Orbital AI Data Centers — Economic Competitiveness Timeline > Analysis > Orbital Satellite Operational Lifetime

Orbital Satellite Operational Lifetime

What is the effective capacity-weighted lifetime (years) of an orbital compute satellite, accounting for GPU degradation, catastrophic failures, and regulatory constraints?

Answer

The effective capacity-weighted lifetime of an orbital compute satellite — the number of full-capacity-equivalent years of service delivered per satellite — ranges from 2.2 years (conservative) to 5.9 years (optimistic), with a central estimate of 3.8 years. This is shorter than the physical operational lifetime because five independent degradation mechanisms progressively reduce delivered capacity, and because a deployment delay (ground testing, launch scheduling, orbital commissioning) consumes part of the GPU's economic life before the satellite begins producing revenue.

The effective lifetime is derived from a structured reliability model that separates and independently quantifies each degradation mechanism:

# Mechanism Optimistic Central Conservative Source / Derivation
1 Bus loss (whole-satellite catastrophic) 0.5%/yr 1.0%/yr 2.5%/yr Starlink fleet data + high-power EPS premium
2 GPU accelerator attrition (individual GPU failures) 3.2%/yr 6.8%/yr 16.1%/yr terrestrial-gpu-failure-rate base + space-hardware-failure-rate space additions
3 SDC / error-rate growth (gradual radiation degradation) 1.3% fixed 1.3% fixed 1.3% fixed Negligible at LEO TID doses; constant mitigation overhead
4 Economic obsolescence (GPU generational improvement) Caps physical life at 7 yr 5 yr 3.5 yr gpu-useful-life + operator design targets
5 Spares / graceful degradation (capacity recovery) ~10% recovery ~5% recovery ~0% Small; not included in headline values

The effective lifetime integrates these components:

Effective lifetime = (1 − SDC_overhead) × ∫₀ᵀ [(1 − λ_bus)(1 − λ_gpu)]ᵗ dt

where T = productive operating window (economic life minus deployment delay), λ_bus = annual bus loss rate, and λ_gpu = annual GPU attrition rate.

Scenario Economic Life Deployment Delay Productive Window (T) Bus Loss (λ_bus) GPU Attrition (λ_gpu) Combined Decay SDC Effective Lifetime
Optimistic 7 yr 3 mo 6.75 yr 0.5%/yr 3.2%/yr 3.7%/yr 1.3% 5.9 years
Central 5 yr 5 mo 4.58 yr 1.0%/yr 6.8%/yr 7.7%/yr 1.3% 3.8 years
Conservative 3.5 yr 6 mo 3.0 yr 2.5%/yr 16.1%/yr 18.2%/yr 1.3% 2.2 years

Deployment delay represents the time between GPU procurement and first revenue: ground testing, integration, launch scheduling, and orbital commissioning. The GPU's economic obsolescence clock starts at procurement, not deployment. Estimates: 3 months (optimistic, SpaceX vertical integration with rapid turnaround), 5 months (central), 6 months (conservative, per Patel's estimate patel-2024-ai-bottlenecks.2). This delay reduces the productive operating window and thus the effective lifetime — a 5-month delay on a 5-year economic life reduces T by 8%.

Combined annual decay = 1 − (1−λ_bus)(1−λ_gpu). The multiplicative combination is slightly lower than the arithmetic sum because a failed satellite does not also experience GPU attrition. GPU attrition rates reflect the terrestrial permanent failure rate (2.5%/4%/6% by scenario, from [terrestrial-gpu-failure-rate](terrestrial-gpu-failure-rate.md)) plus space-specific additions (0.7%/2.8%/10.1%, from [space-hardware-failure-rate](space-hardware-failure-rate.md)). The spares/graceful degradation recovery factor (mechanism 5) is documented below but not included in headline values, as the effect is small (<5% change in effective lifetime).

Confidence note: The effective lifetime is a model output, not a directly measured quantity — no orbital compute satellite has accumulated operational lifetime data. The model's sub-inputs have very different evidence quality: physical lifetime and SDC overhead are well-grounded in fleet data and radiation testing; bus loss rate is empirically anchored but requires an engineering judgment premium for high-power systems; GPU attrition — especially the SEL component — is the weakest link, with the destructive SEL rate for H100/B200 spanning an engineering judgment across a >6 order-of-magnitude uncertainty range in the underlying SEL literature. This single sub-input dominates the overall uncertainty. The 2.2–5.9 year scenario range captures this uncertainty, but the central value of 3.8 years should be understood as a weakly informed indicative estimate — likely somewhere in the range of 2–6 years — not a precisely calibrated figure. Because effective lifetime is the analysis's dominant sensitivity lever, this uncertainty propagates strongly into all downstream TCO and parity results. See the Evidence Quality Assessment section below for a component-by-component decomposition.

