Last updated: 29 May 2026
Purpose: Track the development of orbital AI data centres as a proposed solution to terrestrial compute, power, and cooling constraints — and assess the investment implications for AI infrastructure, chip architecture, and terrestrial capex.
What belongs here:
What does NOT belong here:
companies/<name>/
notesSourcing standard: Every datapoint traces to a primary source (filings, transcripts, investor commentary, engineering publications). Promotional claims from vendors and early-stage companies are flagged. Wikilinks for internal cross-references; inline URLs for external sources.
Maintenance: Update as programme milestones, launch cost data, or new market entrants emerge. Date-stamp additions. Review when a major player announces a concrete programme.
Terrestrial AI data centre expansion is running into three compounding constraints: power grid capacity and permitting timelines (often 4–7 years to new grid connections), cooling density limits in hot or water-stressed regions, and the cost and land footprint of building at the scale hyperscalers now require. See capex-infrastructure-and-power for the current capex and power demand picture.
Against this backdrop, a small number of investors and technologists — most prominently Gavin Baker (Atreides Management), and subsequently Elon Musk via SpaceX/xAI — have argued that Low Earth Orbit (LEO) is a structurally superior environment for AI compute. The argument is not new (space-based computing has been discussed for decades) but has become more investable as launch costs have fallen dramatically with Falcon 9 reusability, and Starship threatens to fall further still.
As of May 2026, this has crossed from thought experiment into early-stage capital deployment: multiple startups and at least one hyperscaler are publicly pursuing orbital data centre programmes.
Solar irradiance in LEO is approximately 1,361 W/m² — roughly 30% more intense than at Earth's surface, with no atmospheric absorption or scattering. More importantly, in a sun-synchronous or other optimised orbit, a satellite can remain in continuous sunlight for extended periods, eliminating the day/night cycling that forces terrestrial solar installations to carry large battery reserves.
The investment implication: orbital data centres are not grid-constrained. The terrestrial bottleneck of substation queues, transmission build-out, and local permitting simply does not exist. This is the most compelling part of Baker's argument from an infrastructure scarcity perspective.
The vacuum of space eliminates convective cooling, but provides a radiative heat sink at approximately 3 Kelvin (−270°C) on the shadow-facing side. A radiator mounted on the dark face of a satellite can passively reject heat without water, refrigerants, or energy-intensive compressors.
Terrestrial data centres spend heavily on cooling — water, mechanical systems, and the energy to run them. In principle, passive radiative cooling eliminates this opex category.
Light travels approximately 47% faster through vacuum than through glass (due to the refractive index of silica). Inter-satellite laser links (ISLs) — which SpaceX already uses commercially on Starlink — communicate at the vacuum speed of light. A dense constellation networked by ISLs would have lower latency and higher throughput per link than terrestrial long-haul fibre.
For distributed inference serving global users, this could meaningfully improve response latency, particularly for cross-continental requests. The cooling advantage is real in principle but inverted in practice when applied to dense compute. Radiative cooling power scales with the surface area of the radiator. As compute density increases — packing more GPU-equivalent silicon into a given volume — the surface-area-to-volume ratio worsens. The practical result is that high-density AI clusters (100+ kW per rack on Earth) cannot simply be moved to orbit; the radiating surface required to reject that heat at orbital temperatures would be impractically large relative to the compute payload.
This forces orbital compute toward distributed, low-density architectures — many small satellites rather than a few dense clusters. That is structurally the opposite of the densification trend in terrestrial frontier AI training. Engineers from Varda Space Industries and others have noted that while the physics does not rule it out, achieving compute density comparable to terrestrial facilities requires radiator engineering that is not yet demonstrated at scale.
Commercial-off-the-shelf (COTS) silicon in LEO is exposed to ionising radiation — trapped electrons, protons, and cosmic ray primaries. Traditional radiation-hardened (rad-hard) chips are ruggedised against this environment but are typically an order of magnitude or more less capable per watt than leading-edge terrestrial silicon — with the gap widening considerably as AI accelerator performance has compounded. If orbital compute must use rad-hard chips, the compute-per-kg advantage largely disappears.
