On May 14, 2025, a CASC Long March 2D rocket left Jiuquan carrying 12 ADASpace satellites.

Only a few major Western publications ran it as a headline.

The satellites entered orbit, began communicating with each other via inter-satellite laser links at a company-reported 100 gigabits per second, and started processing data. By November 2025, they were running Alibaba’s Qwen-3 AI model on that constellation — executing real inference tasks, end to end, without transmitting raw data to the ground. The deployment was confirmed publicly in January 2026.

That same month — November 2, 2025 — a SpaceX Falcon 9 had put Starcloud-1 into orbit carrying a single Nvidia H100 GPU, the first data-center-class chip ever placed in space, 100 times more powerful than any GPU previously flown in orbit. By December, Starcloud-1 had trained NanoGPT on the complete works of Shakespeare and was running Google’s Gemma language model from orbit. An AI, in space, responding to queries from Earth.

Both milestones happened within weeks of each other.

Both sides claim firsts — Starcloud for training an LLM on a high-powered GPU in space, ADASpace for deploying a general-purpose model across a multi-satellite constellation.

The distinction matters less than the shared implication: orbital AI is no longer a whitepaper.

Two countries.

Two different bets.

Two architectures that cannot both be right.

This is the orbital compute race — and unlike most tech races you read about, this one has hardware in orbit, operational systems running real workloads, and genuinely competing visions for what winning looks like. What it does not have yet is a clear leader.

Understanding why requires going back to why this race is happening at all.

AI broke the energy math…

A single Nvidia H100 GPU draws approximately 700 watts.

A hyperscale AI training cluster runs tens of thousands of these cards simultaneously — all day, every day, without pause. Almost all that electricity converts to heat. Heat that must be removed, continuously, at scale, using water and more electricity, in a loop that never ends.

The U.S. Energy Information Administration puts data center electricity consumption at roughly 8% of commercial sector usage in 2024. By 2050, that number could hit 20%!

The International Energy Agency measures global data center power draw at 415 terawatt hours. A large U.S. data center consumes as much water as 2,600 households just for direct cooling.

In Northern Virginia — the data center capital of the world — the regional power grid is already saturated. The largest AI companies are not optimizing their way out of this. They are buying nuclear power plants. Not because nuclear is cheap or fast.

Land. Water. Power. Regulation. Terrestrial data centers are hitting every wall simultaneously.

The AI boom did not create this problem. It detonated it.

So where should all that compute go?

Up.☝️

The concept of an Orbital Data Center is simpler than it sounds: move the server infrastructure off the ground and into orbit.

Instead of leasing land in a data center park, negotiating power purchase agreements, and piping millions of gallons of cooling water through concrete facilities, you deploy compute hardware to a satellite operating at 700-800 kilometers altitude.

The reason companies with serious capital are pursuing this comes down to three physical properties of the orbital environment that no terrestrial site can replicate.

Ground-based data centers compete with cities for grid capacity, sign long-term energy contracts at climbing rates, and still face brownout risk during peak demand.

A satellite in a Dawn-Dusk Sun-Synchronous orbit rides the terminator line — the moving boundary where day meets night on Earth’s surface. Its solar panels stay in continuous sunlight with no shadow interruptions. Solar irradiance at that altitude runs 36% stronger than what reaches the ground, unfiltered by atmosphere or weather. The result is a power supply that is unmetered, unregulated, and structurally impossible to interrupt.

No grid. No utility contract. No brownout. No-brainer.

Cooling infrastructure can consume billions of dollars and millions of gallons of water annually.

There is no air in space, which means convective cooling — every fan and cooling tower on Earth — does not exist here.

What does exist is radiative cooling: any surface exposed to the vacuum sheds heat through electromagnetic radiation according to the Stefan-Boltzmann law.

Well, although this is not a passive free lunch — orbital compute systems need large, flat radiator panels with enough surface area to dump the heat output of dense AI clusters.

Starcloud describes the vacuum of deep space as an effective heat sink for their H100, using infrared radiation to shed thermal load without water or fans. The geometry has to be deliberately engineered — maximize radiating surface area while keeping launch mass manageable, since every kilogram of radiator panel is a kilogram not available for compute hardware. The mechanism that costs terrestrial operators billions of dollars and millions of gallons of water is, in orbit, a function of surface area and physics.

