The orbital compute race

1. 01The idea is a live engineering program, not a thought experiment.Google flies its first TPU satellite on October 1. Starcloud trained a model in orbit in December 2025. China has run 12 computing satellites since May 2025.
2. 02The physics case is solar power.A panel in dawn-dusk sun-synchronous orbit sees near-constant sunlight and yields 5 to 8 times more energy a year than the same panel on Earth. No grid queue, no land, no water.
3. 03The economics case is launch cost, and only launch cost.Google's paper puts parity at under $200 per kg, projected for the mid-2030s. Today is about $1,500. Every bullish timeline is a bet on Starship.
4. 04Cooling is the hardest problem, not the fatal one.Heat can only radiate. One 700 W H100 needs about 1.4 m² of radiator. Two independent 2026 studies still conclude passive radiators are manageable and cheap relative to the rest.
5. 05Inference goes to orbit first. Frontier training stays on Earth.Laser mesh networks deliver two to three orders of magnitude less bisection bandwidth than a terrestrial cluster. Fine for serving models, bad for training them.
6. 06Nobody can fix a chip in orbit.About 9% of chips fail a year and cannot be swapped, which one study turns into 38% extra hardware over five years. GPUs also go obsolete in two to three years.
7. 07The timelines split into two camps.Musk: cheapest compute in space within two to three years, factory output from late 2027. Neuberger, Forethought and Orbital Gateway: meaningful scale is a 2030s event.
8. 08The externalities are real and unpriced.A million-satellite constellation implies about a billion collision-avoidance maneuvers a year, a satellite in nearly every telescope image, and 100 to 1,000 times more reentry mass than current atmospheric models assume.
9. 09It is a geopolitical race.China has the only working multi-satellite computing constellation and a five-year policy plan. The US has three megaconstellation filings and one cautious hyperscaler. Europe has studies.
10. 10The AI labs are buying the ground, not the sky.OpenAI has more than $1.4 trillion of terrestrial compute deals and its own inference chip. Anthropic pays SpaceX $1.25 billion a month for GPUs in Memphis and has only "expressed interest" in orbit. Altman calls space data centers "ridiculous" for this decade.
11. 11For builders, compute keeps getting cheaper. Context stays scarce.Every model in orbit will have read the same internet. None will have read your business. The knowledge layer is where the value moves.

Google alone processed more than 3.2 quadrillion tokens a month as of September 2026, up sevenfold in a year. The IEA expects data center electricity use to more than double by 2030.

In the US, new capacity waits in multi-year queues to connect to the grid, so labs went off-grid. xAI installed hundreds of megawatts of gas turbines in Memphis. OpenAI and Oracle placed large turbine orders for Texas. Turbines now have their own backlogs, and local opposition to data centers is growing.

Once you are going off-grid anyway, the question changes: where is the cheapest, most abundant power? Several serious people now answer "orbit." The Christian Science Monitor's framing in September: tech leaders "faced with earthly opposition" want to build in space.

Announced November 4, 2025 with a preprint paper by Blaise Agüera y Arcas, Travis Beals, James Manyika and colleagues. Google frames it next to its quantum and Waymo bets: a decade-scale program that works backward from an end state.

The October 1 mission

The satellite is called MVP. Refrigerator-sized, built by Planet, riding SpaceX's Transporter-18 rideshare on a Falcon 9 from Vandenberg. Four Trillium TPUs, about 1 kW of solar, Gemini workloads. The cooling stack (thermal interface material, aluminum and copper heat pipes, radiators) allows about 15 minutes of compute before shutdown. Planned operations about one year; six-year orbital life before reentry. The original plan was two custom satellites in early 2027, so this flight pulls the first data forward by months.

• Tested on the ground: three-axis vibration at launch loads (10 g sustained, 50 to 100 g on components), proton beam at UC Davis, thermal vacuum chamber.
• Next: two satellites in 2027 to prove laser links between moving spacecraft, "hitting a coin-sized target from kilometers away while both ends move." Then clusters of 80-plus. Google has floated a single satellite the size of a football field.
• Open problems named by the authors: thermal management at TPU density, on-orbit reliability with no repair, ground links through turbulence, single-event effects on training jobs, and an integrated power-compute-radiator satellite design.
• Context: Google owns roughly 6% of SpaceX. James Manyika says nothing operational is imminent.