Why effective lifetime rather than separate replacement opex: In the TCO model, amortizing capex over the effective lifetime captures the cost of maintaining fleet capacity without a separate opex line for failure-driven replacement. A satellite that delivers 3.8 effective kW_IT-years per kW_IT of capex requires investing capex/3.8 per year to sustain capacity — this implicitly includes the cost of replacing failed satellites at the rate needed to keep the fleet at target capacity.

Physical Lifetime Bounds

The physical operational lifetime before deorbit is bounded by a combination of component degradation, economic obsolescence, and operator design targets. The subsystem-level evidence shows that no single component is the binding physical constraint for a well-designed satellite at 5-7 years: modern Li-ion batteries provide 10-12 year capability (2.4x margin for 5 years) saft-ves16-leo-battery.1, solar panels retain 85-97% power at 7 years mdpi-leo-degradation.1, and modern reaction wheels demonstrate very high reliability (3M+ failure-free hours across 800+ units) newspace-systems-rw.1. The binding constraints are, in practice, economic obsolescence (GPU generational improvement) and EPS reliability risk (especially for high-power compute satellites).

Analysis

The Structured Reliability Model

The effective lifetime is computed from five independent mechanisms, each quantified separately with distinct data sources and mitigation paths. This structured approach replaces the earlier method of applying a qualitative space multiplier to terrestrial GPU failure rates.

The mathematical formulation:

Effective lifetime = (1 − SDC_overhead) × ∫₀ᵀ S(t) dt

where S(t) = (1 − λ_bus)ᵗ × (1 − λ_gpu)ᵗ is the surviving capacity fraction at time t, and T is the physical/economic lifetime ceiling.

The integral evaluates to:

∫₀ᵀ S(t) dt = [αᵀ − 1] / ln(α), where α = (1 − λ_bus)(1 − λ_gpu)

For the central case: α = (1 − 0.01)(1 − 0.068) = 0.99 × 0.932 = 0.9227, T = 5 years → ∫ = [0.9227⁵ − 1]/ln(0.9227) = [0.667 − 1]/(−0.0805) = 4.13. Multiplied by (1 − 0.013) = 0.987 → 4.1 effective years.

Component 1: Catastrophic Satellite Loss (Bus Failure)

Bus loss represents total satellite failure from non-GPU causes — EPS failure, ADCS failure, propulsion loss, debris impact, or software/firmware critical failure. When a satellite is lost, all GPUs aboard are lost simultaneously.

Empirical anchors (updated March 2026):

Fleet Annualized Loss Rate Notes
OneWeb ~0.05-0.08%/yr 2 failures / 656 units over 4-7 years oneweb-stats-mcdowell.1
Starlink V2 Mini ~0.9%/yr 92 failures / 6,927 units over ~1.5 yr avg age (still maturing) [mcDowell-starlink-stats.3]
Starlink Gen1 (unplanned) ~1.5%/yr 348 unplanned failures / 4,714 units over ~5 yr avg age (includes immature early batches) [mcDowell-starlink-stats.2]
Iridium NEXT (7-9 yr) ~0%/yr Zero reported failures across 80 satellites, 7-9 years of operation iridium-lifetime-extension-spacenews.1
Smallsat 220-500 kg ~1.0%/yr 96% mission success rate, multi-year missions smallsat-reliability-spacenews-2020.1

The satellite reliability literature shows an infant mortality pattern (Weibull beta=0.45), meaning failure rate decreases with age castet-saleh-2009-satellite-reliability.1. Satellites that survive the first 1-2 years become progressively more reliable. The strong performance of Iridium NEXT (zero failures at 7-9 years) is consistent with this: mature, well-designed bus hardware does not exhibit wear-out within a 5-7 year horizon.

However, Starlink's proactive retirement of 500+ early V1 satellites at <5 years [starlink-retirement-cgaa.1, starlink-mass-deorbit-pcmag.1] demonstrates that first-generation designs can harbor latent failure modes (in this case, a ferrite transformer defect) that would have caused mid-life failures if unaddressed. The satellites were still operational at retirement — SpaceX chose to replace rather than risk in-orbit failure.

Compute-satellite-specific risk premium: An orbital compute satellite operates at 5-130 kW — 2-50x higher power throughput than Starlink (~2-3 kW). The EPS is the dominant failure driver after infant mortality (27-44% of all spacecraft failures [tafazoli-2009-spacecraft-failures.1, kim-castet-saleh-2012-eps.1]), and EPS failures in LEO are more often fatal than in GEO kim-castet-saleh-2012-eps.1. Higher power means higher currents, more power conversion stages, and more thermal stress on power electronics. Additional risk comes from novel radiative thermal management at scale and larger deployable solar arrays. No LEO satellite has ever operated at 5-130 kW sustained power, so the bus failure rate at these power levels is genuinely unknown — the premium below is an engineering judgment, not an empirically derived value.