Light in a standard glass fiber optic cable travels troughly two-thirds the speed of light in a vacuum. The refractive index of glass is what causes the slowdown. That means laser communication between satellites, traveling through vacuum, moves approximately 50% faster than data through fiber (link for the paper).

For most internet applications, this difference is irrelevant. But for specific high-value use cases — financial routing between trading centers, real-time Earth observation processing, low-latency global AI inference — 50% faster transit is commercially significant. It is part of why SpaceX built laser inter-satellite links into Starlink at all.

The industry is splitting into two architectures, driven by different assumptions about what the technology can do today vs what it will eventually become.

Add AI compute to satellites already collecting data so the analysis happens before anything hits the downlink.

The economics of satellite bandwidth are brutal. A constellation imaging the Earth continuously generates more raw data than its downlinks can carry. Every byte transmitted costs money and takes time. An onboard AI accelerator that filters, analyzes, and compresses in orbit — transmitting only the actionable output — solves the bottleneck at the source.

This is already operational. ESA’s Φsat-2 satellite, launched August 2024, runs AI accelerators that perform cloud detection, vessel identification, and fire anomaly detection entirely in orbit.

China’s Three-Body Constellation ran a remote sensing model that surveyed 189 square kilometers of infrastructure in northwest China in November 2025 — processing the analysis in orbit and transmitting results rather than raw imagery.

Treating orbit as a location for purpose-built compute infrastructure — just like the way hyperscalers treat Ashburn, Virginia or Dublin, Ireland as places to put servers.

Starcloud’s November 2025 H100 satellite is the clearest proof of concept for this model.

One satellite. One chip. Running real AI inference from orbit.

CEO Philip Johnston’s long-term target is a 5-gigawatt orbital cluster — larger than Palo Verde’s 3.9 GW, the largest nuclear power plant in the United States.

The bigger players are moving behind it.

Google’s Project Suncatcher targets first test satellites by 2027.

In January 2026, SpaceX filed with the FCC for up to one million satellites to build a distributed orbital compute backbone — leveraging Starship launch economics and Starlink’s existing laser relay network.

Blue Origin’s TeraWave constellation plans approximately 5,400 networking satellites.

The physics case is airtight. The engineering has four open problems — and anyone presenting orbital compute as a near-term commercial product is glossing over all of them.

Standard commercial silicon was not built for the particle environment above Earth’s atmosphere. High-energy cosmic rays cause bit flips — single-event upsets where radiation changes a stored value inside a processor. One bit flip in the wrong register can crash a system or corrupt a computation. At scale, these events are a constant background condition, not a rare occurrence.

Space-grade radiation-hardened chips survive this environment but cost orders of magnitude more than commercial equivalents and typically run several performance generations behind what hyperscalers use on the ground. The chip capable of cost-effective AI training is not the chip that reliably survives the radiation environment — and vice versa.

China’s Zhongke Tiansuan is attempting to bridge this through software rather than hardware: error-correction algorithms and redundant chip architectures that allow cheaper industrial-grade silicon to operate reliably in orbit despite bit flip events (in short, they can bypass "bit flip" events by simply disabling or resetting affected nodes without system-wide failure). They have over 1,000 days of operational data on Jilin-1 to validate this approach. If it scales, it changes the economics of the entire field.

Radiative cooling works in principle.

Managing the thermal load of dense AI clusters across orbital cycles demands precise attitude control and carefully engineered radiator arrays.

Too little radiator surface area and the chips overheat.

Too much and launch mass becomes unmanageable — every kilogram of radiator panel is a kilogram not available for compute hardware.

Fluid-loop thermal systems are the leading approach. China’s BAIST-led project team plans to validate this specifically with a 2026 demonstration satellite.

No organization has proven it at AI-cluster scale yet.

AI compute hardware refreshes roughly every two years on Earth.

The H100 was state of the art at the time Starcloud-1 launched in November 2025. By the time you read this, Nvidia’s Blackwell architecture has already superseded it — and Starcloud’s own next satellite, launching October 2026, will carry Blackwell chips.