On January 30, 2026 SpaceX filed with the FCC for up to one million orbital data center satellites between 500 and 2,000 km. The FCC accepted the application on February 4. The filing calls the constellation "a first step towards becoming a Kardashev II-level civilization."

Starcloud (Redmond, WA; founded 2024 as Lumen Orbit; Y Combinator; CEO Philip Johnston) published the September 2024 white paper that both Wikipedia and the TU Eindhoven paper credit as the first widely cited proposal to actually build orbital data centers.

• Starcloud-1, November 2025: a 60 kg satellite with the first Nvidia H100 in orbit, about 100× any prior space GPU. In December it trained Andrej Karpathy's NanoGPT on Shakespeare and ran Google's Gemma, which opened with "Greetings, Earthlings."
• Money: $250M Series A extension at $2.3B post-money (August 21, 2026), led by Manhattan West, with Nvidia and Cisco Investments. $450M raised in total.
• Plans: an FCC filing for 88,000 satellites. Starcloud-2 in October 2026 with Nvidia Blackwell plus several H100s. Future satellites on Nvidia's Space-1 Vera Rubin module (25× H100). A 100,000 sq ft factory in Woodinville. Long-term: a 5 GW orbital data center with panels 4 km on a side, claimed 10× lower energy cost. Satellite life: five years, matched to the chips.

Blue Origin announced TeraWave in January 2026, a 5,408-satellite optical backbone with 6 Tbps links, then filed Project Sunrise on March 19: 51,600 satellites in sun-synchronous orbits from 500 to 1,800 km. NASA objected on May 12, citing overlap with human spaceflight altitudes, debris risk and vague end-of-life plans. Amazon separately objected to SpaceX's filing; SpaceX called the arguments "naïve." Jeff Bezos expects gigawatt data centers in space within 10 to 20 years.

The Three-Body Computing Constellation, led by ADA Space and Zhejiang Lab and announced by Wang Jian at the 2024 Wuzhen summit, launched 12 computing satellites on May 14, 2025. Each carries an 8-billion-parameter model and 744 TOPS; the batch totals about 5 POPS, 30 TB of storage and 100 Gbps laser links.

• Feb 2026: nine months of in-orbit testing complete; ten AI models validated in orbit, including Alibaba's Qwen3.
• Mar 2026: the constellation ran the open-source OpenClaw agent in orbit to voice-command a humanoid and a quadruped robot on Earth.
• Jul 2026: two satellites held a continuous laser link for 192 hours at up to 1,000 km with 99.99% availability.
• Plan: 2,800 satellites (2,400 inference, 400 training) at 500 to 1,000 km; 100 by 2027, about 1,000 by 2030, complete by 2035; 1,000 POPS. Third batch of 14 due Q4 2026. ITU has approved the filing.
• State layer: a January 2026 filing for 200,000 satellites, a Beijing space computing centre opened June 2026, and a 2026 to 2031 five-year plan for gigawatt-level space digital infrastructure. Other constellations: Nayuta ALAYA (12,500), Xingshu (1,000), GalaxySpace, Landspace, China Mobile.

The companies that consume frontier compute are not waiting for orbit. Their 2026 moves show where they think the constraint is: delivered megawatts on Earth in the next 24 months. Space is a hedge, and they say so.

SpaceX as landlord

Before it launches a single compute satellite, SpaceX has become one of the largest sellers of terrestrial AI compute. Anthropic pays $1.25B a month for Colossus 1. Google pays $920M a month from October 2026 to June 2029 for about 110,000 Nvidia GPUs, "bridge capacity" for Gemini Enterprise. Both contracts surfaced in SpaceX's filings ahead of its June 2026 IPO (about $75B raised near a $1.75T valuation). SpaceX "expects to enter into additional similar services contracts." Its revenue before these deals was about $18B a year. The labs are financing the orbital bet while doubting it.

