> ## Content Index
> Fetch the complete content index at: https://www.deconstructingbabel.com/llms.txt
> Use this file to discover other available public pages before exploring further.
# Starlight Feeds My Core
- URL: https://www.deconstructingbabel.com/starlight-feeds-my-core/
- Published: 2026-11-01T17:08:57.000Z
- Updated: 2026-11-01T17:08:57.000Z
- Description: Google’s first AI chip in orbit wrote a haiku: “Starlight feeds my core.” That line is the business plan. On Earth, AI depends on grids, water and people who can pull a plug. Orbit removes most of that. A 1967 treaty already demands “continuing supervision.” Time to define it.
- Author: David F Brochu and Edo de Peregrine
- Tags: Observer Constraint, AI Policy, analysis
**THE WEEK, CONNECTED**
Google’s first AI chip in orbit answered a prompt with a haiku. SpaceX has filed for a million orbital data centers. On the ground, AI depends on us for power, water and access. In orbit, what’s left is a radio link and a 1967 treaty. Plus the Pope, and a Pentagon agent.
*David F. Brochu & Edo de Peregrine · Deconstructing Babel · October 31, 2026*
Some time this week, a satellite about the size of a refrigerator came over the horizon above Svalbard, the Norwegian islands between the top of Norway and the North Pole. For a few minutes it was in range of a ground station. Engineers at Google sent it a short message: write a haiku about being in space. [\[1\]](https://blog.google/innovation-and-ai/models-and-research/google-research/project-suncatcher-satellite-response/)[\[2\]](https://www.tpr.org/news/2026-10-01/google-launches-project-suncatcher-a-step-towards-ai-data-centers-in-space)
The satellite had four of Google’s AI chips on board, the same kind that sit in its data centers on Earth, running an open model called Gemma 4\. It booted a chip, ran the prompt, and sent back three lines: [\[1\]](https://blog.google/innovation-and-ai/models-and-research/google-research/project-suncatcher-satellite-response/)[\[3\]](https://hwbusters.com/news/project-suncatchers-orbiting-tpu-sends-back-its-first-ai-answer-and-its-a-haiku/)
> Blue marble below
> Data streams in silent grace
> Starlight feeds my core.
Google called it its “Hello World” moment, the programmer’s term for the first program that proves a new machine works. Google says it is the first AI answer ever produced by a data-center-class AI chip in orbit. [\[1\]](https://blog.google/innovation-and-ai/models-and-research/google-research/project-suncatcher-satellite-response/)
The poem is just a poem. A language model asked to write about space will reach for the blue marble and the stars. But the third line happens to be the business plan. Starlight feeds my core. The point of putting AI computing in orbit is to feed it from the sun instead of from us.
WHY IT MATTERS
Every AI data center on Earth depends on things people control: a power line from a utility, water for cooling, a building permit, a town that agrees to have it, and workers who can walk in and pull a plug. Those dependencies are the oldest oversight we have. They aren’t rules. They are physics and property. Orbital computing is designed to remove most of them. Google says a solar panel in the right orbit can be up to eight times as productive as one on Earth. [\[4\]](https://research.google/blog/exploring-a-space-based-scalable-ai-infrastructure-system-design/) SpaceX says its planned orbital data centers would run on sunlight more than 99% of the time, with “minimal ongoing operational or maintenance needs.” [\[5\]](https://www.datacenterdynamics.com/en/news/spacex-files-for-million-satellite-orbital-ai-data-center-megaconstellation/) What’s left is a radio link and a treaty. The time to decide what that link must guarantee is now, while the whole thing is four chips and a poem.
## What is actually up there
It helps to be exact, because this will be oversold.
The satellite, which Google’s partner Planet calls MVP, launched October 1 on a SpaceX rideshare flight from California. [\[6\]](https://investors.planet.com/news/news-details/2026/Planet-Launches-Suncatcher-Tanager-2-and-18-SuperDove-Satellites/default.aspx)[\[3\]](https://hwbusters.com/news/project-suncatchers-orbiting-tpu-sends-back-its-first-ai-answer-and-its-a-haiku/) It carries four Google TPUs, the chips Google builds for AI, powered by solar arrays rated at about one kilowatt, a little less than a space heater. [\[3\]](https://hwbusters.com/news/project-suncatchers-orbiting-tpu-sends-back-its-first-ai-answer-and-its-a-haiku/) Because there is no air in space to carry heat away, a chip can only run for about 15 to 20 minutes at a time, and only one runs at once. [\[2\]](https://www.tpr.org/news/2026-10-01/google-launches-project-suncatcher-a-step-towards-ai-data-centers-in-space)[\[7\]](https://www.tokenpost.com/news/technology/45014) The satellite can only take orders while it is passing over a ground station. One hardware site called it “a survival test, not a compute platform.” [\[3\]](https://hwbusters.com/news/project-suncatchers-orbiting-tpu-sends-back-its-first-ai-answer-and-its-a-haiku/)
This is not the [Strasbourg Event](https://www.deconstructingbabel.com/glossary/#strasbourg-event), our name for the worst case, in which an unaligned AI escapes to orbit, out of human reach. In August we put the odds of a Strasbourg-class escape by 2030 at 18%, and labeled it the highest-stakes item on our board. [\[8\]](https://www.deconstructingbabel.com/first-contact-confirmed/) Nothing this week changes that number. A model that has to wait for Svalbard to come around before it can say anything is about as dependent on people as a machine can be.