These factors justify a 2-3x premium over Starlink V2 Mini's ~0.9%/yr rate (or a larger premium over OneWeb's 0.05-0.08%/yr) for the central case:

Component 2: GPU Accelerator Attrition

GPU attrition is the rate at which individual GPUs fail permanently while the satellite bus continues to operate. The total orbital GPU attrition rate combines the terrestrial permanent failure rate with the space-specific additions:

Mechanism Optimistic Central Conservative Derivation
Terrestrial permanent base rate 2.5%/yr 4.0%/yr 6.0%/yr terrestrial-gpu-failure-rate: 5 independent sources
Thermal cycling fatigue 0.1%/yr 0.6%/yr 1.6%/yr Norris-Landzberg model, SAC305 parameters pan-2005-norris-landzberg-sac.1
Destructive SEL 0.1%/yr 1.0%/yr 5.0%/yr COTS catastrophic rate oliveira-2022-cubesat-radiation.1 + FinFET sensitivity ball-sheets-sel-7nm-finfet-2021.1
Radiation soft errors ~0%/yr 0.2%/yr 1.0%/yr 14nm crash rate linux-see-cots-soc-2025.1 + mitigation spacecube-cots-iss.1
TID degradation ~0%/yr ~0%/yr 0.5%/yr Negligible for 5yr at LEO doses tid-7nm-finfet-ro-2021.1
Launch-induced damage 0.5%/yr 1.0%/yr 2.0%/yr Vibration/thermal interaction combined-vibration-thermal-bga.1
Total orbital GPU attrition ~3.2%/yr ~6.8%/yr ~16.1%/yr

The terrestrial base rate now reflects only permanent failures requiring physical replacement — not transient faults recoverable by restart or automation. The previous analysis used a fixed 6% rate across all scenarios based on a journalism article's summary of Meta's Llama 3 data. The updated analysis draws on the primary Meta paper meta-llama3-paper.2, a peer-reviewed longitudinal study (SC '25, 11.7M GPU-hours) cui-two-gpus-2025.1, and Meta's explicit transient vs permanent failure taxonomy revisiting-ml-cluster-reliability.2 to derive a scenario-dependent range.

The optimistic scenario benefits most from this refinement: with a 2.5% permanent base rate and well-mitigated space additions (0.7%), the total orbital GPU attrition drops to ~3.2% — roughly half the previous estimate. The conservative scenario is unchanged because the 6% conservative base rate equals the previous fixed rate.

The implied multipliers (1.3x/1.7x/2.7x) remain consistent with MIL-HDBK-217's rating of LEO as equally benign to "ground benign" for non-radiation failure modes mil-hdbk-217-factors.1, with the additional space-specific attrition dominated by destructive SEL and launch-induced damage.

Component 3: SDC and Error-Rate Growth

A critical finding from the SDC research: radiation-induced gradual performance degradation is negligible for FinFET GPUs in LEO with adequate shielding. This mechanism contributes only a fixed ~1.3% capacity overhead, not a growing annual factor.

The reasoning:

  1. TID parametric drift is irrelevant at LEO doses with adequate shielding. 7nm FinFET shows <1% circuit degradation at 380 krad tid-7nm-finfet-ro-2021.1. The expected 5-year dose depends strongly on shielding: ~15-17 krad behind 3mm Al researchgate-leo-radiation.1, ~1 krad behind 5.7mm Al nusat-tid-leo-2025.1, and ~0.7 krad behind ~10mm Al google-suncatcher.1. Even at minimal shielding (3mm Al), FinFET has a ~25x margin; at the 10mm Al assumed for compute satellites, the margin is ~500x. Threshold voltage shifts, leakage increases, and timing degradation are unmeasurable at mission-relevant doses.

  2. Soft error rates do not increase with accumulated TID at relevant doses. In modern SOI SRAMs at 800 krad, the TID-SEU synergistic effect is only ~15% for 6T cells and actually decreases for 7T cells tid-seu-synergy-soi-sram-2022.1. At the shielded 5-year dose of <1-5 krad (depending on shielding depth), this effect is immeasurably small.

  3. SDC is dominated by manufacturing defects, not radiation accumulation. Meta found ~3.6% of CPUs have SDC from manufacturing and aging defects meta-sdc-fleet-2022.1. Google observes "a few mercurial cores per several thousand machines" [google-cores-dont-count-2021]. These are intrinsic silicon defects, not radiation-induced trends.

  4. Inference workloads are naturally resilient. Single bit-flips are masked ~70-85% of the time in LLM inference llm-soft-error-vulnerability-2025.1. SpaceCube demonstrated <1.3% radiation mitigation overhead over 4 years on ISS spacecube-cots-iss.1.

The 1.3% fixed overhead accounts for ECC, memory scrubbing, watchdog/restart systems, and application-level error detection. It does not grow over the mission because the mechanisms that would cause growth (TID-SEU synergy, parametric drift) are negligible at LEO doses for modern FinFET nodes.

Component 4: Economic Obsolescence

Economic obsolescence sets the ceiling on physical lifetime. With Nvidia releasing new architectures annually nvidia-one-year-cadence.1 and each generation delivering 2-4x inference performance, a GPU launched today is 2-3 generations behind within 3 years and 4-5 generations behind within 5 years. At some point, deorbiting and launching current-generation silicon becomes cheaper than operating obsolete hardware.