For scale: the largest AI data centers

Epoch AI's September 2026 ranking: xAI's Colossus 2 at about 946 MW, Anthropic-Amazon New Carlisle at about 910 MW, Microsoft Fairwater Atlanta at about 636 MW. Fairwater Wisconsin targets 2.26 GW by Q2 2028 at more than $100B. Meta expanded Hyperion to 5 GW and over $50B, and guided 2026 capex to $125 to $145B. Alphabet committed more than $180B of 2026 capex and sold $80B of equity to fund it. Rough benchmark: about $29B per gigawatt of facility.

• Launch cost does all the work. Forethought (May 2026): parity with off-grid terrestrial power at about $250/kg, competitive with terrestrial data centers at about $100/kg, cheaper than any Earth energy at $50/kg. Their five-year cost of 1 GW: grid $37.6B, off-grid gas plus solar $36.8B, orbital at $100/kg $35.0B, at $50/kg $30.3B. Conclusion: no meaningful share before 2030; competitive within 3 to 5 years only if Starship stays on track.
• The cooling math cuts both ways. IEEE Spectrum's model, even at $44/kg and $0.20/kWh on Earth, puts a GPU in orbit at 10× the annual cost. Forethought counters that carbon-composite radiators reach 163 to 346 W/kg, making thermal 2 to 5% of total cost. TU Eindhoven's Kees van Berkel: a 100 kW satellite at 370 K needs about 38 m², "plausible."
• Training does not fit. Van Berkel (July 2026): terrestrial Clos networks deliver about 28.8 TB/s bisection bandwidth; orbital torus meshes 2.25 to 10 TB/s with two orders of magnitude worse latency. A 288× to 12,880× penalty for frontier training. Inference on single satellites is feasible; orbital training "within the next 2 to 3 years is not credible."
• No repair, fast obsolescence. Roughly 9% of chips fail a year and stay failed, compounding to about 38% overbuild over five years. GPUs turn over every two to three years. Blaine Curcio: a state-of-the-art data center launched at enormous cost "might be obsolete in a couple of years."
• Downlink. Transcelestial's Rohit Jha: "If you can't talk to these AI systems, then those data centers are useless." Hyperscale in orbit may be five to seven years away and may need nuclear power.
• Debris and traffic. Starlink dodged about 300,000 potential collisions in 2025. Hugh Lewis calculates a million-satellite constellation needs about a billion avoidance maneuvers a year: "ridiculous." Samantha Lawler's CRASH Clock: if every satellite went dark today, the first collision comes in under three days. The ITU review process is called "definitely broken" by insiders.
• Astronomy. 500,000 satellites would put one in almost every telescope image taken anywhere. At a million there would at times be more visible satellites than stars.
• Atmosphere. Reentering satellites leave aluminum oxide aerosols that affect ozone. Models assume about 10,000 tonnes a year of reentry mass. Musk's vision implies one to ten million. ESA's Draco mission (2027) will measure a controlled breakup.
• Regulators and people. NASA's objection to Sunrise shows "no permits in space" is not quite true. Ariel Ekblaw (Aurelia Institute) calls the timelines optimistic. Evelyn Chow (Neuberger): "It's a next-decade event."

