A 60-kilogram satellite carrying an Nvidia H100 GPU successfully trained the first large language model in space on December 10, 2025, marking a pivotal moment in computing history. Starcloud's achievement demonstrates that the most power-hungry chips on Earth can operate in the vacuum of space, potentially unlocking unlimited solar energy for AI workloads that currently strain terrestrial power grids.
TL;DR
Starcloud became the first company to train an LLM in orbit, running NanoGPT on an Nvidia H100 GPU aboard its Starcloud-1 satellite. Google plans to launch TPU-equipped satellites by early 2027 through Project Suncatcher, while China's Three-Body Computing Constellation aims to deploy 2,800 AI satellites by 2030. The economic case hinges on launch costs dropping below $200 per kilogram and the promise of solar panels generating up to 8x more power in orbit than on Earth. For data center operators facing a projected tripling of power demand by 2030, orbital infrastructure represents a potential escape valve from terrestrial grid constraints.
The Terrestrial Power Crisis Driving Space Ambitions
Data centers consumed 4.4% of total U.S. electricity in 2023 and could reach 6.7% to 12% by 2028, according to the Department of Energy. Global electricity consumption for data centers will double to 945 TWh by 2030, with AI-optimized servers growing from 21% of data center power usage in 2025 to 44% by 2030.
Power Demand Projections
| Region | 2024 | 2030 | Growth |
|---|---|---|---|
| U.S. Data Centers | ~45 GW | 134.4 GW | ~3x |
| Global Data Centers | 460 TWh | 945-980 TWh | ~2x |
| AI Servers (Global) | 93 TWh | 432 TWh | ~5x |
Local officials have begun rejecting new data center proposals that strain power grids and consume cooling water. The U.S. alone faces a potential 2.3 GW gap between projected data center load and expected new generation capacity in the PJM interconnection by 2030.
Space offers a compelling alternative. The Sun emits more power than 100 trillion times humanity's total electricity production. In the right orbit, solar panels operate nearly continuously and generate up to 5-8x more output than equivalent systems on Earth, without atmospheric interference.
Starcloud: First LLM Trained in Space
The Historic Achievement
Nvidia-backed startup Starcloud launched the Starcloud-1 satellite aboard a SpaceX rocket on November 2, 2025. The 60-kilogram satellite, approximately the size of a small refrigerator, carries the first Nvidia H100 GPU to reach orbit.
"The H100 is about 100 times more powerful than any GPU computer that has been on orbit before," Philip Johnston, CEO and cofounder of Starcloud, told IEEE Spectrum.
The company trained NanoGPT (a large language model created by OpenAI founding member Andrej Karpathy) on the complete works of Shakespeare, producing a model that speaks in Shakespearean English. Starcloud-1 also runs and queries Google's Gemma LLM in orbit.
Starcloud-1 Technical Specifications
| Specification | Details |
|---|---|
| Satellite Mass | 60 kg |
| Primary GPU | Nvidia H100 (700W TDP) |
| Compute Performance | 100x previous space GPUs |
| Launch Vehicle | SpaceX Falcon 9 |
| Launch Date | November 2, 2025 |
| Orbit | Terminator line (day/night boundary) |
Solving the Thermal Challenge
Putting a 700-watt GPU into orbit presented a massive thermal challenge. On Earth, H100 chips require complex water and air cooling systems. In space, no air exists to carry heat away through convection.
Starcloud CTO Adi Oltean and his engineering team designed a system relying entirely on radiative cooling, using large specialized panels to radiate the intense heat generated by the GPU directly into the freezing void of deep space (average temperature: 2.7 Kelvin or -270.45°C).
"A lot of innovation and hard work" went into the solution, Oltean stated.
Starcloud Roadmap
| Phase | Timeline | Specifications |
|---|---|---|
| Starcloud-1 | November 2025 | Single H100, 60 kg satellite |
| Starcloud-2 | October 2026 | Multiple H100s + Blackwell platform |
| Commercial Satellite | 2026 | 1 MW solar array |
| Hypercluster | When Starship enters service | 5 GW, 4km x 4km solar array |
The company raised more than $10 million through Y Combinator backing and Nvidia Inception program participation. Nvidia's Blackwell platform integration will deliver up to 10x improvements over the current Hopper architecture.
Google Project Suncatcher: TPUs in Orbit
The Vision
Google announced Project Suncatcher in November 2025, a moonshot exploring solar-powered satellite constellations equipped with TPUs and free-space optical links to scale machine learning compute in space.
