Orbital Compute

The thesis

Why Compute in Orbit: The Case for Orbital Data Centers

Power, cooling, and land are scarce on Earth. They are not scarce above the atmosphere.

The argument for orbital compute rests on three physical facts. First, a satellite in sun-synchronous orbit receives near-continuous sunlight — more than 99% of each orbit — with no clouds, no weather, no nighttime, and no grid interconnection. Second, in vacuum, waste heat radiates directly to the 2.7 K cosmic background without water, refrigerant, or evaporative cooling. Third, orbital space is effectively unlimited: there is no zoning board, no grid queue, no water right, and no utility interconnection delay. For AI workloads whose growth is limited primarily by power, cooling, and permitting, orbit removes all three constraints simultaneously.

The trade-off is launch cost. Every kilogram of compute hardware must be lifted from Earth’s surface. At current Falcon 9 rideshare pricing (~$7,000/kg), orbital data centers are economically unviable at scale. Google’s own analysis pins the equilibrium at roughly $200/kg — a 35× reduction from today. That gap is why every serious orbital compute program is contingent on Starship achieving its designed economics. Sam Altman called orbital data centers “ridiculous for now.” The “for now” is the operative phrase.

The programs

Orbital Data Center Programs Filed or Announced

China flew first. The US programs are larger on paper. Both are real.

SpaceX — Starmind. Filed with the FCC on January 30, 2026, for a constellation of up to one million satellites operating as orbital data centers. Altitudes from 500 to 2,000 km in sun-synchronous and 30-degree inclination orbits. The first-generation satellite, AI1, was unveiled June 8, 2026: 20 meters tall, 75-meter deployed wingspan (solar arrays), 30-meter radiative cooling panels, peak compute of 250 kW per satellite. Mass exceeds two tons. The AI1 compute payload is powered by Nvidia Vera Rubin NVL72. Two prototypes are targeted for early 2027, volume production at the Gigasat facility in Bastrop, Texas targeted for late 2027. The Starmind constellation routes data through the existing Starlink laser mesh to ground stations. SpaceX requested a waiver of the standard FCC 6-year/9-year deployment milestones.

Blue Origin — Project Sunrise. Filed with the FCC on March 19, 2026, for up to 51,600 satellites in sun-synchronous orbits at 500–1,800 km. Each orbital plane carries 300–1,000 satellites, with planes separated by 5–10 km in altitude. Optical inter-satellite links route data through Blue Origin’s planned TeraWave broadband constellation (5,408 satellites, announced January 2026). No satellite specifications, compute capacity, timeline, or cost estimates disclosed. Enabled by New Glenn launch capacity.

Google — Project Suncatcher. Announced by CEO Sundar Pichai in late 2025. A research initiative to operate solar-powered AI compute clusters in orbit using Google’s Trillium TPU v6e chips. The concept envisions 81-satellite formation-flying clusters spanning a 1 km radius, connected by free-space optical inter-satellite links. Planet Labs is the satellite manufacturer. Two prototype satellites targeted for early 2027. Google has confirmed the TPU v6e survived simulated five-year LEO radiation exposure in a particle accelerator. Lab demonstrations achieved 1.6 Tbps on a single optical transceiver pair. Google is in talks with SpaceX for launch services, owns approximately 6.1% of SpaceX. Google is also paying SpaceX $920 million per month to lease terrestrial Colossus 2 data center capacity for AI training and inference.

Starcloud. Filed with the FCC on February 3, 2026, for an 88,000-satellite constellation. Y Combinator-backed startup (formerly Lumen Orbit). Raised $250 million total including a $170M Series A at $1.1B valuation in March 2026 — the fastest YC company to reach unicorn status. Starcloud-1 (60 kg, launched November 2, 2025) carried the first Nvidia H100 GPU to orbit and ran the first LLM training in space (NanoGPT on Shakespeare). Starcloud-2, with multiple H100s plus Blackwell hardware and 8 kW compute capacity, targets October 2026 launch. Plans to deploy first AWS Outposts hardware in space. Envisions a 5 GW orbital hypercluster.

