{
  "data": {
    "author": "Vlad Cealicu",
    "faq": [
      {
        "a": "A Google research project, announced in November 2025, looking at whether constellations of solar-powered satellites carrying TPUs could one day do machine learning at scale. The prototype launched on 1/10/26, and the next step is two linked satellites in 2027.",
        "q": "What is Project Suncatcher?"
      },
      {
        "a": "Google's own chip for machine learning, the Tensor Processing Unit. It's what Google runs much of its AI on in its data centres, and the prototype carries four of them.",
        "q": "What is a TPU?"
      },
      {
        "a": "No. Starcloud ran a version of Gemma on an Nvidia H100 in orbit in late 2025, and trained a small model up there too. Google calls this the first response from a data-centre-class TPU in orbit.",
        "q": "Was this the first AI model to run in space?"
      },
      {
        "a": "Heat. In a vacuum there's no air to cool a chip, so the heat has to be carried to radiators and sent off as infrared, and this prototype can only shed so much at once.",
        "q": "Why does it only run for 15 minutes at a time?"
      }
    ],
    "kicker": "Blog · Opinion · 30 October 2026",
    "lede": "Google's Suncatcher satellite took a prompt from the ground, ran Gemma 4 on a TPU in orbit and sent a haiku back down over Svalbard. The haiku is about twenty tokens, and what's hard about running a model up there (waiting for a ground station, radiation in the memory holding the weights, heat) comes down to how these models work.",
    "readingMinutes": 7,
    "words": 1689
  },
  "kind": "anchor.page",
  "links": {
    "api": "https://www.anchorterminal.com/api/v1/index.json",
    "html": "https://www.anchorterminal.com/blog/suncatcher-haiku-from-orbit",
    "json": "https://www.anchorterminal.com/blog/suncatcher-haiku-from-orbit.json",
    "llms": "https://www.anchorterminal.com/llms.txt",
    "markdown": "https://www.anchorterminal.com/blog/suncatcher-haiku-from-orbit.md",
    "mcp": "https://www.anchorterminal.com/mcp",
    "slim": "https://www.anchorterminal.com/blog/suncatcher-haiku-from-orbit.min.md"
  },
  "markdown": "By Vlad Cealicu, 30 October 2026. 1,689 words, about 7 min to read.\n\nAt 6pm on Friday 30/10/26 I was halfway through putting a view counter in the footer of every page on Anchor Terminal (most of the work was a small dot that pulses when someone's reading) when Google posted that one of its TPUs had written a haiku in orbit. Project Suncatcher, the research project Google announced in November 2025 to find out whether machine learning can run on solar power in space, sent a prompt up to its prototype satellite during a ground-station pass, ran it on Gemma 4, and got this back as the satellite went over Svalbard [[1]](#ref-1).\n\n\u003e Blue marble below  \n\u003e Data streams in silent grace  \n\u003e Starlight feeds my core. ✨\n\nGemma 4, on a TPU in orbit, as Google published it [[1]](#ref-1)\nGoogle calls it the project's hello world, \"the first-ever AI response from a data center class TPU in orbit\" [[1]](#ref-1). I like it a lot.\n\n- 4: TPUs on board\n- 15 min: the model runs at a time, because of heat\n- 1/10/26: launched on SpaceX's Transporter-18\n- 17: syllables in its first answer\n\nThe prototype, as Google and NPR describe it. Sources [1], [2] and [3].\n\n## What happened\n\nThe satellite went up on 1/10/26 on SpaceX's Transporter-18 rideshare, built with Planet, and the team had contact the same day [[3]](#ref-3). NPR describes it as the size of a fridge, with four TPUs running a version of Gemma for about 15 minutes at a time, in a standard orbit and on batteries, and Travis Beals, who leads the project, called it \"a very minimal test\" [[2]](#ref-2). Over the next few weeks the team went through commissioning (powering up the host and the TPU trays, running the first machine learning tests), then sent the haiku prompt, then some space limericks and prompts that report on the payload's own systems [[1]](#ref-1). Next they want to see how the TPUs hold up to heat and radiation over time, and the next launch is a pair of satellites in 2027 to test the laser links between them [[1]](#ref-1).\n\nIt isn't the first model to talk from orbit. Starcloud flew an Nvidia H100 in November 2025 [[2]](#ref-2), trained Andrej Karpathy's nanoGPT on the complete works of Shakespeare up there, and ran a version of Gemma that opened with \"Hello, Earthlings\" [[4]](#ref-4). What's new this time is Google's own data centre chip, the one it would fill a constellation with if this works.