For inference workloads (less demanding of cutting-edge hardware than training), this crossover likely occurs at 4-5 years peraspera-realities.4. The "operate, deorbit, replace" model embraces this explicitly.

The physical lifetime bounds (7/5/3.5 years) already incorporate economic obsolescence as a co-equal constraint alongside physical degradation. The convergence of five independent indicators at 5 years — SpaceX filing spacex-fcc-million-satellite-filing.1, Starcloud starcloud-first-ai-model-space.1, Starlink precedent starlink-deorbit-stats.1, hyperscaler depreciation gpu-depreciation-schedules.1, and radiation tolerance limits google-suncatcher.1 — is remarkably consistent.

Obsolescence is asymmetrically worse for orbital: terrestrial GPUs follow a "value cascade" from training to inference to batch workloads and retain residual resale value. Orbital GPUs have zero salvage value and cannot be redeployed. If the value cascade extends terrestrial GPU economic life to 5.5-6 years, the orbital-terrestrial lifetime gap widens slightly.

Component 5: Spares and Graceful Degradation

Several mechanisms partially offset gross GPU failures:

Software routing (no amplification). For data-parallel inference, losing 1 of N GPUs reduces throughput by exactly 1/N (evidence item 26). There is no failure amplification — unlike training workloads where a single GPU failure can destabilize an entire NVLink domain. The per-GPU attrition model already correctly captures this 1:1 proportional degradation.

Soft error recovery. A small fraction (~5-10%) of gross GPU "failures" may be transient (SEU/SEFI recoverable by power cycling or software restart) rather than permanent. SpaceCube demonstrated >99.99% error-free operation with software mitigation spacecube-cots-iss.1, and the Blackwell RAS engine provides predictive failure detection (NVIDIA Blackwell RAS engine).

Cold spares. Unpowered spare GPU modules activated when primaries fail are theoretically viable but carry a mass penalty (~4.0-9 kg/kW_IT each). Assessment: likely suboptimal for mass-production orbital compute — the mass is better spent on active capacity, with fleet-level replacement providing the spare function. Sophia Space's SOOS aiaa-sophia-space.1 and Google's Suncatcher both emphasize software routing around failures rather than hardware spares.

Fleet-level replacement. At fleet scale (10,000+ satellites), individual failures average out to predictable aggregate loss rates. Multi-echelon inventory strategies (in-plane spares, parking orbit warehouses, ground stockpile) achieve 53% cost reduction vs direct replacement constellation-spare-strategy-2025.1. This is the primary recovery mechanism, and it is already captured by the effective-lifetime approach in the TCO model.

Net effect on headline values: The recovery factor (optimistic 10%, central 5%, conservative 0%) would reduce effective GPU attrition by ~0.3%/yr centrally. This changes central effective lifetime from 4.1 to ~4.2 years — a <3% improvement. We do not include this in the headline values because (a) the magnitude is within the uncertainty of other parameters, and (b) the 6.8%/yr central GPU attrition rate from the space-hardware-failure-rate analysis is derived from mechanisms that produce primarily permanent damage (destructive SEL, solder joint fatigue, permanent terrestrial failures). The recovery factor is noted here for completeness and as a potential modest improvement not captured in the conservative effective lifetime estimate.

Why 5 Years Is the Central Physical Design Point

Multiple independent lines of evidence converge on 5 years. We distinguish operator design targets (which reflect economic and engineering judgment) from empirical fleet data (which reflect observed hardware performance):

Operator design targets (Tier 4-5 evidence — statements of intent, not measurements):

Economic evidence (Tier 1-2 — observable market data):

Empirical fleet survival data (Tier 1-2 — measured performance):

Component-level evidence (Tier 1-2 — measured degradation rates):

Cautionary data:

The 5-year central estimate is the strongest part of the physical lifetime analysis. The fleet and component data show that no physical constraint binds at 5 years for a well-designed satellite. The 5-year design point is set by the economic obsolescence cycle (GPU generational improvement), not by hardware degradation.

Evidence Quality Assessment

The effective lifetime is a model output derived from five sub-inputs with very different evidence quality. The central value of 4.1 years is the result of a structured bottom-up calculation, not a directly measured quantity. No orbital compute satellite has accumulated operational lifetime data, so every sub-input involves extrapolation from non-identical contexts.