China — multiple programs. China was the first nation to deploy operational space computing hardware and has the most programs in flight. Zhongke Tiansuan (Comospace), founded in 2024, has logged over 1,000 days of on-orbit operation with its Aurora 1000 space computer aboard a Jilin-1 satellite; its next-generation Aurora 5000, featuring a domestic high-performance GPU, is set for in-orbit trials. ADA Space and Zhejiang Lab launched 12 satellites for the Three-Body edge-computing constellation in May 2025 — each carrying an 8-billion-parameter AI model at 744 TOPS, connected by 100 Gbps laser ISLs, forming the first stage of a planned 2,800-satellite network. Beijing’s Astro-future Institute is building a centralized space data center in dawn-dusk orbit (700–800 km) targeting 1 GW capacity, with the Chenguang-1 tech-demo satellite planned for 2026. In June 2026, Beijing launched a Space Computing Industry Innovation Center to unify multiple firms. Orbital Chenguang, a separate startup, received $8.4 billion in Chinese state credit lines (April 2026) for a phased orbital data center constellation. CASC’s 15th Five-Year Plan proposes gigawatt-scale space-based computing infrastructure. Shanghai Bailing Aerospace is working toward 100 kW-class computing satellites. The Chinese approach is coordinated across state, municipal, and commercial entities in a way the Western programs are not.

FCC filings represent requested authorization, not approved deployment. Amazon has petitioned to block SpaceX’s Starmind filing; SpaceX has responded; the proceeding is ongoing. None of these constellations have received FCC authorization to deploy. The total filed satellite count across all ODC programs exceeds 1.3 million.

The silicon

Orbital-Grade AI Compute Hardware

Three chip families are targeting orbit. Only one has flown.

Nvidia Space Computing. Announced at GTC 2026 (March 16), Nvidia’s space product line has three tiers. The Space-1 Vera Rubin Module is the flagship — a flight-rated package delivering up to 25× the AI compute of an H100, designed for orbital data centers running LLMs and foundation models. Below it, IGX Thor targets mission-critical edge processing with functional safety. Jetson Orin handles real-time vision, navigation, and sensor data onboard satellites — Kepler Communications has already deployed 40 Jetson Orin modules across 10 satellites. The Vera Rubin Space-1 has no public ship date, pricing, or published radiation tolerance rating. SpaceX has confirmed that the AI1 Starmind satellite compute payload uses Nvidia Vera Rubin NVL72.

Google Trillium TPU v6e. Google’s custom tensor processing unit, tested in a particle accelerator simulating five years of LEO radiation exposure. Survived without damage. The TPU is purpose-built for Google’s AI workloads and would fly exclusively in the Suncatcher constellation. No standalone market offering.

SpaceX / Tesla D3. Terafab, a SpaceX-associated semiconductor facility under construction at Gibbons Creek, Texas, targets a radiation-qualified 2 nm process node for orbital AI accelerators. The D3 chip is designed for the AI1 platform with an interchangeable compute module that can accept processors from multiple vendors. D3’s timeline is gated by fab construction, radiation qualification, and yield optimization — an optimistic first operational date is 2030. In the interim, AI1 satellites will fly Nvidia silicon.

The constraint triangle

Power, Thermal, and Radiation: The Three Engineering Problems

Every watt of compute is a watt of heat, and in vacuum the only way to dissipate heat is radiation.

Power. The AI1 satellite carries a 150 kW deployable solar array operating at approximately 250 W/m², sized to deliver peak compute capacity directly from solar power. Sun-synchronous orbits at higher altitudes maintain sunlight more than 99% of each orbit, eliminating the eclipse problem for baseload AI training. Lower-inclination 30-degree orbits face ~30 minutes of eclipse per 90-minute orbit and carry onboard batteries to bridge shadow passes or run interruptible workloads.

Thermal. The AI1 design carries a 110 m² deployable liquid radiator system with redundant pumping loops and micrometeoroid shielding — approaching the radiator area of the International Space Station on a single satellite. Without it, the compute payload would overheat within minutes at full power. The radiator and the solar array are most of what an orbital compute satellite physically is; the compute module itself is small. LEO thermal cycling — 16 sunrise-sunset transitions per day, each imposing rapid temperature swings — accumulates tens of thousands of fatigue cycles over a 5–10 year operational life.

Radiation. LEO is the easiest orbital regime for electronics — Earth’s magnetic field still deflects most galactic cosmic radiation and solar particle flux. But advanced logic at 2 nm is dominated by single-event upsets from energetic particles striking transistor structures small enough that one ionization track can flip a bit. Analog devices suffer cumulative dose degradation. Power electronics face displacement damage. Memory, sensors, optics, and photovoltaics each degrade on their own schedules. A space-qualified compute platform is not a hardened chip — it is a hardened supply chain across accelerators, power management, memory, networking, sensors, and packaging, all qualified for the operating regime.