The first satellite’s expected mission lifetime is 11 months. It will deorbit and burn up in the atmosphere approximately when its replacement launches carrying the next generation of hardware.

That is the orbital obsolescence problem in miniature: by the time a constellation reaches commercial scale, the chips that launched first are already multiple generations behind.

In orbit, swapping hardware means another launch — or a satellite designed from the start to accept modular hardware upgrades via autonomous robotic servicing. Neither is cheap. Neither has been demonstrated at production scale.

Google’s November 2025 feasibility study — the most rigorous published analysis of orbital data center economics — puts the commercial breakeven at $200 per kg to LEO (Low Earth Orbit), projected achievable around 2035 if Starship scales to roughly 180 launches per year. This is a specific modeled threshold, not an industry consensus, but it is the only published figure grounded in detailed cost modeling. Below that threshold, the solar power and cooling economics make orbital data centers cost-competitive with terrestrial alternatives for the right workloads. Above it, the capital cost does not close.

A separate analysis from Nanyang Technological University found the carbon footprint of launching an orbital data center’s hardware could be offset within five years of operation. The sustainability math closes — if the hardware survives long enough to cross the breakeven line.

This is where most coverage goes wrong. The answer depends entirely on which dimension you measure.

As of February 2026, China is the only country with an early-stage multi-satellite AI compute network operating in orbit. The Three-Body Constellation’s 12 satellites provide a claimed combined 5 peta operations per second of compute, connected by real-time inter-satellite laser links. For rough context, El Capitan at Lawrence Livermore delivers 1.72 petaFLOPS (FLOPS means floating-point operations per second, indicates computer performance) — though comparing these figures directly is approximate, since operations per second and floating-point operations per second measure different things depending on precision format. Alibaba’s Qwen-3 ran inference tasks on that network in November 2025, publicly confirmed in January 2026. The second and third satellite clusters launch in 2026. The target is 100 satellites by 2027.

This is real. It is operational.

No US player has matched it at the constellation level.

Here is the number that China-leads narratives consistently omit. Starcloud’s H100 GPU delivers approximately 3,958 TOPS (trillions of operations per second) at peak sparse FP8 (8-bit floating point) compute. Each Three-Body satellite delivers 744 TOPS — though the specific precision format behind that figure is not fully disclosed in public reporting. Using peak figures for both, one Starcloud satellite carries roughly five times the raw compute of one Three-Body satellite. Even adjusting for different precision formats, the per-satellite compute gap is substantial — and the H100 represents current commercial AI chip generation, while China’s accelerators are domestically developed alternatives with limited independent benchmarking.

China has more satellites. The US has more compute per satellite. These are different products solving different problems.

SpaceX’s Starship is the only launch vehicle currently on a credible trajectory to break the $200/kg threshold that makes orbital data centers commercially viable at scale. China’s ZhuQue-3 — its first reusable methane-fueled rocket, launched December 3, 2025 — reached orbit on its maiden flight but failed to recover its first stage, which crashed near the landing pad. A genuine technical milestone, but years behind Starship in payload capacity, launch cadence, and booster reuse. Whoever controls the launch ramp controls the deployment economics for everyone, including themselves.

Starcloud, Axiom Space, Google, Blue Origin, SpaceX — independent commercial actors competing and innovating across a well-capitalized ecosystem. China’s players are largely state-directed programs with commercial structures. State backing ensures execution certainty. It does not guarantee the kind of edge-case innovation that changes the architecture of the solution.

Raw compute metrics and launch economics tell one story. There is another story that those metrics do not capture — and it is the reason China’s orbital compute program deserves attention beyond the satellite count.

China is not building this infrastructure because space is a better place for servers. The strategic logic — stated explicitly by Chinese aerospace officials and analysts — is that space offers a processing environment where sensitive national data never has to transit infrastructure outside China’s direct control.