The company will partner with Planet Labs to launch two prototype satellites by early 2027 into low Earth orbit approximately 400 miles above Earth.
Technical Architecture
| Component | Specification |
|---|---|
| Orbit | Dawn-dusk sun-synchronous, 650 km altitude |
| Cluster Design | 81 satellites, 1 km radius |
| Satellite Spacing | 100-200 meters between nearest neighbors |
| Inter-Satellite Links | 800 Gbps each-way (1.6 Tbps total) via DWDM |
| Target Bandwidth | Tens of terabits per second |
| TPU Model | Trillium v6e Cloud TPU |
Radiation Testing Results
Google tested its Trillium TPUs in a 67 MeV proton beam to simulate low-Earth orbit radiation levels:
| Test Result | Details |
|---|---|
| Most Sensitive Component | High Bandwidth Memory (HBM) subsystems |
| Irregularity Threshold | 2 krad(Si) cumulative dose |
| Expected 5-Year Mission Dose | ~0.7 krad(Si) (shielded) |
| Safety Margin | ~3x expected exposure |
| Maximum Tested Dose | 15 krad(Si) with no permanent failures |
Economic Projections
Google estimates space-based AI clusters could become economically feasible by 2035, contingent on launch costs dropping below $200 per kilogram (currently ~$1,400/kg via SpaceX).
The Global Orbital Data Center Race
Major Players and Timelines
| Company/Initiative | Status | Target Scale | Timeline |
|---|---|---|---|
| Starcloud | First LLM trained | 5 GW orbital facility | 2030s |
| Google Suncatcher | Planning | 81-satellite clusters | 2027 demo |
| China Three-Body | 12 satellites launched | 2,800 satellites | 2030 |
| SpaceX Starlink V3 | Development | Starlink with AI compute | 2026 |
| Blue Origin | R&D | Gigawatt-scale | 10-20 years |
| Axiom Space | Development | Free-flying ODC nodes | End of 2025 |
China's Three-Body Computing Constellation
China launched 12 satellites on May 14, 2025, marking the debut of the "Three-Body Computing Constellation." Named after the science fiction novel and the gravitational physics problem, the constellation represents a collaboration between Zhejiang Lab, Alibaba Group, and other partners.
| Specification | Details |
|---|---|
| Current Satellites | 12 |
| Target by 2030 | 2,800 satellites |
| Onboard AI Model | 8 billion parameters |
| Per-Satellite Performance | 744 TOPS |
| Current Fleet Performance | 5 POPS (12 satellites) |
| Target Performance | Exa-scale (2030) |
| Storage per Satellite | 30 TB |
| Inter-Satellite Links | 100 Gbps laser |
SpaceX and Blue Origin
Elon Musk confirmed SpaceX plans to build orbital data centers by scaling up Starlink V3 satellites, which feature high-speed laser links. The company pitched orbital computing payloads as part of a share sale valuing SpaceX at $800 billion.
Jeff Bezos' Blue Origin has been engineering orbital data center components since late 2023. Bezos predicted gigawatt-scale space data centers within 10 to 20 years: "We will be able to beat the cost of terrestrial data centers in space in the next couple of decades."
Technical Challenges and Engineering Solutions
The Radiation Problem
Space radiation represents the biggest reliability challenge for electronic hardware in orbit. Radiation causes problems ranging from complete burnout to bit flips that corrupt data.
| Environment | Radiation Level | Hardware Requirements |
|---|---|---|
| Low Earth Orbit (LEO) | Moderate | Enhanced COTS with shielding |
| Geosynchronous Orbit | Extreme | Fully radiation-hardened |
| Deep Space | Severe | Ground-up rad-hard design |
The performance gap between space-rated and commercial hardware remains significant. RAD5500 processors commonly used in space achieve 0.9 GFlops, while commercial Nvidia A100 GPUs reach 156 TFlops. Space chips typically lag three to four orders of magnitude behind terrestrial counterparts.
Thermal Management in Vacuum
Without air, convection cooling becomes impossible. Heat pipes, fluid loops, and radiative panels must conduct heat to radiators that emit infrared radiation into space.
| Cooling Method | Application | Limitations |
|---|---|---|
| Solid Conduction | Most common | Limited capacity |
| Heat Pipes | Medium power | Weight constraints |
| Fluid Loops | High power (ISS, Shenzhou) | Weight, micro-vibrations |
| Radiative Panels | All space applications | Requires large surface area |
High-power AI chips require larger radiation cooling plates, increasing satellite weight and driving up launch costs. Starcloud's innovation in radiative cooling for a 700W H100 demonstrates that consumer-grade AI accelerators can operate in orbit with proper thermal engineering.
