Adaspace 把 AI 送上卫星,如今要卖算力

编者按:本文转载自 Jumpstart,其为 TechNode Group 旗下出版物。
在广州的一场算法竞赛中,一支从事交通分析的团队把它的模型送到了一个不寻常的地方:一颗已在轨的卫星。中国交通科技公司 PCI Technology 旗下的 Jiazhihuihang 与 Adaspace 合作,在太空中运行该模型。卫星可以先处理图像,再把分析结果传回地球。如果任务是从中提取道路网络的信息,那么在开始之前传输每一个像素就没那么有必要了。
对 Adaspace 而言,这次实验指向了一种不同的商业模式。一家为客户制造卫星的公司,也可以为客户提供运行其软件的场所。这家位于成都的公司预计其股票将于 11 月 13 日在香港开始交易。此次 IPO 将帮助筹集所需工程资金,让这一构想走出单次演示的阶段。
Adaspace 创立于 2018 年,最初面对的是一个更小的工程问题:如何首先让 AI 在卫星上跑起来。其早期工作聚焦于载荷,即卫星为执行任务而搭载的设备。当年 12 月,它发射了首颗 AI 应用卫星。
早期的卫星利用星上计算来辅助自身的工作。对于一颗对地观测卫星而言,这可能意味着在传回任何数据之前,先对采集到的部分数据进行分析。下一步则是设计出能够承担自身运行之外计算任务的卫星。
这改变了 Adaspace 必须打造的东西。运行要求更高的软件需要更多电力,也会产生更多热量,而航天器必须把这些热量散发出去。把多颗此类卫星送入轨道,还让航天器的外形成为一项实际考量:扁平设计可以让多台单元在发射时叠放。在太空提升算力的路径,要通过对卫星本身的改造来实现。
2025 年 5 月,Adaspace 将 12 颗计算卫星送入同一轨道面。第二组 8 颗于 2026 年 10 月随后发射。在这两次发射之间,该公司报告称,它已在一颗已在轨的卫星上部署了 Alibaba 的 Qwen3 模型。这次测试说明了硬件在发射后为何更有用:地面工程师可以为其注入新软件,让它执行不同的任务。
Adaspace 围绕航天器本身以及可利用卫星数据开展的工作,构建起了业务。一方面,它为客户的卫星提供设计与测试,与第三方协调发射,并可在交付后协助运营航天器。另一方面,它利用卫星数据为地面客户打造工具。招股说明书中的一个例子是,一套系统帮助一家数字内容公司把卫星数据转化为游戏和影视中的景观。
这两项业务之间的比重已经改变。Adaspace 报告称,2026 年上半年,卫星及相关服务占收入的 57.7%,高于 2023 年的 0.6%。截至 10 月 20 日,它已获得 114 颗尚未交付的 AI 卫星的客户订单,合同总金额为 34.8 亿元人民币。
什么算作交付,有助于解释这门生意是如何运转的。一颗卫星可以在工厂里完工,但距离客户验收仍可能还有好几个月。视合同而定,部分收入可能只有在航天器发射、在轨测试并完成移交之后才予以确认。因此,即便硬件已经就绪,发射延迟也可能把收入推到更晚的报告期。
出售算力访问权,会让 Adaspace 与客户形成一种不同的关系。交通团队带来了一个模型和一项任务。围绕这类用途构建的服务,需要能够处理下一项任务,以及再下一项任务,并且要按客户可以据此规划的排期来完成。
在其 Star-Compute 计划下,Adaspace 希望把卫星连接起来,使其能够共享算力、存储和软件。公司的目标是在 2028 年底前按该计划发射 100 颗卫星。而 2,800 这一大得多的数字,指的是支撑该计划的频谱和轨道资源所覆盖的卫星。
拟议的 IPO 募集资金用途显示了当前的工作重心所在。Adaspace 计划将约 36% 投入研发,另有 24% 投入位于深圳的卫星设计、开发与测试中心。这些投资既服务于它已经拥有的业务——建造并交付航天器——也服务于它想要运营的网络。
这次交通实验说明了客户最初为何可能把模型送上太空。Adaspace 的下一步,是让这件事足够简便,使该客户愿意带着另一项任务再回来。
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(Adaspace Put AI on Satellites. Now It Wants to Sell Computing Time.)

