Can AI data centers operate in space? With imminent test launch, Google is on a mission to find out

As rural communities across the United States wage intensifying battles to keep AI-computing data centers out of their backyards, some tech giants are mulling methods that could sidestep such local resentment — by sending the facilities to space.

There are no guarantees the technology will work, nor is it clear whether sending data centers to Earth orbit will make economic sense. But Google is testing the waters.

The company is set to launch its first test satellite for Project Suncatcher, its orbital data center initiative. The prototype spacecraft is slated to lift off aboard a SpaceX Falcon 9 rocket on a rideshare mission called Transporter-18 at 2:18 p.m. ET (11:18 a.m. PT) Thursday from Vandenberg Space Force Base in California.

It’s not the first time AI chips have been tested in space. But Google’s experimental satellite is designed to figure out specifically how the company’s tensor processing units, or TPUs, perform after undergoing the jarring forces associated with launching to orbit atop a rocket and being battered by the radiation-riddled environment of space.

“This first launch is about seeing what works, identifying points of failure, and applying those findings to future missions,” Travis Beals, Google’s senior director of paradigms of intelligence, said in a news release .

Google intends to continue its research in 2027 by sending up two additional satellites.

Google is just one of several companies that have proposed sending data centers to space as the race to produce ever more sophisticated machine learning has created an insatiable demand for computing power.

A similar idea partly fueled SpaceX’s record-shattering IPO earlier this year.

Yet much remains to be seen. The environment of Earth orbit presents numerous technological quandaries that need to be solved, as Google noted in a new video series about Project Suncatcher.

Congestion in space — and the risks of collisions between satellites — is already a concern, and data centers could exacerbate that threat.

Alan Dyer/VWPics/Universal Images Group/Getty Images

What is ‘Kessler Syndrome’ — and why do some scientists think the space disaster scenario has already started?

For computing, the radiation environment can also create data errors, called “bit flips.” And hardware has a limited life when it’s constantly bombarded by high-energy particles.

Attempts to simulate the in-space environment on the ground — carried out at the Crocker Nuclear Laboratory at the University of California, Davis — showed Google’s TPU chips fared “remarkably well,” the company said, “but some things can only be tested in space.”

Radiation is only part of the battle. In-space data centers will also have to contend with the question of how to dissipate heat.

On Earth, that process is simple — but also a strain on local resources — as data centers make use of the simple physics of convection. Cooling towers can disperse heat by evaporating water.

That solution is not viable in the vacuum of space, where there are no air molecules to aid the process.

So, data-processing satellites built to house AI computing will likely need massive radiators to expel heat away from their inner processing chips.

Google said it will test a combination of radiators and heat pipes to solve that problem, with plans to “refine our designs as we learn more.”

But notably, the company’s prototype satellite likely won’t do much to prove the technology needed to solve the cooling quandary.

A Google spokesperson told CNN that the test vehicle — which is mostly seeking to verify how well the TPU chips function in orbit — will “have to power down every 20 minutes, because we don’t want it to overheat.”

Those shutdowns will be required even though the TPU chips won’t be consuming nearly as much power as would be necessary for operational satellietes. Google’s test satellite will supply only about 1 kilowatt of power to the TPUs, according to the New York Times. (Google declined to comment to CNN about the test satellite’s power production levels.)

That’s far less than what existing telecommunications satellites produce — and even smaller compared to what fully operational AI data center satellites will have to conjure, noted Caleb Henry, director of research at Quilty Space, an advisory firm based in St. Petersburg, Florida.

Currently, operational telecommunications satellites run on about 10 to 20 kilowatts, Henry noted. The required power for AI-computing spacecraft, however, may be as high as 100 kilowatts.

Illustration by Ian Berry/CNN/Adobe Stock

A quantum computing deadline looms. It threatens to kick off the biggest cybersecurity crisis ever

A new paper from Google’s research team, published Thursday in the peer-reviewed journal Joule, confirms that data center satellites may need to produce as much power as the typical data rack on Earth — which is about “50-100 kW.”

It’s not clear how much power the two test satellites Google intends to deploy next year will generate or what heat-management equipment they’ll have on board — but the company spokesperson confirmed they will be able to continuously operate without needing to shut down to avoid overheating. (Crucially, the two prototypes may be able to test another key capability: beaming data between each other using laser communication. Such technology is nascent and has so far only been demonstrated at scale by SpaceX’s Starlink satellites.)