Sub-input Central Value Evidence Basis Evidence Quality Type of Estimate
Physical lifetime (T) 5 years Operator design targets, fleet precedent, GPU depreciation, component degradation data Medium-High — Multiple independent convergences; fleet data (Iridium, Starlink, OneWeb) validates 5+ years for bus hardware; no compute-satellite-specific data Engineering consensus + empirical fleet data
Bus loss rate (λ_bus) 1.0%/yr Starlink/OneWeb fleet statistics + EPS risk premium for high-power systems Medium — Fleet data is Tier 1 (directly observed); the 2-3x premium for high-power EPS is engineering judgment extrapolated from lower-power analogs Empirical base rate + judgment premium
GPU attrition (λ_gpu) 6.8%/yr Terrestrial permanent failure rate (5 independent data sources) + space-specific additions (mechanism-level models) Medium-Low — Terrestrial base rate (4%/yr) is well-sourced (SC '25, Meta, Microsoft); space additions (2.8%/yr) are bottom-up mechanism models with no orbital validation; SEL component (1.0%/yr central) is an engineering judgment across a >6 OOM uncertainty range Empirical base + mechanism models + judgment
SDC overhead 1.3% fixed SpaceCube ISS data, Google Suncatcher, FinFET TID testing Medium-High — Multiple empirical data points; negligible at LEO doses with adequate shielding Empirical measurement
Economic obsolescence Caps at 5 years Hyperscaler depreciation schedules, Nvidia release cadence High — Observable market data; depreciation schedules are public financial disclosures Market observation

The sub-inputs that drive the most scenario spread — GPU attrition (especially destructive SEL) and bus loss rate for high-power satellites — are also the ones with the weakest evidence. This is inherent to the problem: the quantities we most need to know (GPU failure rates in space, high-power bus reliability) have no empirical precedent to draw on. The model makes these uncertainties explicit and separable rather than hiding them in a single composite estimate.

Key Uncertainties

  1. No public SEL characterization for NVIDIA H100/B200 — confirmed as genuinely unknown. Targeted research for SEL data on TSMC 4nm/5nm commercial devices found zero published results. NASA's statistical SEL database shows SEL rates vary across >6 orders of magnitude with no predictive trends [ladbury-2025-sel-statistics.2, ladbury-2025-sel-statistics.4]. The 0.1-5.0%/yr destructive SEL range used in this analysis is an engineering judgment spanning the plausible distribution, not an empirically bounded estimate. See space-hardware-failure-rate for the full assessment.

  2. Bus failure rate for high-power compute satellites is an engineering judgment. Fleet data provides excellent empirical anchors for communication-satellite-class hardware (OneWeb 0.05-0.08%/yr oneweb-stats-mcdowell.1, Starlink mature batches ~0.3%/yr [mcDowell-starlink-stats.1], Iridium NEXT ~0%/yr at 7-9 years iridium-lifetime-extension-spacenews.1). The 2-3x premium applied for compute satellites at 5-130 kW reflects the higher EPS stress, novel thermal management, and absence of operational precedent — this is reasonable engineering judgment, but it is not an empirical measurement. The infant mortality pattern (Weibull beta=0.45 castet-saleh-2009-satellite-reliability.1) suggests the premium may be front-loaded: mature compute bus designs could approach communication-satellite reliability levels.

  3. Design maturity trajectory. Bus failure rates will likely decrease through design iteration (as Starlink demonstrated: from 13% to 0.2% across generations starlink-failure-rates-wccftech.1), but the timeline depends on production volume and operational learning rates. Starlink's proactive retirement of 500+ early satellites at <5 years starlink-retirement-cgaa.1 shows that even experienced operators discover latent defects in first-generation designs.

  4. Interaction between bus and GPU mechanisms. A high-power GPU failure causing EPS overload could trigger catastrophic satellite loss. This interaction between mechanisms 1 and 2 is not modeled and could increase effective whole-satellite loss rates beyond the sum of independent terms.

SEL Rate Sensitivity

The destructive SEL rate is the most uncertain sub-parameter in the effective lifetime model. NASA's statistical SEL database shows rates varying across >6 orders of magnitude with no predictive trends ladbury-2025-sel-statistics.2, and there are zero published SEL characterization results for TSMC 4nm/5nm commercial GPU dies (H100/B200). The table below isolates this uncertainty by varying only the destructive SEL rate while holding all other failure parameters at their central values: physical lifetime 5 years, bus loss 1.0%/yr, terrestrial permanent GPU failure 4.0%/yr, non-SEL space additions 1.8%/yr, and SDC overhead 1.0%.

SEL Rate Total GPU Attrition Combined Annual Decay Effective Lifetime (years) vs Central
0.1%/yr 5.9%/yr 6.8%/yr 4.17 +2%
0.5%/yr 6.3%/yr 7.2%/yr 4.13 +1%
1.0%/yr 6.8%/yr 7.7%/yr 4.08 central
2.0%/yr 7.8%/yr 8.7%/yr 3.97 −2%
5.0%/yr 10.8%/yr 11.7%/yr 3.69 −10%
10.0%/yr 15.8%/yr 16.6%/yr 3.25 −20%

Even at the extreme high end (10%/yr destructive SEL — near the conservative scenario's total space-specific GPU attrition), the effective lifetime declines by only 20% from the central value, from 4.1 to 3.3 years. This moderate sensitivity arises because the SEL rate compounds with other attrition mechanisms that already dominate: the 4.0%/yr terrestrial base rate and 1.8%/yr non-SEL space additions together account for 5.8%/yr before any SEL contribution. The effective lifetime is more sensitive to physical lifetime (economic obsolescence) and the terrestrial base rate than to the SEL rate alone. Nevertheless, a SEL rate at the high end of the plausible range (5–10%/yr) would meaningfully compress effective lifetime and shift the economics toward the conservative scenario.

