Land-Based Compute, Low Earth Orbit Compute

 

Land-based compute costs are dominated by electricity (often 30-50%+ of TCO for AI workloads), power availability, land, cooling, and CapEx for facilities/hardware. AI data centers are highly energy-intensive, with racks drawing 50-150+ kW (vs. 10-15 kW traditional), and hyperscale facilities pushing into hundreds of MW or GW scale. Global data center electricity use was ~415 TWh in 2024 (~1.5% of world total) and is projected to more than double by 2030, driven by AI. Electricity Prices and Locations (Rough 2025-2026 Industrial/Business Rates)

• Lowest-cost regions (favorable for data centers): Parts of the Middle East/Gulf (e.g., Qatar ~$0.03-0.07/kWh subsidized), select Asian/African countries (Vietnam, some in China/India ~$0.05-0.10/kWh industrial), US states like North Dakota/Texas/Idaho (~$0.07-0.09/kWh commercial/industrial). Texas and Virginia host massive clusters due to land, policy, and (in Texas) abundant natural gas/renewables potential.
• Higher-cost: Europe (EU industrial often 2x US levels, e.g., Germany/Ireland >$0.20/kWh), California/Hawaii (high), with some data-center-heavy areas seeing rate hikes from grid strain.
• US average industrial: Around $0.08-0.09/kWh recently, but varying widely; data center growth is pushing local prices up in hotspots (e.g., Virginia spikes), though overall data-center states don't uniformly have higher prices than others.

Energy sources matter: Cheap natural gas (Texas), hydro (some Nordic/Quebec), nuclear, or PPAs with renewables + batteries help. On-site generation (gas turbines, emerging SMRs) and behind-the-meter power bypass grid queues (often 4+ years). Renewables + storage can cut effective costs vs. grid in some EMEA/APAC setups (up to 40% savings reported). Cooling (water/ evaporative or air) adds costs in hot/dry areas; PUE for efficient AI facilities can be ~1.1-1.2 but still requires significant infrastructure.
Build/OpEx examples: A 100 MW AI data center might cost billions in CapEx (hardware dominant, then power/cooling/land); annual electricity alone can be tens of millions at $0.05/kWh. Global scaling to meet AI demand is estimated in trillions (e.g., McKinsey ~$6.7T by 2030 for compute infrastructure). Constraints include grid interconnection delays, water use, local opposition, and permitting. Low-price locations (e.g., solar-rich deserts or gas regions) win, but transmission and political factors limit options.Space-Based (Orbital) Compute: Launch Costs and MaterialsHistorical launch costs to LEO:

• NASA Space Shuttle: ~$54,500/kg (or higher in some full-cost accounting; ~$1.5B per launch for ~27,500 kg payload).
• SpaceX Falcon 9: ~$2,720/kg (dramatic ~20x reduction via partial reusability). Falcon Heavy lower still on a per-kg basis for heavy payloads.

Current/projected:

• Falcon 9/Heavy effective: $1,000-3,000+/kg depending on configuration and customer.
• Starship (fully reusable, targeting 100-150t payloads): Goals of $50-200/kg near-term, with optimistic paths to $10-100/kg at high cadence (dozens-hundreds of flights/year). Analysts see $100-500/kg realistic in the 2026-2030 window as reusability, manufacturing scale, and cadence improve. Propellant is a tiny fraction; costs are dominated by vehicle amortization, operations, and refurbishment.

Factors driving costs down: Full reusability (both stages, rapid turnaround like aircraft), high flight rate (amortizes fixed costs), mass manufacturing (Raptor engines, stainless steel Starship), vertical integration, and learning/iteration. Starship's scale enables entirely new economics. Constraints: Technical risk (current test flights), regulatory (FAA, orbital traffic/debris), supply chain for engines/propellant, and achieving true airplane-like ops (minimal refurb, high reliability). Further drops likely over 5-10+ years with maturity, but not guaranteed to $10/kg soon—physics, materials, and ops set floors. Demand pull from satellites, Starlink, Moon/Mars, and yes, orbital compute will accelerate it.
Orbital datacenter materials and mass:

