AI Data Centers Are Moving to Space: Google's Suncatcher Satellite, Starcloud, and the Race to Compute in Orbit (November 2026)

On October 30, 2026, an artificial intelligence model orbiting hundreds of kilometers above Earth wrote a poem — and beamed it back down to the ground. The poem was a haiku, the model was Google DeepMind's Gemma 4, and the computer running it was a Google Tensor Processing Unit (TPU) bolted to a satellite built by Planet Labs and launched on a SpaceX Falcon 9. It was the first working proof of a wild idea whose time may finally be coming: AI data centers in space. The haiku read:

Blue marble below
Data streams in silent grace
Starlight feeds my core.

It is a small poem with an enormous implication. For the first time, a data-center-class AI chip has run a complete inference workload in orbit and returned the answer to Earth. Google's Project Suncatcher — the company's prototype "space data center" — is online, communicating, and, as a Google spokesperson told TechCrunch, "operating nominally."

And Google is not alone. A startup called Starcloud has already flown an NVIDIA H100 GPU to orbit, trained an AI model in space, and raised $450 million at a $2.3 billion valuation — with NVIDIA itself as an investor. SpaceX and xAI have filed paperwork for up to one million orbital compute satellites. Another startup, Cowboy Space, raised $275 million to turn spent rocket stages into megawatt-scale data centers.

Something strange and serious is happening above our heads: the AI industry is starting to move its computers off the planet. In this complete guide, I explain what Google actually achieved this week, who else is racing to build AI data centers in space, the physics of why orbit might beat Ohio for running AI, the very real reasons it might not work, and what to watch next.

The 60-Second Version: Orbital Data Centers Quick Facts

Fact	What Is Confirmed
What happened	Google's Project Suncatcher prototype satellite — carrying four TPUs — completed commissioning and ran its first AI workloads in orbit (TechCrunch, Oct 30, 2026)
The first task	Engineers asked the onboard Gemma 4 model to write a haiku about being in space; the answer was downlinked over Svalbard, Norway
The satellite	Built by Planet Labs; launched Oct 1, 2026 on a SpaceX Falcon 9 (Transporter-18 rideshare) with 100+ other spacecraft; generates up to 1 kW of power
The catch	Only one of the four TPUs can run at a time, for about 20 minutes per burst — the chip's power demand exceeds what the satellite generates
What's next	Two more Suncatcher satellites planned for 2027 with higher power and laser links between spacecraft; Google targets infrastructure "economical in the 2030s"
The startup race	Starcloud: $450M raised, $2.3B valuation, H100 flown Nov 2025, NVIDIA's Space-1 Vera Rubin Module partnership, FCC filing for 88,000 satellites
The giant	SpaceX + xAI have filed for up to one million orbital compute satellites (project reportedly called Starmind)
Why space	Free solar power, passive radiative cooling, no land or water use — Starcloud projects energy costs of $0.002–$0.005/kWh in orbit
Why it might fail	Radiation damage, thermal stress, launch-cost dependence, impossible maintenance, space debris, regulation, latency

The key takeaway: orbital data centers have crossed from slide decks to working hardware. But every system flying today is a prototype measured in kilowatts, while a single Earth data center draws hundreds of megawatts. The race is real; the finish line is at least a decade away.

What Exactly Happened: Google's Satellite Wakes Up

The story begins not this week, but on October 1, 2026, when a SpaceX Falcon 9 lifted off on the Transporter-18 rideshare mission carrying more than 100 spacecraft. Tucked among them was an unassuming satellite built by Planet Labs for Google's Project Suncatcher — a prototype "space data center" carrying four Google Tensor Processing Units, the same type of AI accelerator chip Google uses in its Earth-bound data centers.

For about three weeks, the satellite went through commissioning — engineers testing each subsystem in turn, making sure the spacecraft survived launch and could talk to the ground. Then, this week, came the first real work. Ground engineers uplinked a prompt during a ground-station pass, asking the pre-loaded Google DeepMind Gemma 4 model to write a haiku about being in space. The satellite thought — running inference on its onboard TPU — and the response came down as the spacecraft passed over Svalbard, Norway.