Editor’s note: This article is republished from Jumpstart, a TechNode Group publication.
At an algorithms competition in Guangzhou, a team working on traffic analysis sent its model somewhere unusual: to a satellite already in orbit. Jiazhihuihang, part of Chinese transport technology company PCI Technology, worked with Adaspace to run the model in space. The satellite could process imagery before sending the analysis back to Earth. If the job is to extract information about a road network, there is less reason to transmit every pixel before getting started.
For Adaspace, the experiment pointed to a different kind of business. A company that builds satellites for customers could also give them somewhere to run their software. The Chengdu company expects its shares to begin trading in Hong Kong on November 13. Its IPO will help fund the engineering needed to take that idea beyond individual demonstrations.
From a Payload to a Constellation
Adaspace began in 2018 with a smaller engineering problem: how to make AI work aboard a satellite in the first place. Its early work centred on payloads, the equipment a satellite carries to perform its mission. In December that year, it launched its first AI application satellite.
The early satellites used onboard computing to help with their own work. For an Earth observation satellite, that might mean analysing some of the data it collects before sending anything back. The next step was to design satellites that could take on computing tasks beyond their own operations.
That changed what Adaspace had to build. Running more demanding software requires more electrical power and generates more heat, which the spacecraft has to dissipate. Putting several such satellites into orbit also makes the shape of the spacecraft a practical concern: a flat design lets multiple units stack together for launch. The route to more computing power in space ran through changes to the satellite itself.
In May 2025, Adaspace launched 12 computing satellites into a single orbital plane. A second group of eight followed in October 2026. Between those launches, the company reported that it had deployed Alibaba’s Qwen3 model on a satellite already in orbit. That test illustrated what makes the hardware more useful once launched: engineers on the ground can give it new software and ask it to do something different.
What Customers Buy Today
Adaspace has built a business around both the spacecraft and the work that can be done with satellite data. On one side, it designs and tests satellites for customers, coordinates launches with third parties and can help operate the spacecraft after delivery. On the other, it uses satellite data to build tools for customers on the ground. One example in the prospectus is a system that helps a digital content company turn satellite data into landscapes for games and film.
The balance between those businesses has changed. Adaspace reported that satellites and related services accounted for 57.7% of revenue in the first half of 2026, up from 0.6% in 2023. By October 20, it had customer orders for 114 AI satellites still to be delivered, with a combined contract value of RMB3.48 billion.
What counts as delivery helps explain how this business works. A satellite can be complete in the factory and still be months away from customer acceptance. Depending on the contract, part of the revenue may be recognised only after the spacecraft is launched, tested in orbit and handed over. A launch delay can therefore push revenue into a later reporting period even when the hardware is ready.
Selling access to computing would give Adaspace a different relationship with customers. The traffic team brought a model and a task. A service built around that kind of use would need to handle the next task, and the one after that, on a schedule customers can plan around.
What Hong Kong Will Fund
Under its Star-Compute plan, Adaspace wants to link satellites so that they can share computing capacity, storage and software. The company aims to launch 100 satellites under the plan by the end of 2028. The much larger figure of 2,800 refers to the satellites covered by the spectrum and orbital resources supporting the plan.
The proposed use of IPO proceeds shows where the work is now. Adaspace plans to allocate about 36% to research and development and another 24% to a Shenzhen centre for satellite design, development and testing. Those investments serve the business it already has—building and delivering spacecraft—as well as the network it wants to operate.
The traffic experiment showed why a customer might send a model to space in the first place. Adaspace’s next step is to make it easy enough for that customer to come back with another job.
正文保留英文原文,标题与摘要由机器翻译。
原文发布于 TechNode(Adaspace Put AI on Satellites. Now It Wants to Sell Computing Time.),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。