The economics gap

Launch Cost and the $200/kg Threshold

The gap between here and viability is 35×.

Google’s analysis identifies approximately $200/kg to LEO as the financial equilibrium for orbital data centers. SpaceX’s current rideshare pricing is ~$7,000/kg. The gap is real. Every orbital compute program is a bet that Starship’s cost reduction trajectory closes it within the 3–5 year window the prototype programs use to validate architecture before committing to full constellation deployment.

Starship’s target of sub-$100/kg would make orbital compute economically viable — but that target assumes full and rapid reuse of both stages at high cadence, which has not been demonstrated. The current state: 13 Starship test flights, booster catches demonstrated, ship reentry under iteration, first Starlink V3 payload deployed July 2026, Flight 14 (first targeted orbital flight) NET September 2026. The dependency is explicit and acknowledged by every program: without Starship reaching its designed cadence and economics, orbital compute remains a compelling architecture whose economics do not close.

What is already in evidence: SpaceX (via xAI) has leased the Colossus 1 terrestrial data center in Memphis to Anthropic for $1.25 billion per month, and Colossus 2 capacity to Google for $920 million per month. Revenue relationships are assembling ahead of the orbital infrastructure — customers before deployment, commercial agreements before operational capability.

The startups

Orbital Compute Startups and Early Hardware

Eight organizations filed plans, launched hardware, or committed funding within 90 days in early 2026.

Starcloud has the strongest claim to operational hardware. Its Starcloud-1 satellite (November 2025) ran the first H100 in orbit, trained the first LLM in space, and ran inference on Google’s Gemma model and Capella Space SAR imagery. Starcloud-2 (October 2026) carries multiple H100s plus Blackwell hardware at 100× the power generation of the first satellite. Raised $250M total at a $1.1B valuation. The company plans to deploy the first AWS Outposts hardware in space and has an FCC filing for 88,000 satellites.

Aetherflux (now Cowboy Space Corporation), founded by Robinhood co-founder Baiju Bhatt, raised at a reported $2B valuation in late March 2026. Targets a first commercial compute node in Q1 2027 combining solar-power beaming (infrared laser to terrestrial solar farms) with on-orbit AI inference. Uses Nvidia Space-1 Vera Rubin Module. SpaceX Falcon 9 or Starship launches.

Axiom Space deployed the first dedicated free-flyer orbital data center nodes in January 2026 with 2.5 Gbps optical links, running compute payloads on the ISS and on standalone platforms. Kepler Communications operates the world’s first commercially operational optical data relay network, with 40 Nvidia Jetson Orin modules across 10 satellites. Lonestar Data Holdings sent a mini data center with 8 TB storage to the Moon aboard an Intuitive Machines lander.

The skeptics

The Case Against Orbital Data Centers

The critics are not uninformed.

OpenAI CEO Sam Altman called orbital data centers “ridiculous for now.” AWS CEO Matt Garman has expressed skepticism. Short seller Jim Chanos has cited them. Gartner analysts have questioned the economics. The objections are not trivial:

Bandwidth. An orbital data center must downlink results to Earth. Even at 1 Tbps per optical link, the aggregate bandwidth of a million-satellite constellation serving billions of inference requests is an unsolved engineering problem. Terrestrial data centers sit directly on fiber backbones.

Latency. LEO round-trip latency is 5–15 ms — comparable to terrestrial inter-datacenter links but worse than intra-datacenter GPU-to-GPU latency. Distributed AI training across satellites faces coordination overhead that ground-based clusters do not.

Serviceability. A terrestrial server that fails can be replaced by a technician in minutes. An orbital satellite that fails is debris. Satellite lifetimes of 5–10 years mean the entire constellation must be continuously replaced, creating a perpetual launch demand that compounds cost.

Debris. Over 1.3 million satellites have been filed for across all ODC programs. The current total of all active satellites in orbit is roughly 16,000. The collision risk, spectrum coordination, and orbital congestion implications are without precedent. See Orbital Congestion.

The honest framing: orbital compute is a real engineering program being pursued by serious operators with real capital. It is also entirely contingent on a launch cost reduction that has not been achieved. Both are true. The prototypes flying in 2026–2027 will validate the architecture. The economics wait on Starship.