Google also tested the model with space-themed limericks (TechCrunch dryly noted the engineers "should have sent a poet?") and, more practically, asked the spacecraft to report on the orbital conditions around the satellite — telemetry-style prompts that test whether the system can do useful work, not just party tricks.

Why a haiku matters: it was the first complete AI inference response generated in orbit and returned to Earth from this class of hardware. It proves the full loop works: prompt up, compute in space, answer down. Everything else — the scale, the economics, the reliability — is still to be proven. But the loop is closed, and that is a genuine milestone.

Google's senior director Travis Beals kept expectations grounded, calling the launch "a very minimal test" to confirm the chips can run in space, according to NPR. The mission is meant to last about a year — a long-duration experiment, not a production deployment.

Inside the Satellite: Specs, Power, and the 20-Minute Problem

Here is what is confirmed about the Suncatcher prototype's hardware — and its most important limitation.

The chips: four Google TPUs are onboard. (TechCrunch issued a correction on this point: the satellite has four TPUs but can only operate one at a time.) These are data-center-class accelerators — the Trillium generation (Google's v6e Cloud TPU) was the chip put through radiation testing on the ground. Nothing remotely this powerful has run sustained AI workloads in orbit before.

The power: the satellite generates up to 1 kilowatt of electricity from its solar panels — about as much as a microwave oven. A single TPU's power demand exceeds what the satellite can generate continuously. The result: the TPU can only be switched on for about 20 minutes at a time (TechCrunch's figure; NPR and others report around 15 minutes). Then it has to rest while the batteries recharge.

Why this matters: the power ceiling "complicates the already tricky business of space-to-ground communications," as TechCrunch put it. The satellite must align its brief compute windows with its brief communication windows — talking to the ground only when passing over ground stations. The two more capable satellites Google plans for 2027 will carry higher power generation and inter-satellite optical (laser) links, so spacecraft can relay data to each other instead of waiting for a ground station.

A correction worth noting: early coverage got some details wrong, and NPR later corrected its initial claims — the prototype is not in a sun-synchronous orbit and does not run without batteries, as first reported. It uses a standard orbit and batteries. The full Suncatcher vision does call for dawn–dusk sun-synchronous orbits (orbits that keep the satellite in near-constant sunlight, maximizing solar power), but the prototype is a simpler testbed.

The honest summary: this is a 1-kilowatt science experiment proving that TPUs survive and function in orbit. A single rack in an Earth data center can draw 100+ kilowatts. Suncatcher is to orbital data centers what the Wright Flyer was to aviation — proof that the thing can fly at all, not proof that it can carry passengers.

Radiation and Heat: The Tests That Actually Matter

The haiku got the headlines, but the ongoing test campaign is the real story. Google is using this year-long mission to answer two brutal questions: can AI chips survive space radiation, and can they survive the thermal violence of orbit?

The radiation problem: in orbit, high-energy particles strike chips constantly. The nightmare scenario is the "bit flip" — a particle strike that flips a 0 to a 1 (or vice versa) inside the chip's memory, corrupting data without tripping normal monitoring. For an AI accelerator doing trillions of calculations, silent data corruption is far worse than a clean crash: the model keeps running, confidently producing wrong answers.

Before launch, Google tested its Trillium TPUs in a 67 MeV proton beam on Earth — essentially shooting the chips with a particle accelerator to simulate space radiation. In a preprint paper, Google researchers led by Travis Beals reported that the high-bandwidth memory showed irregularities only after 2 kilorads of radiation exposure — about three times Google's 750 rad minimum requirement. One report described the results as more favorable than anticipated, though other coverage noted some error rates rose. The in-orbit test campaign will now check whether those ground simulations match reality.

The thermal problem: in orbit there is no air, which means no air cooling — no fans, no liquid cooling loops dumping heat into the atmosphere. A chip in space can only shed heat by radiating it away, which is slow and inefficient. Meanwhile the satellite swings between blazing sunlight and deep shadow every ~90 minutes, subjecting every component to relentless thermal cycling — expansion and contraction that cracks solder joints and kills electronics over time. The test campaign is measuring how the TPUs perform under exactly this stress.

