As terrestrial constraints on artificial-intelligence infrastructure harden (2,300 GW of stalled grid interconnection requests, water permitting failures, and quantified community opposition), the technology sector has begun promoting orbital data centres as an escape from planetary limits. Google's Project Suncatcher, SpaceX's FCC filing for up to one million computing satellites, and operational testbeds from Starcloud and Axiom Space frame low Earth orbit as an environment of free solar energy, free cooling, and zero regulation.
This paper synthesises peer-reviewed atmospheric science, regulatory filings, and engineering literature (2022–2026) to test that proposition. Four findings emerge. First, satellite reentry is now a measured source of stratospheric pollution: aluminium-oxide nanoparticles catalyse ozone destruction, with projected emissions of 360 to 10,000 tonnes per year under filed constellation growth, a burden capable of stalling the Montreal Protocol's ozone recovery. Second, launch black carbon exerts a radiative impact approximately 500 times greater per unit mass than surface sources, constituting what atmospheric scientists describe as an unintended geoengineering experiment. Third, the thermal and electromagnetic physics of orbital computing, including radiator mass requirements near 1,200 m² per megawatt and gigawatt-class microwave power transmission through the ionosphere, remain unsolved at scale and unstudied at the proposed density. Fourth, quantum computing, frequently invoked as the alternative relief valve, is architecturally complementary to, not substitutive of, the classical compute that drives AI demand, and arrives a decade after the constraint binds.
The paper concludes that orbital expansion does not resolve the physical constraints documented in the author's earlier work; it relocates them to an unpriced, unregulated, and largely irreversible commons. The rational pathway remains terrestrial: disciplined siting, grid-aware architecture, allocation governance, and externality pricing, extended in this paper with four specific viability conditions that orbital infrastructure must meet before deployment at scale can be called responsible.
Keywords: orbital data centres; satellite megaconstellations; aluminium oxide; stratospheric ozone; black carbon; Kessler syndrome; power beaming; quantum computing; AI governance; infrastructure physics.
| 1 | Introduction: The Escape-Hatch Narrative |
| 2 | Scope and Method |
| 3 | The Reentry Problem: Alumina and the Ozone Layer |
| 4 | Launch Emissions: The Accidental Geoengineering Experiment |
| 5 | The Thermal Question: Heat in Orbit, Honestly Assessed |
| 6 | Orbital Congestion: The Collapsing Kessler Margin |
| 7 | The Electromagnetic Environment: Exposure and Governance |
| 8 | The Quantum Mirage: The Other False Exit |
| 9 | Discussion: Exporting Constraints Is Not Solving Them |
| 10 | Conclusion: There Is No Off-Planet Exit |
| · | References |
When an industry begins planning its exit from the planet, the first question is not whether the engineering is impressive. It is what the exit admits.
The terrestrial constraints on AI infrastructure are no longer projections; they are operating conditions. Grid interconnection queues in the United States reached 2,300 GW in January 2026, with PJM timelines extending past eight years. Water has shifted from a quiet engineering variable to a permitting barrier, with documented project rejection rates of 30% in markets such as Johor. Community opposition has become economically rational, anchored to measured electricity-bill increases. These constraints, documented in detail in the companion paper Surviving the Surge, have produced a predictable response: if Earth will not permit the build-out, build above it.
The commercial momentum is real and accelerating. Google unveiled Project Suncatcher in November 2025, an 81-satellite constellation targeting sun-synchronous orbit, with leadership confirming launches of satellite-based data-centre racks from 2027. SpaceX filed with the FCC in January 2026 for a constellation of up to one million computing satellites; the filing was accepted in February. Starcloud operated the first Nvidia H100 in orbit in late 2025 and plans a "Hypercluster" architecture for late 2026. Axiom Space launched two dedicated orbital data-centre nodes in January 2026. China has begun launching a planned 2,800-unit AI computing constellation. The European Space Agency has commissioned orbital data-centre architecture studies.
The proposition is seductive because it answers every terrestrial constraint at once: continuous solar power with no interconnection queue; heat rejection to vacuum with no water draw; no land, no planning permission, no neighbours. The orbital environment is presented as an infinite, passive receptacle: a place where the externalities of computation can be sent to disappear.