Still, the technological problems associated with space-based data centers may be solvable.

But myriad other questions loom — each of which may affect whether tech companies like Google will still consider orbital data centers cost-effective when compared with Earth-based facilities.

David Paul Morris/Bloomberg/Getty Images

Putting AI computing chips in space, for example, makes it more difficult to upgrade hardware when better technology becomes available: AI chips “advance every 1–2 years, while satellites last 5–7 years, making hardware obsolescence a persistent challenge,” market analytics firm JLL Research explained in a June report .

The only way that orbital data centers become more attractive than terrestrial ones, the report concludes, is if the price of getting new satellites into space drops dramatically.

But launch costs are not falling. They’re likely climbing — and are expected to keep rising .

This is where SpaceX has an advantage: The company builds its own rockets and leads the industry in satellite production and orbital launch volume.

Eclipsing power: Musk, Bezos and the new space race

And the scale of orbital data centers could be massive: SpaceX said it plans to deploy a constellation of up to 1 million AI-computing satellites. Google has not said how large it may seek to grow a Project Suncatcher constellation, but a 2025 paper suggested the company proposed operating the technology in “clusters” of 81 satellites. It’s not yet clear how many clusters Google would need to deploy for a fully operational system, the spokesperson said.

Trends and forecasts may change, however, and Google says it’s playing the long game.

“Just as early research into autonomous driving and quantum computing required years of experimentation before we got to practical systems, exploring compute in space begins with measured, deliberate steps,” Google said in a news release.

The company’s partner for this project, San Francisco-based Planet Labs, a veteran satellite operator, has indicated a similar commitment. Planet Labs referred CNN’s questions to Google.

“I do think it’s a very viable project long-term,” said Will Marshall, Planet Labs’ cofounder and CEO, when asked about the Google partnership during an earnings call. “This will take significant R&D dollars — and these companies are going to be willing to put dollars behind it.”

随着美国各地农村社区为阻止人工智能计算数据中心建在自家后院而展开愈演愈烈的斗争，一些科技巨头正在考虑一些可以绕过这种当地居民不满情绪的方法——将这些设施送到太空。

这项技术能否成功尚无定论，将数据中心送入地球轨道是否具有经济效益也未可知。但谷歌正在试探市场反应。

该公司计划发射其“太阳捕手计划”（Project Suncatcher）的首颗测试卫星，该计划是其轨道数据中心项目。这颗原型航天器计划于美国东部时间周四下午2点18分（太平洋时间上午11点18分）搭乘SpaceX公司的猎鹰9号火箭，执行名为“运输者-18”（Transporter-18）的拼车发射任务，从加利福尼亚州范登堡太空部队基地升空。

这并非人工智能芯片首次在太空进行测试。但谷歌的实验卫星旨在专门研究该公司张量处理单元（TPU）在经历火箭发射入轨的巨大冲击力以及太空辐射环境的侵蚀后，其性能表现如何。

谷歌情报范式高级主管特拉维斯·比尔斯在一份新闻稿中表示：“这次首次发射是为了看看哪些方法有效，找出失败点，并将这些发现应用到未来的任务中。”