Why this is the make-or-break question: if Trillium-class chips degrade too fast in orbit, the entire orbital data center thesis needs radiation-hardened chip designs — slower, more expensive, and years away. If they hold up, the industry can fly essentially commercial silicon. This one satellite's year of data will shape billions of dollars of investment decisions.

Google's Endgame: What Suncatcher Is Supposed to Become

Google is thinking far beyond one poetry-writing satellite. The Suncatcher vision, described in the company's research, looks like this:

• Constellations of satellites flying in compact formations, hundreds of meters apart, in dawn–dusk sun-synchronous orbits that maximize solar exposure.
• Free-space optical transceivers — laser links — connecting the satellites at tens of terabits per second, letting them work as a single distributed computer rather than isolated machines.
• This inter-satellite networking is what makes Suncatcher unusual: Google is "largely alone among visions for space-based data centers in planning on high-speed links between the satellites to support coherent distributed workloads," as TechCrunch noted. Everyone else is essentially planning individual servers in space; Google wants a cluster.
• Two more satellites in 2027, with advanced chips and those laser links, as the next step.
• Google's stated target: infrastructure that is "economical in the 2030s." Not next year. Not 2028. The 2030s.

The honest read: Google is running a decade-long research program and telling you so upfront. That candor is refreshing — and it sets the right expectations. Anyone promising profitable orbital data centers before 2030 is selling something.

The Startup Race: Starcloud's $450 Million Bet

If Google is the careful researcher, Starcloud is the sprinter. The Redmond, Washington startup has become the poster child of the orbital data center rush — and its story is worth understanding in detail, because it shows how fast private capital is moving.

From Y Combinator to unicorn in 17 months

Starcloud was founded in January 2024 (originally named Lumen Orbit) by CEO Philip Johnston, CTO Ezra Feilden (previously of Airbus Defence and Space), and chief engineer Adi Oltean (who worked on SpaceX's Starlink program after two decades at Microsoft). It went through Y Combinator's Summer 2024 batch — and then did something no YC company had done before: it reached unicorn status ($1 billion valuation) just 17 months after demo day, the fastest in the accelerator's history (a program that produced Airbnb, Stripe, and Dropbox).

The funding has compounded since:

• March 2026: $170 million Series A at a $1.1 billion valuation, co-led by Benchmark and EQT Ventures.
• August 21, 2026: $250 million Series A extension at a $2.3 billion post-money valuation, led by Manhattan West — more than doubling the valuation in five months. New investors included NVIDIA itself (about $25 million, per a TechCrunch source), Cisco Investments, Cedar Capital, Goanna Capital, and Standard Capital.
• Total raised: $450 million. The company is building production lines for its Starcloud-3 spacecraft at a new 100,000-square-foot manufacturing facility in Woodinville, Washington.

What Starcloud has actually flown

Unlike most space startups, Starcloud has hardware in orbit doing work:

• November 2025: Starcloud-1 launched carrying an NVIDIA H100 GPU — the first data-center-class GPU ever flown in space, roughly 100 times more powerful than any GPU compute previously flown. On it, the company trained an AI model in space for the first time (a model called NanoGPT), ran a version of Google's Gemini in orbit, and demonstrated orbital fine-tuning.
• Starcloud-2: initially targeted for October 2026 with a Blackwell GPU, an AWS server blade, and even bitcoin mining hardware aboard; later reporting put the second generation in 2027, aimed at performing orbital inference for customers including U.S. government agencies. (Timelines have shifted — treat the 2027 date as the current plan.) Crusoe Cloud has signed up to resell Starcloud's orbital GPU capacity, and AWS is contributing a server blade — early signs of real customer demand.
• Starcloud-3: a 200-kilowatt, 3-tonne spacecraft designed for launch aboard SpaceX's Starship — two hundred times the power of Google's current prototype.

The NVIDIA partnership

The deepest validation of Starcloud's approach came from NVIDIA itself. The two companies are collaborating on the Space-1 Vera Rubin Module — a space-rated variant of NVIDIA's next-generation Vera Rubin platform, hardened for an environment where cooling happens only through radiators and every component must survive radiation and launch loads. NVIDIA says it will deliver 25 times as much in-space compute capability as the H100, and Starcloud's satellites will serve as an early flight platform for the module. (For our full guide to the Rubin platform itself, see our NVIDIA Rubin overview.)