Launch Cost Economics
The economic viability of orbital data centers depends on continued reductions in launch costs.
| Launch Vehicle | Cost per kg (2025) | Projected Cost |
|---|---|---|
| NASA Space Shuttle (historical) | $25,000 | N/A |
| SpaceX Falcon 9 | $2,720 | — |
| SpaceX (commercial rate) | $1,400 | — |
| SpaceX Starship (single-use) | — | $250-600 |
| SpaceX Starship (reusable) | — | $100-200 |
| Musk's target | — | $10 |
| Google Suncatcher threshold | — | <$200 |
A 5 GW orbital deployment would require roughly 100 Starship launches at projected costs.
Economic Case for Orbital Infrastructure
Operating Cost Comparison
| Metric | Terrestrial | Orbital | Advantage |
|---|---|---|---|
| Power Cost | ~$0.05/kWh (best case) | ~$0.001/kWh | ~97% lower |
| Solar Panel Output | 1x (mid-latitude) | 5-8x | 5-8x more productive |
| Carbon Emissions | Baseline | 10x lower | 90% reduction |
| Cooling Water | Required | None | Infinite heat sink |
| Expansion Potential | Land-constrained | 50x larger | Unlimited space |
Starcloud projects energy costs 95% lower than terrestrial data centers powered by natural gas, even including launch expenses.
Sustainability Benefits
Orbital data centers eliminate water consumption for cooling. Terrestrial facilities rely on evaporation towers that consume millions of gallons annually. Space-based alternatives radiate waste heat directly into the vacuum.
Analysis suggests the carbon footprint of launching hardware could be offset within five years of operation, after which facilities run indefinitely on renewable solar energy.
Remaining Challenges and Risks
Technical Hurdles
| Challenge | Current Status | Required Breakthrough |
|---|---|---|
| Launch Costs | $1,400/kg | <$200/kg |
| Radiation Hardening | 3-4 order magnitude gap | Near-commercial performance |
| High-Bandwidth Ground Links | Limited | Terabit-class downlinks |
| Orbital Debris | Growing concern | Active debris management |
| On-Orbit Maintenance | Not possible | Robotic servicing |
| Chip Replacement | 5-6 year lifecycle | Cost-effective refresh |
Space Environment Risks
Tin whiskers form more easily in vacuum, potentially causing short circuits between device pins. Outgassing from packaging materials can impair neighboring device performance. In 2022, Starlink lost 38 satellites due to solar activity.
Timeline Uncertainty
Jeff Bezos estimated 20 years before orbital solutions become cheaper than ground-based infrastructure. Google targets 2035 for economic viability. Near-term deployments will serve niche applications rather than replacing terrestrial data centers.
Deployment Considerations for Infrastructure Teams
Orbital data centers present deployment and operational challenges that differ fundamentally from terrestrial facilities. Introl's network of 550 field engineers support GPU infrastructure deployments across 257 global locations, providing the ground-based expertise that complements emerging space-based compute.
Ground Segment Requirements
| Component | Function | Complexity |
|---|---|---|
| Ground Stations | Uplink/downlink data | Multiple global sites required |
| Mission Control | Satellite management | 24/7 operations |
| Edge Caching | Latency mitigation | Co-located with users |
| Terrestrial Backup | Redundancy | Full capacity mirror |
Hybrid Architecture Implications
Early orbital deployments will augment rather than replace terrestrial infrastructure. Organizations should consider:
- Workload Classification: Identify compute tasks tolerant of orbital latency (batch training, large-scale inference)
- Data Locality: Plan for high-bandwidth ground links and edge caching
- Redundancy Design: Maintain terrestrial capacity for mission-critical workloads
- Vendor Evaluation: Assess Starcloud, Google Cloud, and emerging providers
Key Takeaways
For Infrastructure Planners
- Orbital data centers remain 5-15 years from cost parity with terrestrial facilities
- Starcloud's H100 success proves high-power AI accelerators can operate in space
- Plan ground station and edge infrastructure to support future orbital compute
For Operations Teams
- Monitor Three-Body Constellation and Suncatcher progress for commercial availability
- Evaluate hybrid terrestrial-orbital architectures for batch workloads
- Prepare for new latency profiles in space-to-ground data transfers