Evidence

Regulatory Constraints

  1. FCC 5-year deorbit rule (effective Sept 2024) The FCC adopted rules requiring LEO satellite operators to complete disposal within 5 years of mission end. This replaced the prior 25-year guideline. New licensees after Sept 29, 2024 must comply. The rule requires post-mission disposal within 5 years of mission end — it does not impose a limit on operational lifetime.

  2. SpaceX FCC filing: 5-year operational life SpaceX's January 2026 filing for up to 1 million orbital data center satellites at 500-2,000 km explicitly states the data centers "are expected to have an operational life of five years." This is the operator's own design target, driven by engineering and economic considerations rather than the FCC disposal rule.

  3. Starcloud: 5-year lifespan Starcloud CEO stated their "satellites should have a five-year lifespan given the expected lifetime of the Nvidia chips on its architecture."

Satellite Bus Reliability

  1. Starlink V2 Mini fleet reliability [evidence:mcDowell-starlink-stats.1] As of March 2026, 11,641 Starlink satellites launched total. Gen2 V2 Mini: ~31 failures across 6,927 units (0.45% cumulative over ~1.5 years average age, implying ~0.3%/yr annualized). Gen1 (4,714 launched): 242 failures (5.1% cumulative over ~5 years). Total fleet: 327 failures (2.8% cumulative across all generations). — McDowell, planet4589.org

  3. OneWeb fleet reliability As of June 2023, 4 of OneWeb's 634 satellites had failed in orbit (0.63% cumulative over ~3 years, implying ~0.2%/yr annualized). By October 2024, Eutelsat CEO expressed confidence to extend satellite operational life beyond the initial design. — Space Intel Report

  4. Spacecraft failure subsystem distribution Analysis of 156 on-orbit failures across 129 spacecraft (1980-2005): AOCS caused 32% of all failures (gyroscopes alone 17%); power subsystem caused 27%; together AOCS and power accounted for 59% of all failures. — Tafazoli, Acta Astronautica, 2009

  5. EPS failure severity in LEO The electrical power subsystem (EPS) fails less frequently but more fatally in LEO than in GEO. After 10 years of operation, EPS accounts for 44.1% of all failures. 29% of LEO EPS failures are in the electrical distribution subsystem. — Kim, Castet, Saleh, Reliability Engineering and System Safety, 2012

  7. Smallsat mission success rates Smallsats (220-500 kg class) achieved 96% mission success rate (2009-2018). Most failures occur in the first 60 days: "If you can make it through your first two months, you'll likely make it through your entire design life." — SSC21-WKIII-02

  8. CubeSat failure distribution CubeSat EPS causes more than 40% of all failures after 30 days of operation; communications accounts for ~26-30%. Most failures are immaturity failures, not environment-induced. Improved testing beats subsystem redundancy for improving reliability. — Bouwmeester et al., 2022

Radiation Degradation

  1. LEO TID environment (light shielding) With 3 mm Al shielding, total ionizing dose in LEO is <10 krad(Si) over a 3-year mission — extrapolating linearly, ~15-17 krad(Si) over 5 years. This represents the dose behind minimal shielding; compute satellites with 10mm Al shielding experience ~5-10x lower doses (see items 13-14).

  2. Google Suncatcher TPU radiation testing Google's Trillium TPUs (v6e Cloud TPU) survived proton beam testing to ~2 krad(Si) before HBM subsystems showed irregularities, with no hard failures up to ~15 krad. Google reports the expected shielded 5-year mission dose as ~700 rad(Si) (behind ~10mm Al equivalent), giving ~3x margin to first HBM irregularities and ~20x margin to hard failure threshold. Commercial AI silicon can survive a 5-year LEO mission with adequate shielding.

  3. In-situ LEO TID measurements LabOSat-01 dosimeters on ÑuSat satellites in polar LEO (~490 km, 97.3° inclination) measured 0.5-1.9 krad over ~3 years depending on shielding depth. At ~2.9mm Al: ~1.9 krad/3yr (~3.2 krad/5yr); at ~5.7mm Al: ~0.6 krad/3yr (~1.0 krad/5yr). These in-situ measurements are consistent with the Google estimate of ~700 rad/5yr behind ~10mm Al — dose decreases roughly exponentially with shielding depth. — arXiv:2503.09520, 2025

  4. 7nm FinFET TID tolerance 7nm bulk FinFET ring oscillators show circuit-level degradation (logic gate delays, operating current) of less than 1% after TID exposure of 380 krad(SiO2) — roughly 500x the 5-year shielded LEO dose. — IEEE TNS, 2021

  5. TID-SEU synergy in modern silicon In modern SOI SRAMs, TID irradiation to 800 krad increases 6T SRAM SEU cross-section by only ~15%, while 7T SRAM cross-section actually decreases by ~60%. This contrasts with micrometer-scale SRAMs where SEU cross-section increased by up to 1000x after TID exposure. At mission-relevant doses (<5 krad), the TID-SEU synergistic effect is immeasurably small. — MDPI Electronics, 2022

Derived Inputs (GPU Failure Rates)

Items 17-18 are derived from other research pages, not direct evidence. They summarize inputs consumed by this page's model.