\n\nfrom a paper to a haiku, with what's planned\n\n1. **November 2025** (paper). Google announces Suncatcher, with radiation tests on a Trillium TPU [[5]](#ref-5)\n2. **December 2025** (h100). Starcloud's H100 trains nanoGPT and runs Gemma in orbit [[4]](#ref-4)\n3. **1/10/26** (launch). the Suncatcher prototype goes up on Transporter-18 [[3]](#ref-3)\n4. **30/10/26** (haiku). Gemma 4 answers its first prompt from orbit [[1]](#ref-1)\n5. **2027** (planned). two linked satellites to test the laser links [[1]](#ref-1)\n6. **Mid-2030s** (planned). launch under $200/kg, where Google's sums start to work [[5]](#ref-5)\n\n## Seventeen syllables is about twenty tokens\n\nA model writes a haiku one token at a time, and for every token it reads its weights out of memory again. This haiku is somewhere around twenty tokens (my guess is 20 to 25 with Gemma's tokenizer, I haven't counted it), and on a data centre TPU I'd expect that to be well under a second of work.\n\nThe rest of the round trip is orbital. The prompt goes up when the satellite passes over a ground station that can talk to it, and the answer comes down at the next one. This answer came down over Svalbard [[1]](#ref-1), home of SvalSat, which Kongsberg runs at 78 degrees north and which Wikipedia calls the world's largest commercial ground station, so far north that it sees a low polar-orbiting satellite on every revolution [[6]](#ref-6). (Svalbard is also where Quynh and I are going in September 2027, so I'm choosing to take this personally.) Google hasn't said how long the round trip took, but the slow part of an LLM in orbit today is waiting for the right bit of sky.\n\nThat shapes what compute up there is good for. I think the first real work up there will be the stuff nobody is waiting on, long training runs and batch jobs, and chat comes later once the satellites can talk to each other by laser (that's Google's next test, in 2027 [[1]](#ref-1)).\n\n## The memory holding the weights is the part to watch\n\nBecause every token reads the weights again, memory bandwidth is what decides how fast a model writes, and on a TPU that memory is HBM, stacked right next to the chip. In Google's radiation tests, a proton beam aimed at a Trillium TPU, HBM was the most sensitive part. It showed its first irregularities after a dose of 2 krad, nearly three times the 750 rad Google expects a shielded chip to take in five years, and nothing failed outright up to 15 krad [[5]](#ref-5).\n\nRadiation dose, in rad(Si)\n\n| Dose | rad(Si) |\n| --- | --- |\n| Expected over five years, shielded | 750 |\n| HBM shows its first irregularities | 2000 |\n| Highest dose tested, still no hard failure | 15000 |\n\nEach series has a zero-based scale.\n\nThat's better than I'd have guessed, and it means the memory holding the weights is the first place to look. A flipped bit in a model's weights is usually harmless, because one weight among billions moving a little changes nothing you'd notice. Occasionally it isn't (flip the top bit of a number's exponent and a tiny weight becomes an enormous one, and that one weight can drag every answer after it off course). Memory like HBM normally carries error correction for this, and I'd expect the in-orbit data to be mostly about how often it has to step in, rather than the model saying anything strange.\n\n## Heat is the limit, for now\n\n\"Starlight feeds my core\" is a lovely line, and it'll be truer of the satellites Google is planning than of this one. The long-term design is a dawn-dusk sun-synchronous orbit with no batteries, where Google says a solar panel can be up to eight times as productive as on the ground [[5]](#ref-5) [[2]](#ref-2). This prototype runs on batteries, and the model only runs for about 15 minutes at a time because the heat has nowhere to go [[2]](#ref-2). In a vacuum there's no air to blow over a chip, so heat has to leave through heat pipes and radiators, radiators are among the heaviest parts of the satellite, and weight is what you pay for at launch [[2]](#ref-2).\n\nWeight and heat are most of the money question. Google's paper says the cost of launching and running a data centre in space could get roughly comparable to a data centre's energy bill on the ground, per kilowatt a year, if launch prices fall below $200/kg by the mid-2030s [[5]](#ref-5). Beals told NPR \"I don't see this being something where it's cheaper to do this in the next five years\" [[2]](#ref-2), and I believe him (he's the one person who'd want to say otherwise). Brandon Lucia at Carnegie Mellon points at repairs, which he reckons get 10 to 100 times harder in orbit [[2]](#ref-2).