The argument is not anti-space and not anti-AI. Orbital computing has legitimate near-term applications: satellite-native processing, Earth-observation inference, communications switching. The argument is narrower and harder: that orbital infrastructure at the scale now filed with regulators does not solve the physics documented on Earth. It exports the problem to a commons that has no regulator, no price, and no remediation pathway, and it does so before the science of that commons has been established.
Surviving the Surge (Paper I) established the terrestrial constraint set (power, water, grid, social licence) and proposed the A+ Pathway: appropriate siting, grid-aware architecture, power strategy as core competency, liquid and low-impact cooling, and unified measurement. This paper (Paper II) examines the proposed escape from that constraint set, and extends the A+ Pathway with viability conditions for any orbital deployment. The two papers share one thesis: AI fails when physical limits are ignored, and physical limits do not end at the Kármán line.
This paper is a structured synthesis of evidence published between 2022 and June 2026, across four source classes:
The method follows the discipline of the series: every claim anchored to a named source; projections distinguished from measurements; and uncertainty stated rather than smoothed. Where the evidence is genuinely unsettled, notably power-beam ionospheric interaction and non-thermal radiofrequency biology, the paper says so explicitly. A practitioner's paper earns trust by what it refuses to overstate.
Throughout, "megaconstellation" refers to satellite systems of 1,000+ units. "Orbital data centre" (ODC) refers to free-flying compute infrastructure, whether single-satellite or networked constellation. Figures are stated with their measurement year; the deployment landscape is moving quickly enough that vintage matters.
It does not claim that orbital computing is impossible, nor that current ground-level radiofrequency exposure from communications constellations is harmful; the measured evidence indicates otherwise, and Section 7 reports that honestly. It does not claim that quantum computing is unviable, only that it does not address the workload class driving the constraint. Precision about what is not wrong is what gives the findings about what is wrong their weight.
Every satellite is a future reentry event. At constellation scale, disposal is not an afterthought of the system; it is one of the system's principal emissions.
Satellites are predominantly aluminium by mass. On reentry, that aluminium ablates and oxidises in the mesosphere, producing aluminium-oxide (alumina) nanoparticles of 1–100 nm. These particles matter for two reasons. First, alumina surfaces catalyse the chlorine-activation reaction that destroys stratospheric ozone, a mechanism established in laboratory chemistry since the 1990s. Second, the particles are long-lived, circulating in the stratosphere for years, and they alter the radiative balance of the upper atmosphere: reflecting incoming sunlight while trapping outgoing longwave radiation.
The measurement era has now begun. In February 2026, researchers at the Leibniz Institute for Atmospheric Physics published the first study tying a specific reentry event to a detectable atmospheric pollution plume, moving the field from modelled concern to observed fact. The modelling, meanwhile, has converged on a stark range: full deployment of currently approved megaconstellations implies roughly 360 tonnes of alumina per year (a 646% increase over the natural background delivered by meteoric dust), while a 60,000-satellite LEO population by 2040 implies up to 10,000 tonnes per year, heating parts of the mesosphere by as much as 1.5°C and altering polar vortex behaviour.
Stratospheric ozone recovery is the single greatest success of international environmental governance. Converging research now concludes that, under current constellation growth trajectories, the cumulative alumina burden could render that recovery irrelevant, undoing by accident what the Montreal Protocol achieved by design. No reentry-emissions standard exists in any jurisdiction.
For orbital data centres the implication is structural, not incidental. Compute hardware refreshes on a 3–5 year cycle; LEO satellites deorbit on a 5–7 year cycle. An orbital compute fleet is therefore a continuous alumina injection system: every GPU generation launched is a future stratospheric particle load. The industry's preferred end-of-life strategy, "design for demise" (full burn-up), is precisely the strategy that maximises this emission. Cleaning the orbit pollutes the atmosphere; preserving the atmosphere congests the orbit. That trade has no current solution.
The pollution problem is symmetrical: satellites pollute coming down, and rockets pollute going up.