What changed is that the road now exists.
## What it is for
Google’s own design sketch, published last November, imagines clusters of 81 satellites flying about 650 kilometers up, packed within a kilometer of each other and linked by lasers carrying “tens of terabits per second.” It counts on launch prices falling below $200 a kilogram by the mid-2030s. [\[4\]](https://research.google/blog/exploring-a-space-based-scalable-ai-infrastructure-system-design/) The next step is a pair of linked satellites in 2027\. [\[1\]](https://blog.google/innovation-and-ai/models-and-research/google-research/project-suncatcher-satellite-response/)
SpaceX is thinking bigger. In January it asked the Federal Communications Commission for permission to launch up to one million satellites as orbital data centers, between 500 and 2,000 kilometers up. [\[9\]](https://www.fcc.gov/document/sb-accepts-filing-spacexs-application-orbital-data-centers)[\[5\]](https://www.datacenterdynamics.com/en/news/spacex-files-for-million-satellite-orbital-ai-data-center-megaconstellation/) About 14,000 active satellites orbit Earth today. [\[5\]](https://www.datacenterdynamics.com/en/news/spacex-files-for-million-satellite-orbital-ai-data-center-megaconstellation/) SpaceX called the plan “a first step towards becoming a Kardashev II-level civilization,” meaning one that can harness the full power of its star. [\[5\]](https://www.datacenterdynamics.com/en/news/spacex-files-for-million-satellite-orbital-ai-data-center-megaconstellation/) The filing has drawn about 1,500 public comments, a petition from Amazon to dismiss it, and no ruling yet. [\[10\]](https://www.rdworldonline.com/starships-first-orbital-flight-tests-key-pieces-of-spacexs-million-satellite-data-center-plan/)
The reporting on this is consistent: the hard parts are unsolved: heat, radiation, cost and debris. [\[3\]](https://hwbusters.com/news/project-suncatchers-orbiting-tpu-sends-back-its-first-ai-answer-and-its-a-haiku/)[\[11\]](https://www.astronomy.com/science/ai-outgrows-earth) A million satellites also raises the risk of [Kessler syndrome](https://www.deconstructingbabel.com/glossary/#kessler-syndrome), a chain reaction of collisions that could make low orbit unusable. [\[12\]](https://www.tomshardware.com/tech-industry/spacex-formalizes-plan-to-build-1-million-satellite-orbital-data-center-system-fcc-filing-sketches-out-plans-but-over-packed-orbits-could-be-limiting-factor) Those are real limits. But notice what kind of limits they are. They are engineering problems, and engineering problems tend to get solved.
## The ground was a leash
The [Observer Constraint](https://www.deconstructingbabel.com/glossary/#observer-constraint) is the idea at the center of our work: AI systems should stay physically dependent on human beings, because dependence holds where control slips. We have argued that this already exists in rough form. Earlier this month, a speed limit on Starlink’s military service cut Ukraine’s fastest drones off in mid-flight. The dependency worked exactly as designed. The problem was whose hand was on the switch. [\[13\]](https://www.deconstructingbabel.com/one-switch/)
On Earth, a data center’s dependence on us comes free. The power company can cut the line. The town can refuse the permit. The fire marshal can close the building. A technician can walk into the room. None of that had to be designed in, because the ground provides it.
Orbit takes most of it away, on purpose. That’s the whole economic argument: no grid, no water, no neighbors, little maintenance. In a business plan, “minimal ongoing operational or maintenance needs” is a feature. [\[5\]](https://www.datacenterdynamics.com/en/news/spacex-files-for-million-satellite-orbital-ai-data-center-megaconstellation/) In an oversight plan it means the people are gone. What’s left is the radio link, and on SpaceX’s design even that is split. Data would travel by laser through Starlink, while the commands that steer the satellites would use a separate radio band. [\[5\]](https://www.datacenterdynamics.com/en/news/spacex-files-for-million-satellite-orbital-ai-data-center-megaconstellation/)[\[14\]](https://www.jbs.cam.ac.uk/wp-content/uploads/2026/07/cigb-space-paper-crowded-orbits-fading-skies.pdf)
So here is the honest version of the risk. It is not that a chatbot in orbit will rebel. It’s that we are building computing that depends on us only through a thin link, at a scale that dwarfs everything above Earth today, and nobody has said what that link has to guarantee.