For Strategic Decision-Makers
- Track launch cost trajectories as the key economic enabler
- Consider orbital compute as a hedge against terrestrial power constraints
- Watch China's 2,800-satellite constellation as a competitive benchmark
References
-
NVIDIA. "How Starcloud Is Bringing Data Centers to Outer Space." NVIDIA Blog. December 2025. https://blogs.nvidia.com/blog/starcloud/
-
CNBC. "Nvidia-backed Starcloud trains first AI model in space, orbital data centers." December 10, 2025. https://www.cnbc.com/2025/12/10/nvidia-backed-starcloud-trains-first-ai-model-in-space-orbital-data-centers.html
-
Google Research. "Exploring a space-based, scalable AI infrastructure system design." November 2025. https://research.google/blog/exploring-a-space-based-scalable-ai-infrastructure-system-design/
-
Google Blog. "Meet Project Suncatcher, a research moonshot to scale machine learning compute in space." November 2025. https://blog.google/technology/research/google-project-suncatcher/
-
IEEE Spectrum. "NVIDIA's H100 GPU Takes AI Processing to Space." December 2025. https://spectrum.ieee.org/nvidia-h100-space
-
Data Center Dynamics. "Project Suncatcher: Google to launch TPUs into orbit with Planet Labs." November 2025. https://www.datacenterdynamics.com/en/news/project-suncatcher-google-to-launch-tpus-into-orbit-with-planet-labs-envisions-1km-arrays-of-81-satellite-compute-clusters/
-
Space News. "China launches first of 2,800 satellites for AI space computing constellation." May 2025. https://spacenews.com/china-launches-first-of-2800-satellites-for-ai-space-computing-constellation/
-
CGTN. "China launches first batch of space computing satellite constellation." May 14, 2025. https://news.cgtn.com/news/2025-05-14/China-launches-first-batch-of-space-computing-satellite-constellation-1DmXIysPKww/p.html
-
S&P Global. "Data center grid-power demand to rise 22% in 2025, nearly triple by 2030." October 2025. https://www.spglobal.com/commodity-insights/en/news-research/latest-news/electric-power/101425-data-center-grid-power-demand-to-rise-22-in-2025-nearly-triple-by-2030
-
Gartner. "Electricity Demand for Data Centers to Grow 16% in 2025 and Double by 2030." November 2025. https://www.gartner.com/en/newsroom/press-releases/2025-11-17-gartner-says-electricity-demand-for-data-centers-to-grow-16-percent-in-2025-and-double-by-2030
-
IEA. "AI is set to drive surging electricity demand from data centres." 2025. https://www.iea.org/news/ai-is-set-to-drive-surging-electricity-demand-from-data-centres-while-offering-the-potential-to-transform-how-the-energy-sector-works
-
Department of Energy. "DOE Releases New Report Evaluating Increase in Electricity Demand from Data Centers." 2025. https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers
-
Scientific American. "Space-Based Data Centers Could Power AI with Solar Energy—At a Cost." 2025. https://www.scientificamerican.com/article/data-centers-in-space/
-
Tom's Hardware. "Jeff Bezos envisions space-based data centers in 10 to 20 years." 2025. https://www.tomshardware.com/tech-industry/artificial-intelligence/jeff-bezos-envisions-space-based-data-centers-in-10-to-20-years-could-allow-for-natural-cooling-and-more-effective-solar-power
-
TheStreet. "Musk, Bezos & Altman Eye New Data Centers—in Space." 2025. https://www.thestreet.com/technology/spacexs-musk-blue-origins-bezos-and-openais-altman-eye-space-data-centers
-
Interesting Engineering. "Bezos vs Musk: Space race heats up with new orbital data center push." 2025. https://interestingengineering.com/space/orbital-ai-data-centers
-
Singularity Hub. "Data Centers in Space: Will 2027 Really Be the Year AI Goes to Orbit?" December 2025. https://singularityhub.com/2025/12/19/data-centers-in-space-will-2027-really-be-the-year-ai-goes-to-orbit/
-
Singularity Hub. "Future Data Centers Could Orbit Earth, Powered by the Sun and Cooled by the Vacuum of Space." November 2025. https://singularityhub.com/2025/11/03/future-data-centers-could-orbit-earth-powered-by-the-sun-and-cooled-by-the-vacuum-of-space/
-
GeekWire. "Starcloud plans its next moves after training first AI model in space." December 2025. https://www.geekwire.com/2025/starcloud-power-training-ai-space/
-