Hydrocarbon-fuelled launch vehicles deposit black carbon (soot) directly into the stratosphere, an altitude where no rain removes it. Particles persist for months to years, absorbing sunlight and warming the layer in which they sit. The radiative consequence is disproportionate: a 2026 University College London study found that black carbon from launches exerts a climate impact roughly 500 times greater per unit mass than the same emission at the surface, and that megaconstellation deployment, a phenomenon barely six years old, will account for 42% of the space sector's total climate impact by 2029.
Atmospheric scientists studying the phenomenon have reached for a pointed comparison: the launch industry is conducting an unplanned version of stratospheric aerosol injection: the geoengineering approach humanity has so far declined to deploy deliberately, precisely because its side effects on circulation, ozone and regional climate are unpredictable. We declined the controlled experiment and are now running the uncontrolled one.
Scale this to the filings. Maintaining a million-satellite compute constellation with 5–7 year hardware life implies a launch cadence an order of magnitude beyond 2025's record, each launch adding black carbon to the same stratospheric layers receiving the alumina of Section 3. The two emission streams are additive in radiative terms: black carbon absorbs incoming shortwave; alumina traps outgoing longwave. Earlier modelling (Ryan et al., 2022) found just three years of routine launch growth produces measurable global radiative forcing, with damage per kilogram far exceeding any terrestrial industry.
The most common intuition, that orbital data centres will "beam heat at the Earth", is wrong. The real thermal problems are subtler, and in two cases more serious.
In vacuum there is no convection. All waste heat (and effectively all electrical power consumed by compute becomes waste heat) must leave by radiation, governed by the Stefan-Boltzmann law. To reject one megawatt while holding electronics near 20°C requires approximately 1,200 m² of radiator surface, four tennis courts per megawatt. Radiators dominate the mass budget, which dominates the launch count, which feeds Sections 3 and 4. This is the Physics Wall: heat rejection in vacuum scales with area, and area scales with launched mass.
Radiated infrared from orbital platforms disperses omnidirectionally; the fraction intercepted by Earth is negligible against the ~1.36 kW/m² of incident solar flux. A paper that claimed otherwise would not survive review, and this one does not claim it. The thermal threat to Earth is not the radiators.
Three mechanisms do alter Earth's energy balance. First, the particle load: black carbon absorbing shortwave and alumina trapping longwave (Sections 3–4): pollution, not radiators, changes the heat budget. Second, mesospheric heating: NOAA modelling projects up to 1.5°C of warming in parts of the mesosphere from reentry aerosols at 2040 deployment levels, with knock-on effects for polar vortex circulation. Third, and least studied, power and data beaming.
ESA analysis of space-solar-power microwave links found atmospheric attenuation of roughly 2% at 2.45 GHz: small in percentage, large in absolute terms: a single 5 GW geostationary station would deposit ~100 MW continuously into the atmosphere and ionosphere along its beam path. Experimental data on high-power-density beam interaction with the ionosphere is limited to small sounding-rocket experiments (MINIX, 1980s). At the density now filed (up to one million transmitting satellites), cumulative ionospheric energy deposition is, simply, unstudied.
The honest summary: orbital data centres will not cook the planet with infrared. They will, at filed scale, alter the upper atmosphere through particles and beam absorption, by amounts the current science can bound only loosely. In mission-critical engineering, an unbounded variable in a load path is not an acceptable design state. The same standard should apply to a planet.
Unlike carbon, orbital debris has no remediation pathway at scale. It is the only externality in this paper that is, for practical purposes, permanent.
The orbits attractive for solar-powered compute (sun-synchronous, ~650 km) are attractive to every other operator for the same physics, and are already the most congested corridors in low Earth orbit. As of February 2026, roughly 29,800 objects are tracked; the population of lethal but untrackable debris above 1 cm is estimated beyond 1.2 million. Ground tracking sees a fraction of a percent of what can kill a spacecraft.
The Kessler mechanism (collisions producing debris producing collisions) is no longer a thought experiment; it is a margin being consumed in real time. The series' May 2026 research found the average safety margin between conjunction events has collapsed from 164 days in 2018 to 3.8 days in 2026. A compute constellation of one million units would multiply the active-object population by more than an order of magnitude inside the same shells.