The 88,000-satellite filing

In February 2026, Starcloud filed an application with the Federal Communications Commission to operate as many as 88,000 satellites as orbital data centers. That request has not been approved — and 88,000 satellites would be an unprecedented constellation (Starlink, the largest ever deployed, counts in the thousands). The filing is best read as a statement of ambition and a regulatory placeholder, not a deployment plan. But it shows the scale this industry is thinking at.

The Other Players: Cowboy Space, SpaceX, and the Megawatt Dream

Starcloud and Google are the most visible racers, but the field is widening fast:

• Cowboy Space raised $275 million in May 2026 for a vertically integrated design with a clever twist: launch vehicle upper stages that remain in orbit as one-megawatt data centers — turning what is normally space junk into server racks. One megawatt is a thousand times Google's prototype power.
• SpaceX and xAI — now combined into a single entity reportedly valued around $1.25 trillion — have filed plans for up to one million data center satellites, for a project reportedly called Starmind. A million satellites is a number so large it breaks every existing model of orbital traffic management. Treat it as a regulatory filing expressing intent, not a construction schedule.
• The broader context: analysts project the in-orbit data center market growing from $500 million in 2025 to $39.09 billion by 2035 — a 67.4% compound annual growth rate. (That is one analyst projection, not a fact — but it shows where the smart money's expectations sit.)

The pattern across all of them: everyone is betting that launch costs will keep falling — specifically, that SpaceX's Starship will deliver its promised economics (Starcloud targets roughly $500 per kilogram to orbit). The entire orbital data center thesis is, at bottom, a bet on Starship. If Starship works at scale, the math changes completely. If it does not, everyone is stuck with expensive Falcon-class launch economics and the business case collapses.

Why Put Data Centers in Space at All? The Physics Case

It sounds absurd until you look at the numbers. The case for orbital data centers rests on three physical advantages that Earth cannot match:

1. Free, endless solar power. In dawn–dusk sun-synchronous orbit, a satellite is in near-constant sunlight — no night, no clouds, no weather. Solar panels in space also receive more intense sunlight than on the ground (no atmosphere absorbing it). Starcloud projects orbital energy costs of $0.002 to $0.005 per kWh — a tiny fraction of terrestrial industrial electricity rates. For AI workloads that are essentially "electricity converted into intelligence," energy is the dominant cost, and space has the cheapest energy in the solar system.

2. Free cooling. Data centers on Earth spend enormous amounts of energy and water on cooling — AI chips run extremely hot, and a large data center can consume millions of gallons of water per day for cooling. In space, heat radiates away into the cold of the universe passively, through radiators. No water, no cooling towers, no chillers. (The catch, as Google's prototype shows: radiative cooling is slow, which is why thermal design is one of the hardest engineering problems in this field.)

3. No land, no neighbors, no permits. Terrestrial AI data centers are hitting very Earthly limits: they strain regional electricity grids, consume water in drought-prone areas, and face growing local opposition from communities that do not want a gigawatt facility next door. Building permits, grid interconnection queues, and environmental reviews now take years. In orbit, there is no zoning board. (There is the FCC and international spectrum coordination — but no neighbors.)

The honest version of the pitch: AI's appetite for electricity is growing faster than Earth's grid can comfortably feed it, and space offers unlimited power and cooling with zero local politics. That is a genuinely compelling argument — which is why it deserves genuinely tough scrutiny.

Why It Might Not Work: Seven Hard Problems

Every revolutionary infrastructure idea looks inevitable until you list what can kill it. Here are the seven problems orbital data centers must solve, roughly in order of difficulty:

1. Launch cost is everything. The entire economic case depends on launch getting radically cheaper. Starcloud's energy-cost projections assume Starship flying commercially at around $500/kg. Starship is still unproven at that scale and cadence. If launch costs stall, orbital compute can never beat a data center in Iowa on price. This is the single binding constraint on the whole industry.