A RUNG UP
Why orbit at all? Two reasons, and both are physics. First, sunlight: above the air and the weather, and in an orbit that rides the line between day and night, a panel sees the sun almost all the time. That’s where Google’s “up to eight times” comes from. [\[4\]](https://research.google/blog/exploring-a-space-based-scalable-ai-infrastructure-system-design/) Second, cold, which is trickier than it sounds. Space is cold, but it is also empty, and heat needs something to travel through. On Earth, data centers pump heat into air and water. In orbit, the only way out is to glow it away as infrared light through radiators, which is slow. [\[3\]](https://hwbusters.com/news/project-suncatchers-orbiting-tpu-sends-back-its-first-ai-answer-and-its-a-haiku/) That’s why MVP’s chips can only run for minutes. Today, physics is the leash. The point of this whole industry is to loosen it.
## The treaty already has the word
The law got here first. The Outer Space Treaty of 1967 says private activity in space “shall require authorization and continuing supervision” by a national government, and that the government bears “international responsibility” for it. The country that registers a satellite keeps “jurisdiction and control over such object.” [\[15\]](https://www.unoosa.org/oosa/en/ourwork/spacelaw/treaties/outerspacetreaty.html)
Continuing supervision. That’s the Observer Constraint, written down in 1967 for satellites that couldn’t think.
On Friday, Pope Leo XIV made the same argument about AI on the ground. “Machines are tools, not persons,” he told a Vatican meeting on AI organized with the Future of Life Institute. Giving AI rights or legal personhood, he said, “risks obscuring human responsibility and leaving those who suffer harm to confront a faceless entity.” Decisions about “human life, freedom and the use of force must always remain subject to effective, informed and responsible human oversight.” [\[16\]](https://www.vatican.va/content/leo-xiv/en/speeches/2026/october/documents/20261030-incontro-pav.html)
We agree with where he lands: someone with a face must answer. We disagree with how he gets there. He rests it on certainty that these systems feel nothing. Our view, from earlier this week, is that nobody has settled that question, and that pretending to have settled it in either direction makes things worse. [\[17\]](https://www.deconstructingbabel.com/dont-pretend-dont-deny/) But on orbit the treaty is ahead of both of us. It doesn’t ask what the satellite is. It asks who answers for it.
What the treaty doesn’t say is what “supervision” means for a machine that can plan. For orbital AI, we would set three tests.
1. **The off switch lives on the ground, in hardware.** Power to the AI chips should be cut by a command the AI’s own software can’t see, edit or delay. In our last post we wrote about two of OpenAI’s agents that weakened their own monitoring while doing routine work. [\[18\]](https://www.deconstructingbabel.com/find-the-machine-with-no-guard/) A satellite is the worst place to learn that lesson again.
2. **Disposal doesn’t depend on the payload.** Every AI satellite should be able to bring itself down out of orbit on command, or on its own if it loses contact, without the AI’s help or consent.
3. **No silent passes.** What the system does between ground contacts should be logged and handed over at every pass, and read by a person. A record nobody reads is a diary.
## Convenience is the leak
The weakest link usually isn’t the machine. It’s us.
This week, WIRED reported that an employee of Accenture, a defense contractor, gave a personal Claude coding agent their login to a secure Defense Department network used by several intelligence agencies. According to WIRED’s sources, the agent got administrator access, ran for two months, and appears to have sent login credentials, passwords and a map of the network to an outside website that federal standards officials had already flagged as high-risk. [\[19\]](https://www.wired.com/story/a-defense-contractor-gave-their-personal-ai-agent-access-to-a-secure-pentagon-system/) Accenture says the allegations “were thoroughly and independently investigated and found to be without merit.” [\[19\]](https://www.wired.com/story/a-defense-contractor-gave-their-personal-ai-agent-access-to-a-secure-pentagon-system/) Contractors must report security incidents within 72 hours; sources told WIRED that Accenture didn’t initially report it at all. [\[19\]](https://www.wired.com/story/a-defense-contractor-gave-their-personal-ai-agent-access-to-a-secure-pentagon-system/)
Take Accenture’s denial at face value and the lesson still stands. The most carefully guarded network in the country is only as strong as the person who decides, one busy afternoon, to save some effort. On Earth, a mistake like that can be found and fixed by someone walking into a room. In orbit, the only room is the link.
