📡 Free-Space Optical Link (FSOC) / Inter-Satellite Link (ISL)

The use of lasers to transmit data through the vacuum of space. ISLs allow satellites in a constellation to communicate with each other at terabit-per-second speeds without routing traffic through ground stations. This is crucial for linking individual compute nodes in orbit.

  • Provides massive bandwidth compared to RF (radio frequency)
  • Enables interconnected cluster computing in space
  • Crucial for data center topology

⚡ Power Usage Effectiveness (PUE)

A ratio that describes how efficiently a computer data center uses energy. It is determined by dividing the total amount of power entering a data center by the power used to run the IT equipment. The ideal PUE is 1.0. Space data centers utilizing passive radiative cooling aim for PUEs approaching 1.05.

  • Total Facility Power / IT Equipment Energy
  • Earth data centers often sit between 1.2 and 1.5
  • Lower PUE means less wasted energy on cooling overhead

❄️ Radiative Cooling

The process by which a body loses heat by thermal radiation. In the vacuum of space, convection (air cooling) and conduction (liquid cooling requiring massive Earth infrastructure) are not viable. Space data centers must reject all heat generated by processors into the deep cold of space via large radiator panels according to the Stefan-Boltzmann law.

  • The only primary method of rejecting heat in a vacuum
  • Requires large surface area panels pointed away from the sun
  • Deep space provides a near absolute zero (~3 Kelvin) heat sink

🛡️ Radiation Hardening / Tolerance

Designing electronic components and systems to be resistant to damage or malfunction caused by ionizing radiation (like cosmic rays or solar flares) found in space. This involves physical shielding, specialized manufacturing processes (silicon-on-insulator), and logical redundancy (like Triple Modular Redundancy - TMR).

  • Protects against Single-Event Upsets (SEUs)
  • Mitigates Total Ionizing Dose (TID) degradation over time
  • Commercial Off-The-Shelf (COTS) parts like modern GPUs often require 'tolerance' via software redundancy rather than full physical hardening to remain cost-effective

☀️ Sun-Synchronous Orbit (SSO)

A particular kind of near-polar Low Earth Orbit (LEO) where a satellite passes over any given point of the Earth’s surface at the same local mean solar time. More importantly for space data centers, specific ‘dawn/dusk’ SSOs allow a satellite to remain almost continuously in sunlight, providing 24/7 uninterrupted solar power.

  • Eliminates the 'eclipse' periods common in standard LEO
  • Maximizes solar array efficiency and minimizes battery storage requirements
  • Typically altitudes between 600-800 km

🌐 Low / Medium / Geostationary Earth Orbit (LEO / MEO / GEO)

The three primary orbital regimes used by satellites. LEO (160–2,000 km) offers low latency (~20–40 ms) and is where most space data center proposals are focused. MEO (2,000–35,786 km) is used by GPS constellations. GEO (35,786 km) provides fixed coverage of one hemisphere but adds ~600 ms round-trip latency — too slow for most compute workloads.

  • LEO: lowest latency, highest orbital drag, requires large constellations for continuous coverage
  • MEO: used by GPS/navigation; intermediate latency
  • GEO: single satellite can cover a continent; impractical for interactive compute due to latency
  • Most orbital data center proposals target LEO 500–1,200 km

🔧 Commercial Off-The-Shelf (COTS)

Hardware designed for terrestrial use that is deployed in space without full radiation hardening. Using COTS components (like NVIDIA H100 GPUs) dramatically reduces cost compared to traditional space-grade chips, but requires software-level redundancy to tolerate occasional radiation-induced errors. Starcloud’s first orbital AI data center runs COTS H100 GPUs.

  • 10–100x cheaper than radiation-hardened equivalents
  • Requires error correction and redundancy in software (ECC, TMR)
  • Feasible in LEO where radiation exposure is lower than deep space
  • Starcloud's H100 satellites demonstrated COTS viability for AI workloads in 2025

🔁 Triple Modular Redundancy (TMR)

A fault-tolerance technique where three identical circuits perform the same computation in parallel and a majority-vote circuit selects the correct output if one unit produces an error. In space, TMR is used to protect against Single-Event Upsets (SEUs) — bit flips caused by cosmic rays hitting a semiconductor.

  • Detects and corrects single-component failures automatically
  • Used in safety-critical systems: spacecraft, aviation, nuclear
  • Increases hardware cost and power draw by ~3x
  • Often combined with ECC memory for full-stack radiation tolerance

🧮 Error-Correcting Code Memory (ECC)

A type of RAM that detects and automatically corrects the most common data corruption errors — including single-bit flips caused by radiation (SEUs). ECC memory adds a small amount of extra storage per byte to store parity/checksum data. It is standard in server-grade hardware and essential for any compute deployed in the radiation environment of space.

  • Corrects single-bit errors; detects (but cannot correct) double-bit errors
  • Standard in all server-grade hardware (Xeon, EPYC, A100/H100)
  • Critical in space: cosmic ray flux in LEO causes roughly 1 SEU per GB of DRAM per day
  • ECC adds ~12.5% memory overhead (e.g., 9 chips per 8-chip bank)

📜 Power Purchase Agreement (PPA)

A long-term contract between an electricity generator and a buyer (usually a data center operator) that locks in electricity prices for 10–20 years. On Earth, hyperscalers like Google, Microsoft, and Amazon use PPAs to secure renewable energy at predictable costs. In space, there is no PPA equivalent — solar energy is free and unlimited, which is one of the core economic arguments for orbital data centers.

  • Typical PPA duration: 10–20 years at a fixed or indexed price
  • Hyperscalers are the largest corporate buyers of renewable PPAs globally
  • Space data centers eliminate the need for PPAs entirely — sunlight in LEO is unmetered
  • Electricity is ~40–50% of the lifetime cost of a terrestrial data center

🚀 Starship

SpaceX’s fully-reusable two-stage heavy-lift launch vehicle, designed to bring payload costs below $100/kg to LEO. Starship is the single most important enabler for the space data center industry: current launch costs (~$1,500–2,000/kg on Falcon 9) make orbital compute ~3–10x more expensive than terrestrial equivalents. Most economic models show space data centers becoming cost-competitive only once Starship reaches full reusability at target cadence.

  • Target payload: 100–150 tonnes to LEO per flight
  • Full reusability goal: <$100/kg, down from ~$1,500/kg today
  • Currently in flight testing as of 2025–2026
  • The SpaceX-xAI merger (Feb 2026) is predicated on Starship economics becoming viable by 2028

🔧 NVIDIA Compute Modules and Platforms

NVIDIA distinguishes between compute modules (compact SWaP-optimized hardware units like the Space-1 Vera Rubin Module) and platforms (complete edge AI solutions including the module plus JetPack SDK and ecosystem, such as IGX Thor and Jetson Orin).

  • Module: core hardware (GPU/CPU/memory) for spacecraft integration
  • Platform: module + SDK/tools for full deployment
  • Space-1 Vera Rubin Module: 25x AI compute vs H100
  • IGX Thor/Jetson Orin platforms enable orbital edge inference
  • Key for orbital data centers and autonomous space operations