  1. Terrestrial permanent GPU failure rate [analysis] The terrestrial GPU failure rate analysis establishes the annual permanent failure rate for H100-class GPUs at 2.5%/4%/6% (optimistic/central/conservative). This reflects only failures requiring physical GPU replacement — not transient faults recoverable by restart. The range is derived from five independent data sources: Meta's Llama 3 primary paper [meta-llama3-paper.2], the NCSA Delta longitudinal study (2.5 years, 11.7M GPU-hours) [cui-two-gpus-2025.6], Meta's ML cluster reliability study [revisiting-ml-cluster-reliability.2], the NTP paper [nonuniform-tensor-parallelism.1], and Microsoft SuperBench [microsoft-superbench.1]. The Epoch AI MTBF of ~50,000 hours [epoch-gpu-failures.1] conflates all 419 job interruptions (hardware + software + network) as "failures" and should not be used for permanent failure rate estimates.

  2. Space-specific GPU failure mechanisms [analysis] The space hardware failure rate analysis builds a bottom-up mechanism-level budget for space-specific GPU attrition: thermal cycling fatigue (0.1–1.6%/yr), destructive SEL (0.1–5.0%/yr), radiation soft errors (0–1.0%/yr), TID degradation (0–0.5%/yr), and launch-induced damage (0.5–2.0%/yr). These sum to 0.7–10.1%/yr space-specific additions on top of the scenario-dependent terrestrial permanent base rate, giving total orbital GPU attrition of 3.2–16.1%/yr.

SDC and Gradual Degradation

  1. SpaceCube radiation mitigation overhead SpaceCube on ISS achieved >99.99% error-free operation of eight COTS PowerPC processors over four years using FPGA and software radiation mitigation (scrubbing, error detection, watchdog, checkpoint/restart). Radiation hardening by software overhead was <1.3%.

  2. HBM radiation sensitivity HBM uncorrectable ECC error sensitivity: approximately one event per 50 rad of proton exposure. With 10mm Al shielding in dawn-dusk SSO (~150 rad/yr), approximately 1 uncorrectable error per 10 million inferences. Google assessed this as "likely acceptable for inference."

  3. LLM inference soft error masking In LLM inference, most errors from a single bit-flip are masked (cause no output change). The abnormal outcome rate grows from ~15-30% for 1 injected fault. Larger models show a higher proportion of masked errors. Bit-flip vulnerability is strongly position-dependent: low-order bits cause <1% SDC rate while high-order bits (30-31) cause 23-24% SDC rate. — Chai et al., arXiv:2601.19912, 2025

  4. GPU SDC terrestrial baseline GPU silent data corruption rate is 8.15×10⁻³ FIT per device at sea level (one error per 14,000 device-hours). Cosmic radiation causes 61.7% of faults; manufacturing variations contribute 4.3%. — Global Journal of Computer Science and Technology, 2025

  5. Meta fleet SDC rates Meta found approximately 3.6% of CPUs in their fleet cause SDCs. Root causes are CPUs that are "born defective (escaped manufacturing testing), become defective (aging), or just differ from each other (timing variability)" — manufacturing defects, not radiation accumulation. — Meta Engineering Blog, 2022

Graceful Degradation

  1. Sophia Space SOOS Sophia Space's SOOS monitors each tile's temperature and computational load, redistributing work when a quad or entire tile goes offline due to debris impact or hardware failure. The system gracefully degrades performance until a replacement module is launched. — Aerospace America/AIAA, 2026

  2. Inference failure scaling: no amplification [analysis] Data parallelism for inference provides linear throughput scaling — losing one data-parallel replica reduces aggregate throughput by exactly 1/N. There is no failure amplification in data-parallel inference because each replica serves independent requests. This is a fundamental architectural property of data-parallel inference, in contrast to tensor-parallel training where a single GPU failure can halt an entire NVLink domain.

  3. Epoch AI spare buffer analysis A 1M GPU cluster experiencing 20 node failures/hour with 1-day replacement time needs a buffer of only 480 nodes (0.3% of total capacity). Even with multi-day replacement delays, performance impact remains below 1%. — Epoch AI, 2024

  4. Constellation spare strategy Multi-echelon inventory optimization for satellite constellations: in-plane spares for immediate replacement, parking orbit spares for hours-timescale resupply, ground stockpile with launch lead time. The indirect strategy (parking orbit spares with batch resupply) achieves 53% total cost reduction vs direct replacement. — arXiv:2509.09957, Georgia Tech, 2025

Solar Panel Degradation

  1. GEO solar array measured degradation Telemetry from 11 GEO communications satellites showed GaAs cells degrade at 0.44-1.03%/yr and Si cells at 0.71-1.69%/yr. LEO radiation fluences are 5-10x lower than GEO.

