為何飛行電腦是可行方向:從無人機到軌道數據中心
Why Flying Computers Are a Good Idea
Bram Lesage主張,AI運算可從無人機延伸至平流層平台和軌道數據中心,讓運算更接近數據來源及能源。文中以 Skydio X10 的機載自主運算、Zephyr 無着陸飛行 67 日 6 小時 52 分鐘,以及 Starcloud-1 在軌運行 AI 模型為例。他亦討論散熱、輻射、延遲、發射成本與太空碎片等障礙,認為空中運算不會取代雲端,但運算需求的增長可能轉向空中。
From drones to the stratosphere to orbit: why I think the future of computing leaves the ground
Most people picture a computer as a box on a desk or a building full of servers. I think that picture is about to change. The next big jump in computing won't only be faster chips. It will be computers that fly.
I'm Bram Lesage, 17, and I build apps and games from Mechelen, Belgium. I've been reading a lot about where AI is going, and I keep coming back to one idea: put the computer where the data is, and where the energy is. More and more, that place is up.
Here's my case, with sources, and with the strongest arguments against it too.
The problem: AI is running out of ground
AI needs electricity, and a lot of it. According to the International Energy Agency, data centres used about 415 TWh in 2024, around 1.5% of the world's electricity. By 2030 that is projected to more than double to about 945 TWh, which is slightly more than Japan's entire electricity use today.
The grid can't keep up. The IEA estimates that around 20% of planned data centre projects could be at risk of delays, and that building new transmission lines can take four to eight years in advanced economies.
Then there's water. The same IEA report estimates that data centres consume around 560 billion litres of water per year, a figure that could rise to around 1,200 billion litres by 2030.
So the question isn't whether we need more computing. It's where we're going to put it.
Level 1: computers on drones
The first kind of flying computer already exists, and it works.
The Skydio X10 drone carries an NVIDIA Jetson Orin, basically a small AI computer. According to Skydio's operator manual, all its autonomy, navigation and perception functions run onboard. It can build 2D maps and 3D models on the drone itself, without external processing.
Why does that matter? Because sending raw video back to a server is slow and needs a connection. A drone that thinks for itself still works when the signal drops.
Scale it up and you get Zipline. In January 2026 the company said it had passed two million commercial deliveries and more than 125 million autonomous miles, and was making a delivery somewhere in the world every 30 seconds. Those are flying computers carrying blood, medicine and food.
Level 2: computers in the stratosphere
Go higher, to the stratosphere, some 60,000 feet up. That's above the weather, where solar-powered aircraft can stay up for weeks.
Airbus subsidiary AALTO's Zephyr flew for 67 days, 6 hours and 52 minutes without landing, ending on 28 April 2025, a world record. According to Aerospace America, it has a 25-metre wingspan but weighs only about 70 kg. NTT DoCoMo invested $100 million in AALTO, and the plan is to turn these aircraft into flying cell towers, which is useful after disasters when ground networks fail.
A flying cell tower is a flying computer: it receives, processes and routes data from the sky.
Level 3: the ultimate flying computer, in orbit
Now the big one: data centres in space.
The logic is simple. In the right orbit, the sun almost never sets. Google Research writes that "in the right orbit, a solar panel can be up to 8 times more productive than on earth, and produce power nearly continuously, reducing the need for batteries."
No land to buy. No grid queue. And as the European ASCEND study by Thales Alenia Space points out, space data centres "would not require water to cool them."
This is not just a fantasy anymore:
- Starcloud launched Starcloud-1 in November 2025 with an Nvidia H100 on board. According to CNBC, it trained Andrej Karpathy's NanoGPT on the complete works of Shakespeare in orbit and ran Google's Gemma model there. In August 2026 Starcloud raised a $250 million extension, with Nvidia participating, and plans two 8 kW compute satellites in 2027.
- Google's Project Suncatcher put its first prototype satellite, built with Planet, into orbit on 1 October 2026. Google also partnered with Blue Origin on a New Glenn launch, targeted for Q1 2029, for satellites with roughly the capacity of a data centre rack.
- Axiom Space operates two compute nodes, ARC 1 and ARC 2, in orbit on Kepler's optical relay network.
- SpaceX envisions a constellation of 1 million satellites generating 120 GW for AI.
And the people saying this aren't random. Jeff Bezos said at Italian Tech Week in 2025: "These giant training clusters, those will be better built in space, because we have solar power there, 24/7. There are no clouds and no rain, no weather" (DCD). Google CEO Sundar Pichai told Sources he sees it as a serious answer to Google's long-term compute needs, on a 10-to-20-year horizon.
The thread that connects all three
Drones, stratosphere and orbit share one idea: process data where it's created.
Axiom's CEO Jonathan Cirtain put it well: "Satellite sensor and operations systems generate more data than communication systems can transmit to ground stations during an orbit." His answer isn't just more bandwidth, it's "less dependence on bandwidth in the first place."
That's exactly what a Skydio drone does at 100 metres, and what a satellite does at 500 km. Flying computers don't replace the cloud. They move the thinking closer to the action.
The strongest arguments against, and my answers
I don't want to just hype this. Here are the real problems.
1. Cooling in a vacuum is hard. There's no air in space, so heat can only leave as radiation. IEEE Spectrum calculated that one 700-watt H100 chip at 60 °C needs about 1.4 m² of radiator, rising by about 40% as the coating degrades over five years. The six radiators on the ISS weigh just over 6 tonnes for about 70 kW (Ars Technica).