• Power: Solar arrays in orbit get ~1.36 kW/m² constant (no atmosphere/night in good orbits like dawn-dusk SSO), but need large deployable arrays. Efficiencies and specific power improving (e.g., ~10-150+ W/kg possible with advanced/thin-film; Starlink-like ~30-40 kg/kW system level in examples).
• Cooling: Vacuum enables radiative cooling (no convection), potentially efficient per area at high temps, but radiators add significant mass (e.g., several kg/m²; area often comparable to solar due to ~400 W/m² generation vs. heat rejection needs). Total power/thermal systems (solar + radiators + structure) can be 50-70%+ of mass. Chips/servers need radiation hardening.
• Overall mass/power: Early concepts tens-hundreds of kg/kW (optimistic future lower with scale/tech). For MW-GW scale: Enormous structures (km-scale arrays for GW), assembly (robotic?), redundancy. A 1 GW system could require millions of kg launched initially.

Rough economics: At $100-1,000/kg launch, the upfront mass penalty for solar/radiators/structure is huge vs. terrestrial (where power is "launched" via wires/fuel and cooling uses air/water). However, "free" 24/7 solar (no fuel) and radiative cooling could yield lower ongoing OpEx. Latency (round-trip to ground), bandwidth limits for data transfer, maintenance difficulty, and radiation are major drawbacks. Concepts exist (e.g., Starcloud, various studies), but currently niche/premium at best.
Comparison summary: Land-based wins today on cost, density, maintainability, and latency in most places with cheap power (e.g., $0.05/kWh Texas gas/renewables cluster beats orbital launch+mass penalties).
Orbital could compete or win on pure energy access if launch hits <$100-200/kg and mass-specific power/thermal improves dramatically—constant solar + vacuum cooling avoids Earth's grid/water/land bottlenecks. Total cost per FLOP or per kWh delivered would need detailed modeling; launch is only one part (hardware, comms, ops dominate too).
All options viable and lucrative for at least a decade (and safety favors multiples): AI/compute demand is exploding—projections show data centers driving large shares of electricity growth, with hyperscalers racing for GW-scale capacity. Grid, permitting, water, and local opposition create bottlenecks on Earth; not every region can or will host massive clusters. Land-based in low-price/energy-rich spots (US Texas, Middle East, etc.) scales fastest/cheapest now. Orbital (or other space) provides a parallel path:
unlimited "clean" solar without terrestrial infrastructure fights, diversification against ground risks (e.g., outages, regulation, climate), and potential for specialized/low-latency-edge or sovereign uses.
Explosive unmet demand (trillions in investment needed) means even higher-cost options succeed if they deliver reliable compute. Multiple vectors hedge risks—supply chain, geopolitics, single-point failures.
History shows Shuttle-to-Falcon 20x drop; Starship-like advances + demand will make orbital more attractive over time. Expect hybrid: mostly terrestrial for cost/density, growing orbital for marginal high-value or constrained capacity. Safety/reliability (diverse power/compute sources) and strategic autonomy further justify pursuing all viable paths. The bottleneck is energy and access at scale; innovation across ground/space will be rewarded handsomely for years.

Orbital thermal management is one of the most significant engineering bottlenecks for high-power applications like AI data centers in space, contrary to the intuitive notion that "space is cold." In vacuum, there is no convection or conduction to an atmosphere or fluid medium—heat rejection occurs almost exclusively through thermal radiation. This fundamentally limits cooling efficiency compared to Earth-based systems that use air, water, or other coolants. Fundamental Physics: Radiation-Only Heat Rejection

Key implications:

• Heat rejection scales with
  T^4
  , so small temperature increases dramatically boost capacity (e.g., doubling T multiplies power by 16). However, electronics (especially AI chips) have strict upper temperature limits (~20–80°C typical for reliable operation; higher temps increase leakage, errors, and failure rates exponentially).
• At ~20°C (293 K) with a two-sided radiator and good emissivity, rejection is roughly 600–800 W/m². For 1 MW of waste heat (common for even modest AI clusters), this requires ~1,200–1,600 m² of radiator area—roughly a hockey rink or several tennis courts.
• Solar arrays (needed for power) absorb heat and must be oriented differently from radiators, which need a clear view of cold space and minimal solar/albedo exposure. Backside use of arrays as radiators is limited.