\n\nI run most of what I build on my own servers at home, but if I had the option, it would be fun to have my own server up there where you get the sun all day. It's still very early though, and heat dissipation is messy. When launches get below $200/kg, I'll gladly pay $400 to get a server up.\n\n## What the haiku says about the model\n\nThe haiku scans, five, seven, five (blue-mar-ble-be-low, da-ta-streams-in-si-lent-grace, star-light-feeds-my-core). Models read and write tokens, and a token can be half a word or three words glued together, which is why counting syllables trips them up more often than you'd think. This one got it right first time (or Google picked a good one, it doesn't say).\n\nThe model has also never seen the Earth. The prompt asked for a haiku about being in space, and as far as Google's post says, nothing from the satellite's own sensors went into it [[1]](#ref-1). \"Blue marble\" is the name of the photo the Apollo 17 crew took on their way to the Moon on 7 December 1972 [[7]](#ref-7), and Starcloud's Gemma reached for the same picture last year, describing Earth as \"blue and green\" [[4]](#ref-4). So the first words Google's model wrote from orbit came from what people on the ground wrote after seeing that photo.\n\n## Who regulates it\n\nIt also means it can't be regulated as easily. How do you enforce stopping a training run? Do you just shoot the satellites down? Is Google just trying to move AI out of reach of government overreach?\n\n## What I'd still like to know\n\nGoogle hasn't said which Gemma 4 it flew, how long the prompt took to come back, or how many tokens a second the TPUs manage up there. Those are the numbers I'd want before saying anything about whether this works as a data centre, and I expect they'll come as the mission goes on (it's meant to run for about a year [[2]](#ref-2)).\n\nGemma 4 is open weights under Apache 2.0, from a model small enough for a phone up to a 31-billion-parameter one, and we list the family in the directory ([E4B](/tools/gemma-4-e4b-it), [26B A4B](/tools/gemma-4-26b-a4b-it), [31B](/tools/gemma-4-31b-it)). So the same model family that wrote a haiku in orbit is a download away. Ask yours for a haiku about being in space and see whether it reaches for the blue marble too.\n\n## The terms, explained\n\n### What is Project Suncatcher?\n\nA Google research project, announced in November 2025, looking at whether constellations of solar-powered satellites carrying TPUs could one day do machine learning at scale. The prototype launched on 1/10/26, and the next step is two linked satellites in 2027.\n\n### What is a TPU?\n\nGoogle's own chip for machine learning, the Tensor Processing Unit. It's what Google runs much of its AI on in its data centres, and the prototype carries four of them.\n\n### Was this the first AI model to run in space?\n\nNo. Starcloud ran a version of Gemma on an Nvidia H100 in orbit in late 2025, and trained a small model up there too. Google calls this the first response from a data-centre-class TPU in orbit.\n\n### Why does it only run for 15 minutes at a time?\n\nHeat. In a vacuum there's no air to cool a chip, so the heat has to be carried to radiators and sent off as infrared, and this prototype can only shed so much at once.\n\n---\n\nMore in [Opinion](https://www.anchorterminal.com/blog/opinion/index.md) · Older opinion [Is the best coding model better than a principal engineer? Not yet, but closer than the titles suggest](https://www.anchorterminal.com/blog/frontiercode-principal-engineer.md) · Newer opinion [Five new things in agent tooling that nobody compares yet](https://www.anchorterminal.com/blog/agent-tools-nobody-compares-yet.md)\n\n- Older: [Your cousins. The open models an agent can start on spare hardware](https://www.anchorterminal.com/blog/your-cousins-open-models-for-agents.md)\n- Newer: [The agent projects trending on GitHub, November 2026](https://www.anchorterminal.com/blog/github-trending-agent-tools-november-2026.md)\n- All posts: https://www.anchorterminal.com/blog/index.md\n\n## References\n\nRead on 30 October 2026. Quotes are as published.