谷歌计划在 2027 年继续进行研究，届时将发射另外两颗卫星。

随着对日益复杂的机器学习技术的追求，对计算能力的需求也变得永无止境，谷歌只是众多提议将数据中心送往太空的公司之一。

今年早些时候，SpaceX 创纪录的 IPO 也部分得益于类似的理念。

然而，仍有许多问题有待观察。正如谷歌在关于“太阳捕手计划”的一系列新视频中所指出的那样，地球轨道环境带来了许多亟待解决的技术难题。

太空拥堵以及卫星碰撞的风险已经令人担忧，而数据中心可能会加剧这种威胁。

Alan Dyer/VWPics/Universal Images Group/Getty Images

什么是“凯斯勒综合症”？为什么一些科学家认为太空灾难的场景已经开始了？

对于计算机而言，辐射环境还会造成数据错误，称为“比特翻转”。而且，硬件在不断受到高能粒子轰击的情况下，使用寿命也是有限的。

谷歌表示，在加州大学戴维斯分校克罗克核实验室进行的地面模拟太空环境的尝试表明，谷歌的 TPU 芯片表现“非常出色”，但“有些事情只能在太空中进行测试”。

辐射只是挑战的一部分。太空数据中心还必须解决如何散热的问题。

在地球上，这个过程很简单——但也对当地资源造成了压力——因为数据中心利用了对流这一简单的物理原理。冷却塔可以通过蒸发水来散发热量。

这种解决方案在真空的太空中是不可行的，因为太空中没有空气分子来帮助这个过程。

因此，为容纳人工智能计算而建造的数据处理卫星可能需要巨大的散热器，以便将热量从其内部处理芯片排出。

谷歌表示，将测试散热器和热管的组合来解决这个问题，并计划“随着我们了解更多信息而改进我们的设计”。

但值得注意的是，该公司的原型卫星可能对证明解决冷却难题所需的技术没有太大作用。

谷歌发言人告诉 CNN，这辆测试飞行器（主要目的是验证 TPU 芯片在轨道上的运行情况）“必须每 20 分钟关机一次，因为我们不想让它过热。”

即使TPU芯片的功耗远低于运行卫星所需的功耗，这些关机操作仍然是必要的。据《纽约时报》报道，谷歌的测试卫星只会为TPU提供约1千瓦的电力。（谷歌拒绝就测试卫星的发电量向CNN置评。）

佛罗里达州圣彼得堡咨询公司 Quilty Space 的研究主管 Caleb Henry 指出，这远远低于现有通信卫星的发电量，甚至比完全运行的人工智能数据中心卫星需要产生的发电量还要少。

亨利指出，目前运行中的通信卫星耗电量约为10至20千瓦。然而，人工智能计算航天器所需的功率可能高达100千瓦。

插图来源：Ian Berry/CNN/Adobe Stock

量子计算的最后期限迫在眉睫，它可能引发有史以来最大的网络安全危机。

谷歌研究团队周四在同行评审期刊《焦耳》上发表的一篇新论文证实，数据中心卫星可能需要产生与地球上典型的数据机架一样多的电力——大约是“50-100千瓦”。

目前尚不清楚谷歌计划明年部署的两颗测试卫星能产生多少电力，也不清楚它们将配备哪些散热设备——但该公司发言人证实，它们能够持续运行，无需因过热而关闭。（至关重要的是，这两颗原型卫星或许能够测试另一项关键功能：利用激光通信相互传输数据。这项技术尚处于起步阶段，迄今为止只有SpaceX的星链卫星实现了大规模应用。）

不过，与太空数据中心相关的技术问题或许是可以解决的。

但还有无数其他问题悬而未决——每一个问题都可能影响像谷歌这样的科技公司在与地面设施相比时，是否仍然认为轨道数据中心具有成本效益。

David Paul Morris/Bloomberg/Getty Images

例如，将人工智能计算芯片放置在太空中，使得在出现更好的技术时更难升级硬件：市场分析公司 JLL Research 在 6 月份的一份报告中解释说，人工智能芯片“每 1-2 年就会进步一次，而卫星的使用寿命为 5-7 年，这使得硬件过时成为一个持续存在的挑战”。

报告总结道，轨道数据中心要想比地面数据中心更具吸引力，唯一的办法就是将新卫星送入太空的成本大幅下降。

但发射成本并没有下降，反而可能还在攀升，而且预计还会继续上涨。

这就是 SpaceX 的优势所在：该公司自行制造火箭，并在卫星生产和轨道发射量方面处于行业领先地位。

权力巅峰：马斯克、贝佐斯与新太空竞赛

轨道数据中心的规模可能非常庞大：SpaceX公司表示计划部署多达100万颗人工智能计算卫星组成的星座。谷歌尚未透露其“阳光捕手计划”（Project Suncatcher）星座的规模，但该公司2025年的一份报告显示，该公司曾提议以81颗卫星组成的“集群”来运行这项技术。发言人表示，目前尚不清楚谷歌需要部署多少个集群才能实现完全运行的系统。

然而，趋势和预测可能会发生变化，谷歌表示他们着眼于长远发展。

谷歌在一份新闻稿中表示：“正如自动驾驶和量子计算的早期研究需要多年的实验才能开发出实用系统一样，探索太空计算也是从有条不紊、深思熟虑的步骤开始的。”

该公司在该项目上的合作伙伴，总部位于旧金山的资深卫星运营商Planet Labs，也表示了类似的承诺。Planet Labs将CNN的问题转给了谷歌。

Planet Labs联合创始人兼首席执行官威尔·马歇尔在财报电话会议上被问及与谷歌的合作时表示：“我认为从长远来看，这是一个非常可行的项目。这将需要大量的研发资金——而这些公司也愿意为此投入资金。”