2. Radiation degrades everything. Bit flips, memory corruption, shortened component lifespans — space radiation is a tax on every chip, paid continuously. Google's year-long test will quantify it, but radiation-hardened designs are inherently slower and more expensive than commercial silicon. There may be a fundamental trade-off between "fastest chips" and "chips that survive orbit."

3. You cannot send a technician. When a server fails in an Earth data center, a technician swaps the blade in minutes. When a satellite's computer fails in orbit, it is a total loss — a multi-million-dollar brick circling the Earth until it deorbits. Reliability requirements are therefore extreme, and redundancy (flying spare capacity) is expensive.

4. Heat rejection does not scale easily. Radiative cooling works, but radiators are large, heavy, and every kilogram costs launch money. A megawatt-scale orbital data center needs enormous radiator arrays — structures that must survive launch, deploy reliably, and avoid being hit by debris. Nobody has built one at data-center scale yet.

5. Latency. Light takes time to travel. A round trip to low Earth orbit and back adds tens of milliseconds at minimum — fine for batch AI training and some inference, but a real problem for interactive, real-time applications. Orbital data centers will likely serve training and batch inference, not your video call. The addressable market is large but not unlimited.

6. Space debris and traffic. Tens of thousands — or a million — additional large satellites means collision risk, debris generation, and an orbital environment that gets more dangerous with every launch. Regulators have barely begun to grapple with mega-constellations for communications; compute constellations are an order of magnitude larger in ambition. The FCC has not approved Starcloud's 88,000-satellite filing, and the one-million-satellite filing will face even harder questions.

7. Terrestrial competition is not standing still. Earth data centers are getting more efficient too — better chips, better cooling, small modular reactors and dedicated power plants co-located with compute. The bar orbital data centers must clear keeps rising. Space does not just need to be good; it needs to be better than the rapidly improving alternative.

The balanced verdict: none of these problems is obviously fatal, but together they explain why Google says "economical in the 2030s" and not "profitable next quarter." The physics favors space; the engineering and economics are still an open question.

What This Means for the AI Compute Crunch on Earth

Step back from orbit for a moment, because the terrestrial context is what makes this story urgent rather than merely cool.

AI's electricity demand is exploding. Data centers are straining regional power grids across the United States, utilities are scrambling to add capacity, and communities are pushing back against new facilities. The industry's answer so far has been more power plants, bigger grid connections, and eye-watering capital expenditure — the same capex wave that just wiped hundreds of billions off tech stocks as investors questioned whether the spending will ever pay off.

Orbital data centers are, in a sense, the industry's escape valve: if Earth cannot site and power enough data centers, put them where power is free and nobody complains. It is the same logic driving the push for dedicated nuclear plants for data centers and data centers in the Arctic for free cooling — except taken to its ultimate conclusion.

For readers following the AI infrastructure story on this blog — from our coverage of the GPU shortage to the NVIDIA Rubin rollout — the orbital race is the next chapter of the same narrative: compute demand is outrunning Earth's easy answers, and the industry is now willing to consider answers that would have sounded like science fiction five years ago.

How an Orbital Data Center Actually Works: The Architecture

It helps to picture what a real orbital data center would look like, because it is nothing like a server farm with a roof. Start with the satellite bus — the spacecraft's body — carrying racks of AI accelerator chips instead of cameras or telecom gear. Unfolding from the bus are two enormous structures: solar arrays sized to generate hundreds of kilowatts or even megawatts, and radiator panels that glow dull red as they shed the chips' waste heat into space. In Google's vision, the satellite flies in a dawn–dusk sun-synchronous orbit, a special path that keeps it in near-constant sunlight so the arrays never stop producing.

The satellites do not work alone. In the Suncatcher architecture, dozens or hundreds fly in compact formations, hundreds of meters apart, stitched together by free-space optical transceivers — laser beams carrying tens of terabits per second between spacecraft. That turns the constellation into a single distributed computer: a training job can be split across many satellites the way it is split across many racks on Earth, with the lasers playing the role of the data center's internal network fabric.

