director of the Commercial Space Division of the National Defense Agency, in Guobin, revealed at the Space Calculator Industry Congress that the National Space Agency had led the launch of a special space intelligence calculator constellation and an expert group meeting, which was proceeding in an orderly manner.
Why is this important?
The core value of space algorithms is not “calculations” per se, but response speed.
The working links of traditional satellites are as follows: satellite photo-data is sent back to ground-based AI-generation decision-making recommendations are uploaded back to satellites or sent down to users. The time of this link, in hours or even days.
In emergency response scenarios, however, a minute-scale response is required.
After the earthquake, the time for gold relief was 72 hours, but what really saved lives was the hours before. The flood came at short notice and in a matter of hours decided where to send the rescue team. In case of forest fires, when the wind changes, the rescue programme must be adjusted immediately.
If the satellite is “days of counting”, the chain of photography, analysis, decision-making advice can be compressed to a few minutes. Satellites work in the sky and the ground simply receives “where to go, who to save, how to save”.
That is the business logic that lies at the heart of space computing: it is the decision that goes to the front.
But challenges remain.
The “days-by-day” technical path is clear, but the project is difficult.
The heat spread is the first can. Space is a vacuum environment where the heat can only be dispersed by radiation. High-performance calculated chips generate large amounts of heat in space, and how the heat is distributed efficiently determines how high the computing power is. The problem is being addressed by the Vanwa class liquid cryopathic circuit and ultra-large-scale extended irradiated cold wings, which were verified during the launch.
Energy is the second option. Space computing requires a stable power supply. The size of the satellite's solar sails is limited and the greater the calculator, the greater the amount of effort. How to support high performance calculations under a limited area of sailboard is the hard bone of the project.
Interstellar communication is third. A single satellite has limited computing power, and multiple satellites work together to carry out complex tasks. However, the rapid exchange of data between satellites requires a laser communication network. The launch of the Blue Tower Smart Optical Satellite is proving this capability.
The business model is fourth. The cost of space computing is much higher than that of ground computing. What kind of scenario is willing to pay a premium for space algorithms? Emergency response, ocean monitoring and national defence security are currently priorities. The common feature of these scenarios, however, is that responsiveness is much more sensitive than cost. The gold rescue window for an earthquake is only a few hours away, and suspicious ships in the open ocean will not wait for you to send the data back to the ground for analysis — in such scenarios, the premium on space computing is paid. But large-scale commercialization also requires more routine scenes of paid.
At the end.
Back to that 95%.
We throw out 95 per cent of space data every day, not because they are worthless, but because we do not have the capacity to “use them”. Space algorithms are not “calculated fast or not”, they are “already used”.
From “sky” to “days”, essentially a leap from “data collection” to “production decision-making”.
The launch of 10 satellites this time by the Sparrow II was just the beginning. The cluster target of the 2,800 satellites of the Star Count project was completed in 2035. Triangular calculator constellations, space computing clouds, space-based light computing - multiple technology routes are advancing in parallel.
The game of space computing is better than who sends it to heaven first, who gives it “help”.
Ninety-five per cent of the data is still being thrown away. But at least we started picking it up.
Space algorithms, rather than simple “calculations up to heaven”, are the turning point of human satellites from passive sensors to space AI agent. Ninety-five per cent of space data go from idle waste to in-orbit intellectuals and minute-level decision-making, and the second half of China ' s commercial space is long past the number of launch times, but rather the number of algorithms in the sky.
I will keep track of space algorithms, space-based AI, real landings on commercial space tracks and industry dividends, without blowing concepts, but with hard-core logic. Look at me and understand the bottom of China's new infrastructure.
Disclaimer
All of this is only industry information and industry logic analysis and does not constitute any investment advice, business decision-making advice or operational guidance.
The data, cases and industry developments in the text are derived from open sources of authority, and authors do not provide absolute guarantees of the real-time, integrity and applicability of information.
There are uncertainties and policy risks in the capital markets and in the science and technology industry, and all decisions made by readers are based on them, and risk is assumed.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.