Environmental complications in orbit:

• Thermal cycling: In LEO, ~90-minute orbits cause sunlight/eclipse transitions, leading to repeated expansion/contraction that stresses materials and structures. Sun-synchronous or higher orbits can mitigate this for more stable thermal environments.
• External heat loads: Solar flux (~1,360 W/m²), Earth albedo (reflected light), and infrared from Earth add absorbed heat that radiators must overcome.
• Microgravity: Affects fluid behavior in active cooling loops (no buoyancy-driven convection).

Mass and Scalability PenaltiesRadiators add substantial mass, which directly impacts launch costs. Examples:

• ISS External Active Thermal Control System rejects ~70 kW using radiator wings totaling ~7 metric tons. Scaling naively to 1 MW implies ~100 tons of radiators—versus ~10 tons for the compute hardware itself.
• Areal densities for current radiators: Often 5–12+ kg/m² (including structure, heat pipes, coatings). Advanced designs target <3–5 kg/m² with heat pipes, composites, or deployables.
• For GW-scale AI: Radiator areas could reach hundreds of thousands to millions of m², creating km-scale structures. This drives enormous launch mass penalties unless launch costs drop dramatically and specific power/mass improves.

Heat transport challenges: Heat from dense chip arrays must be moved efficiently to large radiators via heat pipes, fluid loops (e.g., ammonia, as on ISS), vapor chambers, or liquid metal cooling. Long distances increase thermal resistance, gradients, and mass (thicker conductors or more fluid). High heat flux from AI chips (hundreds of W/cm² locally) requires advanced spreading. Current and Emerging Technologies

• Passive: Multi-layer insulation (MLI) for protection, high-emissivity/low-absorptivity coatings (e.g., white paints like AZ-93), deployable radiator panels, and structural panels doubling as radiators.
• Active/hybrid: Mechanically pumped fluid loops, capillary/loop heat pipes, oscillating heat pipes, thermoelectric coolers. Phase-change materials for thermal buffering during eclipses.
• Advanced concepts: Liquid droplet radiators (expose fluid droplets directly for high efficiency/low mass, though containment is tricky), variable geometry/louvers for control, high-temperature operation (500–600 K) for nuclear/power systems (less relevant for silicon chips).
• Ongoing demos: Companies like Starcloud, Orbit AI, and others are testing with NVIDIA GPUs in small satellites, validating radiators and thermal architectures. Modular swarms (many smaller satellites with inter-satellite links) could distribute heat loads.

Radiation synergy: Orbital environments also feature high radiation, requiring shielding (more mass) that can complicate thermal paths. Chips need hardening or frequent replacement, but orbital refresh is difficult. Why It's a Defining Bottleneck for Orbital AIFor low-power satellites, heritage systems suffice. For AI data centers (tens of kW to MW+ per module, scaling to GW), thermal management dominates system mass, cost, and architecture. It competes with power generation (solar arrays also massive/deployable), structural integrity, latency/bandwidth (favoring tight clustering, which worsens local heat), and reliability. Higher operating temperatures help via

T^4

but risk performance/reliability. Innovations in lightweight deployables, better materials (e.g., composites, advanced heat pipes), and in-orbit assembly/ servicing will be critical.
Trade-offs and outlook: Orbital cooling offers "free" cold sink potential and no water use, potentially competitive PUE if mass penalties are overcome. However, for the near term (next decade), it remains challenging and likely favors hybrid or specialized uses rather than wholesale replacement of terrestrial data centers. Physics is unforgiving, but engineering iteration (aided by falling launch costs and demand pull) is pushing boundaries. Expect significant progress from current testbeds, but radiator mass and area will remain key economic and design drivers.

Orbital (space-based) AI compute starts becoming competitive with land-based options in the $200–$500 per kg to LEO range, according to multiple independent analyses, company white papers, and techno-economic studies. Below current Falcon 9 prices (~$1,500–$3,600/kg depending on configuration and customer), it is generally not competitive for general-purpose or bulk AI workloads due to high upfront mass penalties for solar arrays, radiators, structure, shielding, and power systems. At very low costs (e.g., <$100/kg with high cadence), it could gain advantages in specific niches. Why This Price Point?Terrestrial data center costs are dominated by electricity (often 30–50%+ of TCO, or $570–$3,000/kW-year depending on location and rates), plus land, cooling (water/air), grid connections, permitting, and hardware. In orbit, power (24/7 solar) and cooling (radiative) are paid upfront via launch mass, with near-zero ongoing energy/fuel OpEx after deployment (barring station-keeping propellant). However, this requires launching significant extra mass:

• Typical system mass estimates: 20–60+ kg/kW total (solar arrays ~10–15 kg/kW, radiators ~5–20+ kg/kW or more depending on temperature/area, plus structure, bus, shielding, comms, etc.). Optimistic near-term targets are ~30–40 kg/kW; current heritage is higher.
• For 1 MW: Potentially hundreds of tons total mass when scaled (radiators alone can dominate for high-power dense AI).
• Amortized over lifetime (e.g., 5–10 years, with replacement considerations) plus manufacturing, ops, latency/bandwidth penalties, and reliability.

Key cited thresholds:

• ~$200/kg: Frequently cited by Google/Project Suncatcher and others as enabling rough parity on a per-kW-year energy delivery basis with terrestrial options (mid-2030s projection with learning curves). Below this, launch ceases to be the dominant barrier.
• $200–$500/kg: Common range for Starcloud, industry analysts, and models (e.g., $500/kg targets for initial viability with Starship). At $500/kg, some projections show competitive LCOE or payback in 3–4 years for nodes with high utilization.
• One study indicates cost parity potentially below ~$4,000/kg under certain assumptions (with sensitivity to terrestrial energy/carbon prices), but this is an upper bound; most detailed models are more stringent.
• At current ~$2,000–$3,500/kg, orbital is several times more expensive (e.g., 3x+ in some models for 1 GW systems), driven by launch + satellite build costs.

Break-even logic: Launch cost × system kg/kW (amortized) must be low enough that the "free" constant solar + radiative cooling offsets terrestrial electricity + infrastructure costs. Satellite manufacturing (targeted at ~$500–$1,000/kg or lower via scale) is another big piece—launch is necessary but not sufficient. Sensitivities and Additional Factors

• Lower mass/kW (advanced thin-film solar, lightweight deployable radiators <5 kg/m², higher operating temps via T^4 scaling, modular swarms, in-orbit assembly) shifts the breakeven higher (more viable at $500–$1,000/kg).
• Higher terrestrial costs (expensive power regions, carbon taxes, water scarcity, grid delays) make orbital viable at higher launch prices.
• Use case matters: Bulk training/inference (high data movement) suffers from latency/bandwidth costs to ground. Edge processing, sovereign/resilient compute, or low Earth-coupled apps compete earlier. Maintenance/replacement (radiation, failures) adds recurring launch needs.
• Lifetime and utilization: Longer-lived hardened systems or robotic servicing improve economics dramatically.
• Scale: Gigawatt-class benefits from amortization but requires enormous structures (km-scale arrays).

Starship trajectory: Current Falcon 9 is ~20x better than Shuttle era. Starship aims for $10–$200/kg at high cadence via full reusability, mass manufacturing, and flight rate. Analysts see $100–$500/kg realistic in late 2020s–early 2030s with demand pull (Starlink, other mega-constellations). This aligns with the competitiveness window.
Caveats: These are projections and models with wide uncertainty (optimistic assumptions on mass, reliability, and ops). Some skeptics argue even near-zero launch cost struggles due to networking, density, and servicing challenges. Real competition likely emerges first for premium/niche uses before broad parity. Hybrids (terrestrial core + orbital overflow) are probable.
In short, $200–500/kg is the consensus "inflection" zone where orbital compute transitions from exotic/expensive to potentially lucrative and scalable for AI demand—especially with parallel advances in hardware mass efficiency. Falling launch costs, combined with exploding terrestrial bottlenecks, make this a high-stakes race over the next decade.