On the ground, the system needs gateway stations — large antennas that receive results and send up new workloads — plus the ordinary terrestrial infrastructure for everything latency-sensitive. The current prototype downlinked its haiku over Svalbard, Norway, one of the few places with the geography to see many polar-orbiting satellites per day. A production system would need a global network of such gateways, or relay satellites that pass data along until one of them is over a ground station.

The key architectural insight: an orbital data center is not a server farm launched into space — it is a fundamentally different machine, where power generation, heat rejection, networking, and orbital mechanics are all part of the computer's design. That is why the engineering challenge is so deep, and why the companies that solve it first will own a genuine moat.

The Timeline: What to Watch Next

This story will be decided by hardware milestones, not press releases. Here is the calendar that matters:

• Ongoing through ~October 2027: Google's Suncatcher prototype test campaign — watch for published results on radiation effects and thermal performance. Good results de-risk the entire field; bad results push timelines years out.
• 2027: Google's two follow-on Suncatcher satellites with higher power and inter-satellite laser links — the first test of the "cluster in orbit" concept.
• 2027: Starcloud-2, targeting orbital inference for paying customers including U.S. government agencies — the first real commercial test of the model.
• 2027–2028: Starship's commercial cadence — the variable everything depends on. Watch launch cost per kilogram, not headlines.
• Late 2020s: Starcloud-3 (200 kW) and Cowboy Space's megawatt upper-stage data centers — the first systems at a scale where the economics can actually be measured.
• 2030s: Google's target window for economical orbital data center infrastructure — if the physics, the rockets, and the regulators all cooperate.

Frequently Asked Questions About AI Data Centers in Space

Did Google really put a data center in space?

Google launched a prototype satellite (Project Suncatcher) carrying four TPUs. On October 30, 2026, it ran its first AI workloads in orbit — including a haiku written by the Gemma 4 model — and returned the results to Earth. It is a 1-kilowatt testbed, not a production data center. Google targets economical orbital infrastructure in the 2030s.

What is Project Suncatcher?

Google's research program for space-based AI data centers. The long-term vision is constellations of satellites flying in close formation, connected by laser links at tens of terabits per second, running distributed AI workloads on solar power. The current prototype is step one: proving the chips survive and function in orbit.

Why did the satellite write a haiku?

Engineers chose a haiku as the first prompt because it is a complete, verifiable inference task: prompt goes up, the model generates text on the onboard TPU, the answer comes down. It proves the full compute loop works in orbit. They also tested limericks and telemetry-reporting prompts.

Why can the satellite only run its TPU for 20 minutes at a time?

The satellite generates up to 1 kW of solar power, but a data-center TPU draws more than that. Only one of the four TPUs can run at once, in ~20-minute bursts, before the batteries need recharging. Future satellites will generate more power.

What is Starcloud?

A Redmond, Washington startup (founded January 2024, Y Combinator Summer 2024) building orbital data centers. It flew an NVIDIA H100 to orbit in November 2025 — the first data-center GPU in space — trained an AI model there, and has raised $450 million at a $2.3 billion valuation, with NVIDIA as an investor and partner on the space-rated Vera Rubin Space-1 module.

How many satellites are planned?

Filings (not approvals): Starcloud has applied to the FCC for up to 88,000 satellites; SpaceX/xAI have filed for up to one million. These are regulatory placeholders expressing ambition. Actual deployed hardware today: a handful of prototypes.

Isn't space too expensive for data centers?

Today, yes — that is why everything flying is a prototype. The economic case depends on launch costs falling dramatically (Starship at ~$500/kg is the target) combined with near-zero energy costs in orbit (Starcloud projects $0.002–$0.005/kWh). Google's own target for economical operation is the 2030s.

What about space junk?

It is one of the seven hard problems. Tens of thousands of large new satellites increase collision risk and debris. Regulators have not approved the mega-filings, and orbital traffic management rules are still being written. This is a genuine constraint, not a footnote.

Will orbital data centers replace Earth data centers?

No — not for decades, if ever. Latency alone rules out interactive workloads. The realistic near-term role is AI training and batch inference — workloads that can tolerate delay. Earth data centers will remain the backbone; orbit would be an overflow valve for the most power-hungry jobs.

Should I invest in orbital data center companies?

