Orbital AI Data Centers — Economic Competitiveness Timeline > Glossary

Glossary

Definitions of terms and units used throughout this research.

Power Units

kW_IT
IT load power: the electrical power consumed by GPU/TPU compute hardware only (excluding cooling, power conversion losses, bus systems, station-keeping).
kW_total
Whole-of-system power: total power consumption including IT load, cooling, power conversion, and all auxiliary systems.
kW_gen
Nameplate generation capacity: rated output of solar panels or other power sources under standard conditions (AM0 for space, STC for terrestrial).

Cost Basis

TCO
Total Cost of Ownership: amortized capital expenditure plus annual operating expenditure, expressed as $/kW_IT/year.
Capex
Capital expenditure: one-time costs to deploy and commission the system.
Opex
Operating expenditure: recurring annual costs (energy, maintenance, replacement, communication, monitoring, insurance).
LCOE
Levelized Cost of Energy: $/MWh, used for terrestrial electricity cost comparisons only.
WACC
Weighted Average Cost of Capital: the blended rate at which an enterprise finances its assets, reflecting both debt and equity costs. Higher WACC increases the annual cost of owning capital assets. Expressed as a decimal fraction (e.g., 0.08 = 8%).
CRF
Capital Recovery Factor: the annuity factor that converts a one-time capex into an equivalent annual cost accounting for the time value of money. CRF(r, n) = r × (1+r)^n / ((1+r)^n − 1). When r=0, CRF = 1/n (simple amortization). When r>0, CRF > 1/n because capital has a cost. The model uses CRF rather than simple 1/n division to reflect the different cost of capital for orbital vs terrestrial assets.
Pre-PUE
Electricity rate at the meter, before applying PUE to compute the effective rate per kWh of IT load. Multiplying a pre-PUE rate by PUE gives the cost per kWh actually consumed by IT equipment.
Orbital PUE
Ratio of total satellite power to IT load power (1.035-1.10, central 1.05), covering DC-DC conversion losses and housekeeping power (station-keeping, ADCS, communications). Analogous to terrestrial PUE but excludes cooling (handled separately in the thermal mass budget). See [orbital PUE page](pages/orbital-pue/) for derivation.

Mass Units

kg/kW_IT
Satellite mass per kilowatt of IT load delivered. The primary mass efficiency metric for orbital systems.
W/kg
Specific power: power output per unit mass. Used for solar arrays (W_gen/kg) and thermal rejection systems (W_rejected/kg).

Scenarios

Optimistic
Low-cost / high-performance end of the plausible range for the specific parameter being estimated.
Central
Middle-of-the-road assumptions reflecting current trajectories.
Conservative
High-cost / low-performance end of the plausible range for the specific parameter being estimated.

Inference Workload Tiers

Tier 1 (1-8 GPUs)
Small to medium models (up to ~70B quantized). Feasible on any satellite architecture.
Tier 2 (8-72 GPUs)
Large dense models and frontier MoE with wide EP. Fit within a monolithic 72-GPU satellite's internal NVLink domain; exceed a small (8-16 GPU) satellite's domain, requiring cross-satellite parallelism with ISL bandwidth constraints.
Tier 3 (72+ GPUs)
NVL144+ workloads. Approach or exceed even monolithic satellite capacity (~300-500 kW ceiling). The terrestrial roadmap (NVL144 → NVL576 → NVL1152) pushes domains beyond near-term single-satellite capability.