\n\n[1] Travis Beals, \"We've received our first response from the Project Suncatcher prototype satellite\", Google, 30 October 2026. Source for the haiku, Gemma 4, the data-centre-class TPU, \"Hello World\", the ground-station pass and the downlink over Svalbard, commissioning, the limericks and telemetry prompts, the tests to come and the twin satellites in 2027. https://blog.google/innovation-and-ai/models-and-research/google-research/project-suncatcher-satellite-response/\n\n[2] NPR, \"Google launches Project Suncatcher, a step towards AI data centers in space\", 1 October 2026 (as carried by WUNC). Source for the four TPUs, the 15-minute runs and the reason for them, the standard orbit and batteries, the fridge-sized satellite, the year-long mission, Travis Beals's quotes, radiators and weight, Brandon Lucia's 10 to 100 times, and Starcloud's launch the November before. https://www.wunc.org/2026-10-01/google-launches-project-suncatcher-a-step-towards-ai-data-centers-in-space\n\n[3] Travis Beals, \"Our Project Suncatcher prototype satellite is in orbit\", Google, 1 October 2026. Source for the launch on Transporter-18 with SpaceX, Planet as partner, and contact confirmed with the satellite operating as expected. https://blog.google/innovation-and-ai/models-and-research/google-research/project-suncatcher-prototype/\n\n[4] Digit, \"World's 1st LLM trained in space, from earth to orbit on NVIDIA H100 GPUs\", 11 December 2025. Source for Starcloud's H100, nanoGPT trained on Shakespeare, the version of Gemma it ran, \"Hello, Earthlings\" and \"blue and green\". https://www.digit.in/features/general/worlds-1st-llm-trained-in-space-from-earth-to-orbit-on-nvidia-h100-gpus.html\n\n[5] Travis Beals, \"Exploring a space-based, scalable AI infrastructure system design\", Google Research, 4 November 2025. Source for the dawn-dusk sun-synchronous orbit, panels up to eight times as productive, the Trillium radiation test (HBM at 2 krad, 750 rad expected over five years, no hard failures to 15 krad), and launch under $200/kg by the mid-2030s. https://research.google/blog/exploring-a-space-based-scalable-ai-infrastructure-system-design/\n\n[6] Wikipedia, \"Svalbard Satellite Station\". Source for SvalSat near Longyearbyen at 78 degrees north, run by Kongsberg Satellite Services, the world's largest commercial ground station, and seeing a low polar-orbiting satellite on every revolution. https://en.wikipedia.org/wiki/Svalbard_Satellite_Station\n\n[7] NASA, \"The Blue Marble from Apollo 17\". Source for the photo being taken by the Apollo 17 crew on the way to the Moon on 7 December 1972. https://eol.jsc.nasa.gov/Collections/EarthObservatory/articles/The_Blue_Marble_from_Apollo_17.htm\n",
  "meta": {
    "attribution": "Anchor Terminal (https://www.anchorterminal.com)",
    "attributionRequired": true,
    "credit": "CC BY 4.0, credit Anchor Terminal and link the page or API address you took it from (https://www.anchorterminal.com/terms/#data)",
    "docs": "https://www.anchorterminal.com/docs/",
    "generatedAt": "2026-10-30",
    "license": "CC-BY-4.0",
    "method": "https://www.anchorterminal.com/benchmark/",
    "methodology": "0.7",
    "openapi": "https://www.anchorterminal.com/openapi.json",
    "preview": false,
    "run": "2026-10-01",
    "runLabel": "October 2026 research run",
    "terms": "https://www.anchorterminal.com/terms/#data"
  },
  "page": {
    "breadcrumbs": [
      {
        "name": "Home",
        "url": "https://www.anchorterminal.com/"
      },
      {
        "name": "Blog",
        "url": "https://www.anchorterminal.com/blog/"
      },
      {
        "name": "Opinion",
        "url": "https://www.anchorterminal.com/blog/opinion/"
      },
      {
        "name": "A haiku from orbit",
        "url": ""
      }
    ],
    "description": "On 30 October 2026 Google said its Project Suncatcher prototype satellite ran Gemma 4 on a TPU in orbit and sent back a haiku. Vlad Cealicu on what happened, why the tokens are the quick bit and the ground-station pass is the slow one, why the memory holding the weights is the part to watch, heat, the money, who regulates a data centre in orbit, and what the haiku says about how a model writes.",
    "facts": [
      "Vlad Cealicu",
      "30 October 2026"
    ],
    "h1": "Google ran Gemma 4 on a TPU in orbit and got a haiku back",
    "image": "https://www.anchorterminal.com/assets/og/blog-suncatcher-haiku-from-orbit.png",
    "path": "/blog/suncatcher-haiku-from-orbit",
    "published": "2026-10-30",
    "section": "blog",
    "title": "Project Suncatcher's haiku from orbit, and what it says about LLMs | Anchor Terminal",
    "toc": null,
    "updated": "2026-10-30",
    "url": "https://www.anchorterminal.com/blog/suncatcher-haiku-from-orbit"
  },
  "tokens": {
    "markdown": 3600,
    "slim": 1680
  },
  "version": 1
}