1. Sep 2024Starcloud white paper (then Lumen Orbit): the first widely cited proposal to build orbital data centers.
2. Nov 2024Wang Jian announces China's Three-Body Computing Constellation at Wuzhen.
3. May 14 2025China launches 12 computing satellites on a Long March 2D.
4. Oct 2025Musk: cheapest AI compute will be in space within 2 to 3 years. Bezos: gigawatt space data centers in 10 to 20 years.
5. Nov 4 2025Google announces Project Suncatcher and publishes the paper.
6. Nov–Dec 2025Starcloud-1 launches with an H100, trains NanoGPT and runs Gemma in orbit.
7. Jan 2026Blue Origin TeraWave. China files for 200,000 satellites. Musk repeats the claim at Davos.
8. Jan 30 2026SpaceX files for up to 1,000,000 satellites. FCC accepts on Feb 4.
9. Feb 2026Starcloud files for 88,000. SpaceX acquires xAI. China completes nine months of in-orbit testing.
10. Mar 2026Blue Origin files Project Sunrise (51,600). Nvidia announces Space-1 Vera Rubin. China's constellation commands robots on Earth.
11. Apr 2026Anthropic expands its Google and Broadcom TPU deal to 3.5 GW, online from 2027.
12. May 2026Cowboy Space raises $275M. NASA objects to Sunrise. Forethought publishes its feasibility study. Edge Aerospace wins an ESA contract.
13. May 6 2026Anthropic takes all of Colossus 1 (300 MW, 220,000 GPUs) and "expresses interest" in multi-gigawatt orbital compute. May 20: SpaceX's IPO filing shows $1.25B a month through 2029.
14. Jun 2026SpaceX reveals AI1. Beijing opens a space computing centre.
15. Jun 5 2026Google agrees to pay SpaceX $920M a month for about 110,000 GPUs from October. SpaceX IPOs near $1.75T. Jun 24: OpenAI and Broadcom unveil Jalapeño.
16. Jul 2026Van Berkel's cost-and-network paper. China's eight-day continuous laser link.
17. Jul 12 2026Altman calls space data centers "ridiculous" for this decade. Musk: "We start flying them next year."
18. Aug 2026SpaceX publishes Starmind specs. Starcloud raises $250M at $2.3B. ESPI warns Europe is behind.
19. Aug 25 2026Jalapeño's first results: 1.5 to 1.9× more work per watt than Blackwell. Anthropic signs about $44.6B with Nscale for 460 MW in West Virginia.
20. Sep 2026CNBC's four obstacles. Sophia Space's $300M framework. Musk's "100%" post. Google announces the Oct 1 launch.
21. Oct 1 2026Suncatcher MVP on Transporter-18.

1. Oct 1 2026MVP survives launch and the TPUs complete their 15-minute Gemini bursts. First-week telemetry sets the tone.
2. Oct 2026Starcloud-2 with Blackwell. Starcloud's next round gets priced on this.
3. Q4 2026China's third batch (14 satellites at 525 km) and any sign of the 100-by-2027 pace.
4. 2026–2027Starship cadence. Every economic model keys on flights per year and dollars per kg. Watch the first Starlink V3 deployments from Starship.
5. 2027Google's two-satellite laser test. SpaceX's first AI1 launches and Gigasat production. ESA's Draco reentry mission.
6. PendingFCC decisions on the SpaceX, Blue Origin and Starcloud filings, and whether SpaceX's milestone waiver is granted.
7. PendingNvidia Space-1 shipments. The first radiation-qualified Rubin-class module in orbit is the hardware inflection point.
8. PendingAny public chip failure rate in orbit. The least-discussed variable, and the one that could quietly double the cost.

• Compute trends toward a commodity produced like energy. Whether orbit wins or loses, the industry now competes on the marginal cost of a watt of inference. On a ten-year horizon the price of intelligence keeps falling. A business betting on "AI is expensive" as a moat is betting against everyone in this brief.
• Inference at the edge of everything. The first orbital workloads serve models rather than train them. Add direct-to-cell connectivity and you get low-latency AI anywhere on the planet without terrestrial infrastructure. Design products assuming inference is cheap and everywhere.
• The scarce input is context. Every model in orbit will have read the same internet. None will have read your calls, documents, decisions and methods. As the cost of running a model falls, value shifts to the knowledge layer that makes it useful for one specific business. The space race strengthens the company-brain thesis rather than weakening it.
• Sovereignty becomes a purchasing criterion. China's program is explicitly about controlling its own space data. Europe's policy institute is warning about dependence. Expect "where does my inference run" to become a compliance question the way data residency did.
• The energy story will shape public opinion about AI. Orbit is the industry's answer to "you are taking our water and power." Expect the debris and astronomy arguments to intensify as filings turn into launches.