Analytics India Magazine. "Starcloud Becomes First to Train LLMs in Space Using NVIDIA H100." December 2025. https://analyticsindiamag.com/ai-news-updates/starcloud-becomes-first-to-train-llms-in-space-using-nvidia-h100/
-
Data Center Dynamics. "Starcloud-1 satellite reaches space, with Nvidia H100 GPU now operating in orbit." November 2025. https://www.datacenterdynamics.com/en/news/starcloud-1-satellite-reaches-space-with-nvidia-h100-gpu-now-operating-in-orbit/
-
Benzinga. "Nvidia-Backed Starcloud Trains First LLM In Space Amid Orbital Datacenter Buzz." December 2025. https://www.benzinga.com/news/space/25/12/49325156/nvidia-backed-starcloud-trains-first-llm-in-space-amid-orbital-datacenter-buzz-ceo-calls-it-significant-first-step
-
News9live. "AI model trained in Earth orbit for first time as Nvidia backed Starcloud tests succeed." December 2025. https://www.news9live.com/technology/artificial-intelligence/starcloud-ai-in-space-first-llm-training-nvidia-h100-satellite-tests-2911048
-
Y Combinator. "Starcloud: Data centers in space." 2025. https://www.ycombinator.com/companies/starcloud
-
CNBC. "From data center spas to servers in space: How the energy crunch is reshaping cloud computing." December 2025. https://www.cnbc.com/2025/12/29/future-of-the-cloud-from-spas-to-orbital-space-data-centers.html
-
Silicon Angle. "Google to launch satellites equipped with its TPU AI chips in 2027." November 2025. https://siliconangle.com/2025/11/04/google-launch-tpu-equipped-satellites-2027/
-
The Conversation. "Google's proposed data center in orbit will face issues with space debris." 2025. https://theconversation.com/googles-proposed-data-center-in-orbit-will-face-issues-with-space-debris-in-an-already-crowded-orbit-270410
-
Engineering.org.cn. "Computing over Space: Status, Challenges, and Opportunities." 2025. https://www.engineering.org.cn/engi/EN/10.1016/j.eng.2025.06.005
-
Semi Engineering. "Radiation, Temperature, Power Challenges For Chips In Space." 2025. https://semiengineering.com/radiation-temperature-power-challenges-for-chips-in-space/
-
Per Aspera. "Realities of Space-Based Compute." 2025. https://www.peraspera.us/realities-of-space-based-compute/
-
Avnet Silica. "Overcoming the radiation challenges of space electronics." 2025. https://my.avnet.com/silica/resources/article/overcoming-radiation-challenges-of-space-electronics/
-
SpaceEyeNews. "China Launches World's First Space-Based Supercomputer: Why Now, and Why in Orbit?" 2025. https://spaceeyenews.com/china-space-based-supercomputer-three-body-constellation/
-
Unz Review. "Three-Body Computer Constellation – China is building the world's first orbital supercomputer network." 2025. https://www.unz.com/bhua/three-body-computer-constellation-china-is-building-the-worlds-first-orbital-supercomputer-network/
-
Technology.org. "Space Race Heats Up: Bezos and Musk Chase Orbital AI Data Centers." December 2025. https://www.technology.org/2025/12/11/space-race-heats-up-bezos-and-musk-chase-orbital-ai-data-centers/
-
Utility Dive. "US data center power demand could reach 106 GW by 2035." December 2025. https://www.utilitydive.com/news/us-data-center-power-demand-could-reach-106-gw-by-2035-bloombergnef/806972/
-
AI News Hub. "Space-Based Data Centres: The Future of AI Computing in 2025." 2025. https://www.ainewshub.org/post/space-based-data-centres
-
NextBigFuture. "SpaceX Starship Roadmap Lower Launch Costs by 100 Times." January 2025. https://www.nextbigfuture.com/2025/01/spacex-starship-roadmap-to-100-times-lower-cost-launch.html
-
The Strategist. "The Starship revolution in space." 2025. https://www.aspistrategist.org.au/the-starship-revolution-in-space/
-
PatentPC. "Rocket Launch Costs (2020-2030): How Cheap Is Space Travel Becoming?" 2025. https://patentpc.com/blog/rocket-launch-costs-2020-2030-how-cheap-is-space-travel-becoming-latest-pricing-data
-
Wikipedia. "Space launch market competition." https://en.wikipedia.org/wiki/Space_launch_market_competition
-
Concept to Cloud. "SpaceX Plans Orbital Data Centers with Starlink V3 by 2026." 2025. https://concepttocloud.com/news/spacex-orbital-data-centers-starlink-v3
-
Oninvest. "Musk's SpaceX and Bezos's Blue Origin begin race for orbital data centers." December 2025. https://en.oninvest.com/article/musk-s-spacex-and-bezos-s-blue-origin-begin-race-for-orbital-data-centers-wsj
Published: December 30, 2025