The alternative is COTS chips with redundancy and error correction (SpaceX does this on Starlink). Starcloud's November 2025 launch flew an NVIDIA H100 — a fully commercial, non-rad-hard chip — relying on this approach rather than hardened silicon. Early operational reports suggest it functioned (running NanoGPT and a version of Google Gemini in orbit), but long-run degradation rates in the LEO radiation environment at H100-class densities are not yet known. This is viable for some workloads but increases failure rates and complicates cluster management. There is no demonstrated multi-year solution at AI accelerator densities as of mid-2026.
Training frontier models requires tight, low-latency synchronisation across thousands of accelerators — sub-microsecond collective operations within a NVLink or equivalent fabric. Coordinating a training cluster across dozens of satellites in motion, even with ISLs, introduces synchronisation latency that simply does not exist in a terrestrial NVLink fabric. Orbital training is not obviously viable in the near term.
Inference is more tractable. Individual inference requests are parallel and largely independent — there is no collective synchronisation requirement analogous to training. The near-term orbital compute opportunity is inference, not training. This is a materially narrower TAM than the bull case often implies.
Terrestrial data centres benefit from cheap, fast hardware replacement — a failed GPU is swapped in hours. Orbital hardware is either non-serviceable (most LEO satellites) or extremely expensive to service (requires a crewed mission or robotic servicing). Failure rates in the radiation environment are higher than on the ground, and the inability to cheaply replace failed hardware increases the effective hardware cost per compute-hour.
The orbital compute investment case reduces almost entirely to a single variable: launch cost per kilogram to LEO.
| Vehicle | Approximate cost/kg to LEO | Status (May 2026) |
|---|---|---|
| Falcon 9 (reusable) | ~$3,000–4,500/kg | Operational at scale |
| Falcon Heavy | ~$1,000–1,500/kg | Operational |
| Starship (target) | ~$65–200/kg | Development; not yet operational at cadence |
| Starship (aspirational) | ~$10–20/kg | SpaceX's long-run aspiration; not near-term | At Falcon 9 prices (~$3,600/kg per TechCrunch's Feb 2026 analysis), orbital data centres are unambiguously uneconomic vs terrestrial colocation. SpaceX itself has stated Starship targets in the range of $10M per flight at 100–150 tonnes to LEO — implying ~$65–100/kg at gross flight cost. A research fellow at the European Space Policy Institute has stated that a $10M/flight Starship is "unrealistic in the near-term" and that most orbital data centre cost models assume Starship economics that have not yet been demonstrated. The economic path to cost parity with terrestrial colocation for most workloads is generally projected in the 2029–2031 window, with significant uncertainty. See launch market structure for Starship development status.
| Entity | Programme | Status | Notes |
|---|---|---|---|
| Starcloud (Redmond, WA, NVIDIA-backed) | Orbital AI data centre constellation | Operational (proof-of-concept) |
Starcloud-1 launched Nov 2025 with NVIDIA H100 GPU; first LLM trained in space (NanoGPT on Shakespeare corpus); first satellite to run a version of Google Gemini in orbit. Raised $170M at $1.1B valuation. FCC filing for 88,000 satellites. Next satellite (Starcloud-2) targeting Oct 2026 with Blackwell GPU. | | Kepler Communications | Compute cluster (LEO) | Operational | Launched largest orbital compute cluster as of early 2026: 40 Nvidia Jetson Orin processors across 10 satellites linked by laser communications. 18 paying customers. Commissioned March 2026. Inference workloads; not frontier-AI class. | | Axiom Space | Orbital data centre nodes | Deployed | First two orbital data centre nodes launched to LEO 11 January 2026. Targeting cloud computing, AI/ML, and edge processing applications. | | Orbital (LA startup, a16z-backed) | Orbital AI inference constellation | Early stage | Emerged from stealth April 2026; targeting launch 2027, manufacturing facility 2028. | | SpaceX / xAI | Orbital data centre / merged entity | Early stage | Musk has signalled merger of xAI compute needs with SpaceX orbital infra; FCC filing for up to 1M data centre satellites. | | Google / Project Suncatcher | 2-satellite TPU prototype (Planet Labs); 81-satellite cluster long-term vision | Announced / Development | Google announced Project Suncatcher Nov 2025; 2 prototype satellites with TPUs launching with Planet Labs by early 2027. Long-run design envisions 81-satellite clusters in dawn-dusk SSO. Separately reported (May 2026, WSJ) to be in talks with SpaceX on Starship launch for larger constellation. |
Sources: Tom's Hardware on Google/SpaceX talks, IEEE Spectrum on Orbital, TechCrunch on orbital economics Note: February 2026 has been reported as the first month in which orbital data centre operators ran production workloads in space, though at very small scale. These are proof-of-concept deployments, not commercial-scale operations.