Consider the architecture of most satellite intelligence today: raw data collected in orbit beams to ground stations, routes through terrestrial networks to processing centers, and then back out. Each hop in that chain is a potential point of foreign jurisdictional exposure. An orbital compute network that processes data entirely in orbit and transmits only finished results to Chinese-controlled ground stations eliminates those hops by design. Whether China’s current constellations achieve this in practice is a separate question — but it is clearly the architecture they are building toward, and it has no equivalent in the US commercial orbital compute programs, which are built around commercial returns, not national data control.

ADASpace’s confirmed roadmap makes this explicit: 2,400 inference satellites and 400 training satellites by 2035, spanning multiple orbit types, targeting 100,000 petaflops of inference compute and 1 million petaflops of training compute.

When ADASpace’s full constellation reaches 2,800 satellites, the projected combined compute is 1,000 peta operations per second. Compared to El Capitan’s 1.72 petaFLOPS, the headline figure often cited is “580 times more powerful” — but this comparison is approximate, since both figures use different precision formats and measurement methodologies. The honest read is that the scale gap is enormous regardless of how you normalize the units. That number is also a 2035 target, not a current reality. Today’s 12 satellites deliver roughly 5 POPS (parallel operations per second) claimed — meaningful as a proof of concept, not yet transformational at scale.

China has put the first constellation in the sky. The US has put the most powerful chip in space. Both are working toward a future where the 2035 targets actually ship.

In January 2026, China filed proposals for approximately 200,000 satellites. SpaceX, the same month, filed for up to one million.

These are not satellites in orbit. They are claims on orbital slots and radio spectrum — finite resources governed by the International Telecommunication Union on a first-come, first-served basis.

The technology race and the orbital real estate race are happening simultaneously. The governance frameworks to manage congestion, debris liability, and data jurisdiction in orbit are not keeping pace with either.

The satellites going up now will define the orbital infrastructure landscape for decades. The rules governing how that infrastructure operates — who can access which orbits, whose data sovereignty claims apply in space, who is liable when a satellite collision creates a debris field — are being written by absence. The actors moving fastest are setting the defaults.

This is not unique to China. It applies equally to SpaceX’s one-million-satellite ambition. But it is worth naming clearly: the most consequential decisions about orbital compute governance are not being made in policy discussions. They are being made in FCC filings and launch schedules. I haven’t spent enough time reading about this (since this is just an extended topic), but let me know if you want me to.

The orbital compute market is projected to reach approximately $39 billion (sounds small, but remember it is near zero today). Analysts estimate commercial viability by 2030. The realistic trajectory:

Proof of concept complete. Starcloud demonstrated an H100 in orbit. Axiom Space nodes deploy. Google and ESA launch test satellites. China’s constellation expands past 100 satellites. BAIST validates thermal management.

The question shifts from “can this work” to “what does it cost at scale”.

Starship cadence increases. Per-kilogram launch costs compress toward $200/kg. First commercial revenue from orbital compute services. ADASpace targets its 1,000-satellite network. The hardware obsolescence problem either finds a solution or begins killing business models.

Large constellations operational for specific high-value workloads — Earth observation, sovereign AI, financial routing. China’s gigawatt-class network approaches commercial parity for its target use cases. The governance question gets answered, one way or another, by whoever has the most satellites already in orbit.

The race to put AI in orbit is real, it is funded, and it has hardware flying.

What it does not have is a clear winner — because the two leading players are winning on different dimensions.

China has the most satellites in orbit performing actual AI workloads today. The US has the most powerful chip in space and the launch economics advantage that will ultimately determine who can deploy at scale. Both have 2035 targets that dwarf anything currently operational. Neither has proven the full business case.

What China has that no Western competitor has built is an orbital compute program with data sovereignty as a structural design goal — not a marketing claim, but an architecture choice for themselves. That matters differently depending on what you think the dominant use case for orbital AI will be.

If it is commercial cloud services and large-scale AI training, the US ecosystem wins on compute density and capital. If it is sovereign geospatial intelligence and national data infrastructure, China has a head start it is building on deliberately.

The cloud went above the clouds. Both sides are building the infrastructure. The question of who owns what — and by what rules — has barely been asked yet.

Zero Address covers Chinese technology for English readers. I read the Mandarin tech docs so you don’t have to.

If this reframed the story for you — share it with one person who is still reading the headline version. And of course, subscribe to receive free articles like this every week.