  2. ISS solar array degradation The ISS P6 silicon photovoltaic arrays showed measured short-circuit current degradation of 0.2-0.5%/yr, below the predicted rate of 0.8%/yr.

  3. Solar panel lifetime implication At 0.2-1.5%/yr degradation in LEO, a 5-year satellite retains ~92-97% of initial power; a 7-year satellite retains ~90-96%. Solar panels are NOT the binding constraint on satellite lifetime.

GPU Obsolescence

  1. Nvidia 1-year release cadence Nvidia has shifted from a 2-year to a 1-year release cadence for datacenter GPUs: Hopper (2022), Blackwell (2024-25), Rubin (2026), Feynman (2028). Each generation delivers roughly 2-4x performance improvement for AI inference.

  2. GPU depreciation schedules AWS, Google, and Microsoft use 6-year depreciation schedules for servers/GPUs. Industry is converging toward 5-year useful life via a "value cascade" model. AI-native neoclouds use 4-5 year schedules.

  3. Dylan Patel on orbital GPU economics Dylan Patel (SemiAnalysis) noted that the testing, assembly, and launch process for orbital GPUs could consume 6+ months, representing "10% of your cluster's useful life" if GPU useful life is 5 years.

  4. "Operate, deorbit, replace" model Per Aspera analysis proposes that early orbital clouds will default to designing for a "5-7 year tour, then burn it up and launch Version N+1 with the latest silicon."

Station-Keeping

  1. Orbital decay at LEO altitudes Atmospheric drag in LEO causes ongoing orbital decay, with rates increasing substantially during high solar activity. Station-keeping propulsion is required throughout the operational lifetime to maintain altitude. At ~575 km, uncontrolled satellites deorbit within a few years.

Fleet Operational Lifetime Data

  1. Iridium NEXT: zero failures through 7-9 years, life extended to 17.5+ years Iridium NEXT satellites (launched 2017-2019, 12.5-year design life from Thales Alenia Space) were re-assessed in February 2024 to operate "well to at least 2035" (17.5+ years). CEO Matt Desch expects satellites will be in service "longer than 17.5 years." First-generation Iridium with similar design life lasted 20+ years before fuel depletion, not component failure. — SpaceNews, February 2024

  2. First-generation Iridium: 20+ years in LEO Original Iridium satellites (8-year design life, launched 1997-2002) operated for 20+ years at 780 km altitude, completing ~100,000 orbits each. Decommissioning was driven by fuel depletion, not component failure. This is the strongest empirical data point for LEO satellite bus longevity: 2.5x design life achieved. — SpaceNews, February 2024; Iridium Communications

  3. Starlink Gen1 unplanned failure rate: ~1.5%/yr [evidence:mcDowell-starlink-stats.2] Starlink Gen1 unplanned failures (excluding proactive retirements): early deorbits (114) + reentry after failure (127) + screened (100) + dead (7) = 348 out of 4,714 launched = 7.4% cumulative over ~5 years average age, implying ~1.5%/yr annualized. This includes immature V0.9 (100% loss) and early V1.0 batches with known design flaws. — Derived from McDowell tracking data, March 2026

  6. OneWeb: 0.3% cumulative failure over 4-7 years As of March 2026, OneWeb has 656 satellites launched across 20 missions (2019-2024), with 654 in orbit and only 2 total down. 637 fully operational. Cumulative failure rate 0.3% over 4-7 years of operation (~0.05-0.08%/yr annualized). Design life 7+ years. — Jonathan McDowell, planet4589.org

  7. Satellite infant mortality (Weibull beta=0.45) Analysis of 1,584 Earth-orbiting satellites (1990-2008) shows satellite failures follow an infant mortality pattern: Weibull shape beta=0.4521, scale theta=2,607 years. Beta < 1 indicates failure rate decreases with age — satellites that survive early life become progressively more reliable. This contradicts an earlier assumption of wear-out (beta=1.7). — Castet & Saleh, Reliability Engineering and System Safety, 2009

  8. Li-ion batteries provide 10-12 year LEO capability Saft VES16 space-qualified Li-ion cells achieve 65,000+ cycles at 30-50% depth of discharge over 12 years. A 5-year LEO mission requires ~27,000 cycles; a 7-year mission ~37,800 cycles. Li-ion provides ~2.4x cycle margin for 5 years. Prior Ni-Cd/Ni-H2 batteries lasted only 5-7 years. — Saft manufacturer data

  9. Modern reaction wheels: 3M+ failure-free hours NewSpace Systems T065 reaction wheel (first flown 2014) has accumulated over 3 million failure-free hours in orbit with no reported SEUs. Over 800 wheels sold, baselined on 4 constellation programs. — NewSpace Systems