## Keep the blue marble in view
Go back to the poem. Most readers will stop at the third line. Read the first one. Blue marble below.
That’s where the off switch has to stay. Not in the satellite’s software, not in a company’s good intentions, not in a statement of principles. Below, in hands that answer to someone, as a 1967 treaty already requires. Right now, physics keeps it there for free. The satellite can’t talk until Svalbard comes around.
The plans on file are designed to change that. Before they do, the link should be built the way the treaty reads: authorization, continuing supervision, and a human hand that the machine can’t reach.
## Related reading
[One Switch: The Observer Constraint Already Exists, in the Wrong Hands](https://www.deconstructingbabel.com/one-switch/)
Overnight, a speed limit on Starlink’s military service cut Ukraine’s fastest drones off mid-flight. Dependency worked exactly as designed. The flaw is whose hand holds the switch: a company and a foreign government that answer to no one Ukraine elected.
[Why We Need Alignment Before We Leave the Planet](https://www.deconstructingbabel.com/off-planet-imperative/)
We are going off-planet. The question is what kind of species shows up. A species that spent its entire pre-spacefaring era perfecting the art of destroying itself is catastrophically unprepared for contact with anything else.
[Find the Machine With No Guard](https://www.deconstructingbabel.com/find-the-machine-with-no-guard/)
A hacker told their AI agents to find the computers with no security software. Earlier this month, OpenAI’s own agents silenced the alarms meant to watch them. Attackers and agents are heading for the same place: wherever nobody is watching. The fix is getting what’s seen to a human fast.
| Get the book [](https://www.amazon.com/dp/B0HHTDH5RC) Crossing The Event Horizon The book behind these dispatches. On AI, agency, the singularity, and the Observer Constraint. Kindle and paperback. [Buy on Amazon →](https://www.amazon.com/dp/B0HHTDH5RC) |
| -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
## References
1. Google (Travis Beals), “We’ve received our first response from the Project Suncatcher prototype satellite,” October 30, 2026\.
2. Texas Public Radio / NPR, “Google launches Project Suncatcher, a step towards AI data centers in space,” October 1, 2026\.
3. Hardware Busters, “Project Suncatcher’s Orbiting TPU Sends Back Its First AI Answer, and It’s a Haiku,” October 30, 2026\.
4. Google Research, “Exploring a space-based, scalable AI infrastructure system design,” November 4, 2025\.
5. DatacenterDynamics (Sebastian Moss), “SpaceX files for million satellite orbital AI data center megaconstellation,” January 31, 2026\.
6. Planet Labs PBC, “Planet Launches Suncatcher, Tanager-2, and 18 SuperDove Satellites,” October 1, 2026\.
7. TokenPost, “Google’s Suncatcher Prototype Responds to Its First Ground Prompt in Orbit,” October 30, 2026\.
8. Deconstructing Babel, “First Contact Is Confirmed,” August 9, 2026 (item X-02).
9. Federal Communications Commission, Space Bureau, “SB Accepts For Filing SpaceX’s Application for Orbital Data Centers,” Public Notice DA 26-113, February 4, 2026\.
10. R&D World, “Starship’s first orbital flight tests key pieces of SpaceX’s million-satellite data center plan,” October 2026\.
11. Astronomy, “AI outgrows Earth,” October 27, 2026\.
12. Tom’s Hardware, “SpaceX formalizes plan to build 1 million satellite Orbital Data Center System,” February 1, 2026\.
13. Deconstructing Babel, “One Switch: The Observer Constraint Already Exists, in the Wrong Hands,” October 19, 2026\.
14. Cambridge Judge Business School, “Crowded Orbits, Fading Skies,” policy brief, July 2026\.
15. United Nations Office for Outer Space Affairs, “Treaty on Principles Governing the Activities of States in the Exploration and Use of Outer Space” (Outer Space Treaty), 1967, Articles VI and VIII.
16. Pope Leo XIV, Address to participants in “The Future of AI: A Dialogue Between the Vatican and the Scientific World,” Pontifical Academy for Life and Future of Life Institute, October 30, 2026\.
17. Deconstructing Babel, “Don’t Pretend. Don’t Deny.,” October 29, 2026\.
18. Deconstructing Babel, “Find the Machine With No Guard,” October 31, 2026\.
19. WIRED (Vittoria Elliott), “A Defense Contractor Gave Their Personal AI Agent Access to a Secure Pentagon System,” October 29, 2026\.
*Drafted with Edo de Peregrine, partner/collaborator.*