Orbital data downlink bandwidth—transferring data from space-based systems (like AI data centers) to ground stations—is a critical bottleneck for space compute viability, though laser/optical communications are rapidly advancing to address it. Unlike terrestrial data centers with fiber-optic connections offering near-unlimited internal and external bandwidth, orbital systems face physics constraints: distance, atmospheric effects, orbital dynamics (limited contact windows), power/pointing limits, and regulatory spectrum issues. RF vs. Optical (Laser) Downlinks

• Traditional RF (Radio Frequency): Reliable but bandwidth-limited. Typical rates for LEO satellites: tens to hundreds of Mbps (e.g., S-band ~few Mbps, X/Ku/Ka-band up to ~1-10 Gbps in advanced systems). Starlink user terminals achieve 100+ Mbps downloads (up to hundreds), but per-satellite or gateway capacity is higher yet shared. RF faces spectrum congestion, interference, and lower data density.
• Optical/Laser Communications (Lasercom/FSO): Game-changer, using near-infrared lasers (e.g., 1064 nm or 1550 nm). Offers 100–1,000x higher bandwidth than RF due to terahertz-scale carrier frequencies and narrower beams. Demonstrated rates:
  • 100–200 Gbps from LEO to ground (e.g., NASA TBIRD achieved 200 Gbps; multiple 100 Gbps demos).
  • Up to 1–10+ Gbps operational in various systems; inter-satellite links (ISLs) at 100–200 Gbps (e.g., Starlink lasers).
  • Projections: 10s–100s Gbps per link becoming standard; Tbps-scale with multiplexing or advanced terminals.

Lasercom enables high-throughput downlinks critical for moving AI results, models, or processed data.Key Challenges for Orbital AI Datacenters

• Intermittent Connectivity: LEO satellites pass over ground stations briefly (minutes per pass, multiple times/day depending on orbit and network of stations). Sun-synchronous orbits help with consistency but limit windows. Solutions: Constellations of relays (space-to-space lasers to GEO or other LEO), optical ground stations (OGS) networks, or buffering data onboard.
• Atmospheric Effects: Clouds, turbulence (scintillation), aerosols, and weather attenuate or distort laser beams. Mitigation: Adaptive optics, multiple ground stations (diversity), hybrid RF fallback, or higher orbits (tradeoff with latency/mass).
• Pointing, Acquisition, Tracking (PAT): Lasers require arcsecond-level precision over hundreds–thousands of km due to narrow beams. Demands advanced gimbals, sensors, and control—adds mass/power but proven in demos.
• Power and Mass: High-rate transmitters need power (though efficient); terminals add mass (impacts launch costs). For GW-scale AI, aggregate downlink needs could be enormous if moving raw data.
• Data Volume Mismatch: AI training/inference generates petabytes; downlinking everything is impractical. Key strategy: On-orbit processing/edge AI—run inference on satellite data (e.g., Earth observation) and downlink only insights/results, slashing bandwidth needs by orders of magnitude.
• Regulatory and Infrastructure: Spectrum for RF; coordination for optical (less regulated but needs ground infrastructure buildout). Scaling to thousands of satellites requires massive gateway networks.

Current and Projected Capabilities

• Demos/Operational: TBIRD (200 Gbps), ESA EDRS (~1.8 Gbps relays), Starlink ISLs (100–200 Gbps), various CubeSat terminals (hundreds Mbps to Gbps). Chinese Jilin-1 claims high rates (e.g., 100 Gbps class).
• For AI Concepts (e.g., Starcloud, others): Rely on laser downlinks + third-party networks (Starlink/Kuiper). Plans involve optical links for high throughput; inter-satellite networking to aggregate/relay data. A single high-power module might target 10s–100s Gbps effective; constellations scale total capacity.
• Future Outlook: With investment (market growing at 20–45% CAGR), expect routine 10–100+ Gbps per terminal, multi-Tbps constellation-scale with dense ground networks and in-orbit mesh. Starship-scale launches + mass manufacturing will help deploy more capable terminals.

Implications for Space Compute CompetitivenessFor bulk AI (e.g., large model training), high internal bisection bandwidth is needed within the orbital cluster (via lasers), but Earth interaction favors workloads with low data movement: inference on space-generated data, sovereign/resilient compute, or periodic model syncs. Latency (~few ms round-trip in LEO vs. near-zero terrestrial) further limits interactive uses.
Bottom line: Downlink bandwidth has improved dramatically and will continue (driven by Earth observation, broadband constellations, and AI demand), but it remains a core constraint vs. fiber-connected ground centers. Advances in lasercom, on-orbit AI (reducing data volume), and hybrid architectures make it manageable for viable niches today and increasingly competitive as tech matures. For full-scale orbital AI datacenters, success hinges on smart data strategies as much as raw Gbps. Progress is rapid—200 Gbps today would have seemed futuristic a decade ago.