At current launch costs and chip radiation-hardening constraints, orbital compute is not a near-term threat to terrestrial AI infrastructure. The economics are 3–20× worse than terrestrial colocation for most workloads. The primary value of monitoring this theme now is tracking whether Starship achieves its cost targets and whether chip vendors develop COTS AI silicon with sufficient radiation tolerance.
If Starship achieves $65–100/kg at operational cadence, orbital inference nodes for latency-sensitive global applications become economically plausible. This is a real optionality scenario rather than a base case. Key monitorables: SpaceX Starship flight cadence and turnaround economics; first announced commercial orbital inference SLAs; chip vendor radiation-tolerance roadmaps. Orbital compute at scale would likely favour custom ASICs over COTS GPUs. The radiation environment, power-per-watt constraints, and low-density thermal architecture all push toward purpose-built silicon optimised for space. If orbital inference scales, it is a potential long-term headwind for GPU incumbents and a potential opportunity for custom ASIC vendors (Broadcom, Marvell, and hyperscaler-internal teams). This is speculative at this stage.
The bull case for orbital compute includes a scenario where orbital inference crowds out terrestrial data centre demand growth post-2030. This is not close to being investable as a short thesis against data centre REITs, power equipment (Vertiv, Eaton), or hyperscaler capex — the timeline is too uncertain and the near-term constraints too binding. It is worth flagging as a tail risk in longer-duration infrastructure theses.
Bull case (Baker, Musk): Terrestrial compute is gated by power and land; space removes both constraints permanently. Once Starship makes launch cheap, the structural cost advantages of solar + passive cooling + vacuum optics dominate. Orbit is where frontier compute eventually lives.
Bear case (most engineers and infrastructure analysts): Thermal dissipation density, radiation hardening, maintenance costs, and coherent cluster synchronisation make the orbital environment deeply hostile to the compute architectures that matter for AI. The economics require Starship assumptions that are aspirational. Terrestrial data centres will continue to solve their constraints via grid diversification, liquid cooling advances, and nuclear.
The honest answer: the physics is not wrong on either side — the debate is about which engineering constraints prove more tractable over a 5–10 year horizon.
Starship cadence and cost per flight: The single most important variable. Track FCC filings, launch manifests, and SpaceX's stated turnaround economics.
Radiation-tolerant COTS AI silicon: Are any chip vendors publishing roadmaps for LEO-tolerant AI accelerators without full rad-hard cost penalties?
Cooling density demonstrations: Has any orbital programme demonstrated AI-relevant compute density (>10 kW/m²) with passive radiative cooling?
Google Project Suncatcher: Now publicly confirmed (Nov 2025). Watch for the 2-satellite Planet Labs prototype launch (targeting early 2027) and whether the separately reported SpaceX Starship talks lead to a larger follow-on programme — that would establish a market reference point for orbital CaaS pricing.
Orbital vs terrestrial inference economics: First public cost-per-token benchmarks from orbital inference operators, when available.
Press Esc to close