This blog does not give financial advice. What we can say factually: every company in this space is pre-revenue or early-revenue, valued on future expectations (Starcloud at $2.3B with one GPU in orbit), and the key risks — launch costs, radiation, regulation — are outside any startup's full control. Treat it as high-risk speculation, not infrastructure.

The Bottom Line

A satellite wrote a haiku this week, and it might be the most important poem of the decade — not for its literary merit, but for what it proves. A data-center AI chip can run in orbit, do real work, and phone home. The loop is closed.

Everything else is still open: whether the chips survive a year of radiation, whether launch costs fall far enough, whether regulators allow constellations of tens of thousands of servers in the sky, and whether the economics beat a data center next to a power plant in Texas. Google — characteristically — is giving itself until the 2030s to find out. Starcloud and its backers are moving faster, with $450 million saying the answer is yes.

My take: do not think of orbital data centers as a replacement for the data centers being built today. Think of them as a call option on two futures at once: a future where AI's power demand keeps doubling, and a future where launch gets cheap. If both happen, the companies putting servers in orbit now will look prescient. If either fails, the haiku will remain a charming footnote.

The milestones to watch are concrete and dated: Google's radiation results through 2027, the two laser-linked Suncatcher satellites, Starcloud-2's commercial inference debut, and above all the cost per kilogram to orbit. I will update this guide as each one lands — from the ground, for now.

Sources

Key facts in this guide come from the following reporting, retrieved November 2, 2026. Figures that vary by source are presented as reported ranges.

• TechCrunch: "Google's first data center satellite is online and writing poetry" (Oct 30, 2026) — Suncatcher prototype online, four TPUs/one at a time, ~20-minute bursts, 1 kW power, Gemma 4 haiku, "operating nominally," two more satellites in 2027
• Silicon Report: "Google's prototype space data center is online and communicating" — Transporter-18 launch, three-week commissioning, 67 MeV proton-beam tests, 2-kilorad HBM threshold, dawn–dusk orbit vision with tens-of-terabits optical links
• The Laborecon: "Google's first orbital TPU satellite working" (Nov 2026) — Planet Labs build, Oct 1 Falcon 9 launch, NPR corrections (standard orbit + batteries), Travis Beals "very minimal test," ~1-year mission, two more satellites early 2027
• TechWorthIt: "Google's Suncatcher satellite returns first AI response from orbit" — Svalbard downlink, limerick and telemetry prompts, prompt-source discrepancies, 2 krad(Si) vs 750 rad(Si) requirement
• GeekWire: "Starcloud raises $250M for orbital data centers with Nvidia's help" (Aug 21, 2026) — $250M extension, $2.3B valuation, $450M total, NVIDIA/Cisco investments, H100 flight Nov 2025, Space-1 Vera Rubin Module (25x H100), 88,000-satellite FCC filing, SpaceX Starmind filing
• Awesome Agents: "Starcloud Raises $170M to Put AI Compute in Orbit" (Mar 2026) — $170M Series A at $1.1B, fastest YC unicorn (17 months), founders, Starcloud-2/Starcloud-3 plans, $0.002–$0.005/kWh energy projection, Crusoe Cloud resale, AWS blade
• Unite.AI: "Starcloud Raises $250M Series A Extension at $2.3B Valuation" — round details, NVIDIA collaboration, manufacturing and launch-allocation use of funds, Cowboy Space $275M context
• The Investor Society: "Starcloud Raises US$250M" (Aug 2026) — founder backgrounds (Johnston, Feilden, Oltean), Woodinville 100,000-sq-ft facility, Starcloud-2 2027 government inference target, FCC filing status
• AInvest: "Starcloud's October 2026 Launch Could Validate the Orbital Data Center S-Curve" — $500M→$39.09B market projection (67.4% CAGR), first AI model trained in space milestone
• TechEchelon: "Google's First Orbital Data Center Satellite Comes Online" — milestone summary, 2027 follow-on satellites

Related on AI Tech Hub: our NVIDIA Rubin guide (the platform behind the space-rated Vera Rubin module) and The Strata Effect (why AI compute demand is outrunning Earth's easy answers).