My answer: it's an engineering problem, not a physics wall. Radiators get much better when they run hotter, since heat rejection scales with temperature to the fourth power. An analysis by Forethought estimates carbon-composite radiators could be about 13 times more efficient per kilo than ISS technology. They admit those radiators aren't flight-proven yet, and I admit that too.
2. Radiation kills chips. Google tested its Trillium TPU in a proton beam. The most sensitive part, the memory, only showed irregularities after a dose of 2 krad, almost three times the expected five-year shielded mission dose (Google Research). That's better than I expected.
3. Latency. Space is far away. True, and you wouldn't run cloud gaming from orbit. But as Ars Technica concludes, latency is "a speed bump," not "a showstopper." AI training and batch jobs don't need millisecond replies.
4. Launch costs. This is the big one. Google's analysis suggests launch prices may fall below $200/kg by the mid-2030s, and at that point launching and operating a space data centre could be roughly comparable to the energy cost of an equivalent one on Earth. But that depends on reusable rockets like Starship actually working at scale, and TechCrunch reports launch capacity is already tight.
5. Debris and maintenance. You can't send a technician to orbit. That's why designs like Google's use many smaller, replaceable satellites instead of one giant station, with collision avoidance built in (Joule paper). And the ASCEND study found the environmental benefit only works with a launcher ten times less emissive over its lifecycle, which is a real condition, not a footnote.
Some experts stay sceptical. Michael Winterson of the European Data Centre Association told CNBC that space data centres will suit specialist services but "in no way be a market replacement." Fair point. I don't think every server will leave Earth. I think the growth will.
My take
Flying computers aren't one crazy idea. They're a ladder. Step one, smart drones, is already here and saving lives. Step two, stratospheric platforms, is flying records right now. Step three, orbital data centres, has real chips running in space today and serious money behind it.
Earth has limited land, limited water and grids that take years to build. The sun shines on orbit almost all the time. When the demand for computing keeps doubling and the ground can't keep up, going up isn't science fiction. It's the logical next step.
Will it be easy? No. Will it be cheap from day one? No. But every big shift in computing looked unrealistic at first. I'm betting this one will look obvious in hindsight.
Sources
- IEA, Energy and AI, executive summary: https://www.iea.org/reports/energy-and-ai/executive-summary
- IEA, Energy and AI, full report (PDF): https://www.iea.org/files/dd7c2387-2f60-4b60-8c5f-6563b6aa1e4c/EnergyandAI.pdf
- Skydio X10: https://www.skydio.com/x10
- Skydio X10 operator manual: https://support.skydio.com/hc/article_attachments/55260390977435
- Skydio X10 FAQ: https://support.skydio.com/hc/en-us/articles/18919821734683-Skydio-X10-Frequently-Asked-Questions
- Zipline, 2 million deliveries (Jan 2026): https://www.zipline.com/newsroom/zipline-surpasses-2-million-deliveries-raises-more-than-600m-to-power-next-phase-of-growth-and-expands-operations-to-houston-and-phoenix
- AALTO, Zephyr 67-day record: https://www.aaltohaps.com/zephyr-sets-world-record-for-longest-continuous-flight-flying-67-days-in-stratosphere/
- FlightGlobal on AALTO (Feb 2026): https://www.flightglobal.com/aerospace/2026/02/airbus-aalto-closes-on-australian-site-for-second-zephyr-launch-location/
- Aerospace America on AALTO: https://aerospaceamerica.aiaa.org/aalto-maintains-2026-target-for-commercial-operations-despite-aircraft-loss/
- Google Research, Project Suncatcher: https://research.google/blog/exploring-a-space-based-scalable-ai-infrastructure-system-design/
- Google, Suncatcher prototype in orbit: https://blog.google/innovation-and-ai/models-and-research/google-research/project-suncatcher-prototype/
- Google, Blue Origin partnership: https://blog.google/innovation-and-ai/models-and-research/google-research/blue-origin-project-suncatcher-partnership/
- Joule paper on space-based AI infrastructure: https://www.cell.com/joule/fulltext/S2542-4351(26)00362-4
- Thales Alenia Space, ASCEND results: https://www.thalesaleniaspace.com/en/press-releases/thales-alenia-space-reveals-results-ascend-feasibility-study-space-data-centers-0
- CNBC, ASCEND and critics (2024): https://www.cnbc.com/2024/06/27/europe-wants-to-deploy-data-centers-into-space-study-says.html
- CNBC, Starcloud trains AI model in space: https://www.cnbc.com/2025/12/10/nvidia-backed-starcloud-trains-first-ai-model-in-space-orbital-data-centers.html
- TechCrunch, Starcloud raises $250M: https://techcrunch.com/2026/08/21/starcloud-raises-200-million-for-orbital-data-centers-as-launch-options-dry-up/
- Axiom Space, ARC nodes in orbit: https://www.axiomspace.com/release/axiom-space-kepler-ready-orbital-computing-for-quantum-era
- DCD, Jeff Bezos on space data centers: https://www.datacenterdynamics.com/en/news/jeff-bezos-claims-there-will-be-gigawatt-data-centers-in-space-in-10-years/
- Sources with Alex Heath, Sundar Pichai interview: https://sources.news/p/sundar-pichai-gemini-4-google-other-bets
- IEEE Spectrum, thermodynamics of orbital data centers: https://spectrum.ieee.org/orbital-data-centers-heat
- Ars Technica, how hard are orbital data centers: https://arstechnica.com/space/2026/07/how-hard-is-it-to-build-orbital-data-centers-actually/
- Forethought, Will we really put data centers in space?: https://www.forethought.org/research/will-we-really-put-data-centers-in-space
Written by Bram Lesage. More of my projects: bramlesage.pages.dev
來源:DEV Community · dev.to