Other Terms

PUE
Power Usage Effectiveness: ratio of total facility power to IT load power. A PUE of 1.2 means 20% of power goes to cooling and overhead.
MoE
Mixture of Experts: a neural network architecture that activates only a subset of parameters per input, requiring wide parallelism across many GPUs.
NVIDIA's high-bandwidth GPU interconnect (1.8 TB/s per GPU in current generation). Required for efficient multi-GPU inference.
LEO
Low Earth Orbit: altitude of ~400-2000 km. Below the Van Allen radiation belts, offering relatively benign radiation environment.
TID
Total Ionizing Dose: cumulative radiation exposure measured in rad(Si). LEO satellites receive ~150-5000 rad(Si) over 5 years depending on shielding.
ADCS
Attitude Determination and Control System: sensors and actuators (reaction wheels, star trackers, magnetorquers) that maintain satellite orientation.
C&DH
Command and Data Handling: the satellite's flight computer and telemetry subsystem.
TRL
Technology Readiness Level: a 1–9 scale for technology maturity, where 9 means flight-proven in operational missions.
COTS
Commercial Off-The-Shelf: standard commercial components not specifically designed or qualified for space.
MTBF
Mean Time Between Failures: average operating time before a component fails. Often conflates transient and permanent failures; see "permanent GPU failure rate" for the orbital-relevant metric.
Permanent GPU failure rate
Annual rate at which GPUs fail irrecoverably, requiring physical replacement. Excludes transient faults recoverable by restart, driver reset, or node reboot. The critical parameter for orbital deployment where hot-swap is impossible. Estimated at 2.5%/4%/6% (optimistic/central/conservative) for H100-class GPUs in terrestrial operation.
GPU job interruption rate
Annual rate at which GPU-related events interrupt running jobs (~17%/year for H100-class). Includes software bugs, network issues, transient hardware faults, and permanent failures. Not a measure of permanent capacity loss.
Overprovisioning
Maintaining spare GPU capacity (typically ~5%) to ensure continuous availability despite hardware failures. Terrestrial operators replace failed GPUs within 3–5 days from a spare pool, keeping effective capacity at ~100%. Not feasible for orbital deployment at terrestrial replacement cadences.
SEE
Single Event Effect: a radiation-induced event where a single charged particle causes a circuit malfunction (bit flip, latch-up, or burnout).
TP
Tensor Parallelism: distributing a single neural network layer's computation across multiple GPUs, requiring all-reduce synchronization via NVLink-class bandwidth.
PP
Pipeline Parallelism: distributing different layers of a neural network across GPUs in a pipeline, requiring only point-to-point transfers between adjacent stages.
EP
Expert Parallelism: distributing MoE model experts across GPUs, requiring all-to-all communication during each MoE layer.
DP
Data Parallelism: running independent model replicas on separate GPUs, each serving independent requests with no cross-replica communication.
HBM
High Bandwidth Memory: stacked DRAM memory used in AI accelerators (e.g., HBM3e in H100/B200), providing 1–8 TB/s bandwidth per GPU.
ADR
Active Debris Removal: the removal of defunct satellites or debris from orbit using a dedicated servicer vehicle.
ISL
Inter-Satellite Link: optical or radio link between satellites for data transfer. Google Suncatcher bench-demonstrated 800 Gbps each-way (1.6 Tbps total) per optical ISL pair.
NRE
Non-Recurring Engineering: one-time development costs (design, qualification, tooling) amortized across production units.
PJM
PJM Interconnection: the largest regional transmission organization in the US, covering 13 states. Uses a capacity auction that has driven significant cost increases for data centers.
ERCOT
Electric Reliability Council of Texas: manages the Texas electric grid. Uses real-time scarcity pricing instead of capacity auctions.
BTM
Behind-the-Meter: power generation located on-site at the data center, avoiding grid transmission and distribution charges.
LCOS
Levelized Cost of Storage: the per-MWh cost of storing and dispatching energy from a battery system, analogous to LCOE for generation.
Effective lifetime
Capacity-weighted operational lifetime derived from a structured reliability model integrating bus loss, GPU accelerator attrition, SDC overhead, and economic obsolescence. Computed as (1 − SDC_overhead) × ∫₀ᵀ [(1−λ_bus)(1−λ_gpu)]ᵗ dt, where T = physical lifetime. Shorter than physical lifetime. Expressed in years.
Platform manufacturing cost
Cost to manufacture non-compute satellite components (solar arrays, thermal rejection system, power electronics, bus structure, wiring, ADCS, propulsion, communications) per kW_IT. Excludes compute hardware (GPU cost) and launch cost.
SWaP
Size, Weight, and Power: the three primary constraints for space hardware design.
BOL
Beginning of Life: the performance level of a component (typically solar arrays or batteries) when first deployed.
EOL
End of Life: the degraded performance level at the end of a component's operational life.
Terminator orbit
An orbit whose plane lies along the day-night boundary (terminator) of the Earth. A dawn-dusk SSO is the most common type. Radiators facing away from Earth see deep space (~3K) with no direct solar illumination, maximizing heat rejection efficiency.
Minimum domain size
The number of GPUs that must be tightly coupled via NVLink-class interconnect (TB/s bandwidth) to serve a single inference request for a given model. Depends on model size, architecture, quantization, batch size, context length, and throughput targets. Determines whether a workload fits within a single satellite's internal NVLink domain or requires cross-satellite networking.