The economic structure makes restraint irrational for incumbents: orbital slots are free, debris liability is effectively unenforceable, and first occupancy confers advantage. In any commons with those rules, the race is to deploy before regulation arrives. First-mover advantage in an unpriced commons is first-polluter advantage. The pattern is familiar from fisheries, aquifers and the pre-Montreal atmosphere; the difference is that a cascaded orbit cannot be restocked.
Public anxiety about "microwaves from space" deserves an honest engineering answer, and the honest answer indicts the governance, not today's exposure.
Ground-level power density from communications constellations is extraordinarily low. Independent measurements place ambient satellite signal at roughly 10⁻⁶–10⁻⁵ µW/cm² at the surface, below the power density of the cosmic microwave background. Ku/Ka-band downlinks are designed for high-gain dishes, not human tissue, and attenuate rapidly in atmosphere. On current evidence, ambient exposure from satellite internet is not a rational health concern, and this paper will not pretend otherwise.
Three governance findings are harder to dismiss. First, unintended emissions: radio-astronomy facilities (LOFAR; SKA-Low precursors) have measured Starlink satellites radiating well outside licensed bands, and Bassa et al. (2024) found second-generation V2-mini satellites emitting unintended electromagnetic radiation up to 32× stronger than the first generation, exceeding ITU-recommended thresholds and typical electromagnetic-compatibility standards. The fleet is drifting out of its regulatory envelope faster than the envelope is policed. Second, the regulatory framework is dated: in 2021 the US Court of Appeals (D.C. Circuit) ruled the FCC had acted "arbitrarily and capriciously" in reaffirming 1990s thermal-only exposure limits without addressing the submitted evidence on non-thermal effects. The point here is institutional, not biological: the regulator was found not to be doing the work. Third, power beaming changes the regime: energy-transmission beams are a different class from communications signals: orders of magnitude higher power density along the beam path, with ionospheric interaction (Section 5.2) essentially uncharacterised at scale. Extending a communications-era exposure framework to an energy-transmission era without new science is not regulation; it is inheritance.
Today's exposure: trivial, and demonstrably so. Tomorrow's regime (a million transmitters, power-class beams, documented out-of-band emissions, and a standards framework last grounded in 1996) is unmeasured by design. The deficiency is not radiation. It is accountability.
When orbit is questioned, quantum computing is the next exit offered: a machine so efficient the energy problem dissolves. The architecture says otherwise.
AI's energy demand is overwhelmingly dense linear algebra: the training and inference of neural networks on parallel matrix hardware. Quantum processors do not accelerate that workload class. Industry consensus, including from the hyperscalers building both technologies, is explicit: quantum machines are hybrid accelerators requiring co-located classical high-performance computing, suited to narrow problem classes (optimisation, molecular simulation, cryptanalysis), not to transformer inference at consumer scale. Quantum will sit beside the GPU estate as "quantum pods," not replace it.
For the right problem the gains are genuine: analyst assessments describe quantum systems solving in 200 seconds what would occupy a classical supercomputer for 10,000 years, at a fraction of a data centre's energy. But the first systematic resource studies (Oak Ridge National Laboratory, 2026) project that large-scale fault-tolerant quantum computing carries substantial demands of its own (continuous cryogenic power, significant water, and rare materials including helium-3), demands that national quantum strategies have largely failed to examine. The pattern of Paper I repeats: the energy cost of the processor is quoted; the energy cost of the system is footnoted.
Credible industry projections place commercially valuable quantum machines around 2029 and quantum-accelerated data centres in the late 2030s. The terrestrial constraint set binds now: the interconnection queue, the water permits, the bills. A relief valve that opens a decade after the pressure peak is not a relief valve.
| Question | Evidence-based answer |
|---|---|
| Can quantum run LLM training/inference? | No. Wrong workload class; hybrid accelerator architecture only. |
| Would national quantum computers cut AI data-centre load? | No. They address optimisation/simulation niches; AI demand is unaffected. |
| Is quantum energy-free? | No. Cryogenics, control electronics and error correction are energy-intensive; helium-3 and water demands unquantified until 2026. |
| Is national quantum investment irrational? | No, but it is a sovereignty investment (cryptography, materials, defence), not an AI-capacity investment. Conflating the two misallocates both. |
The strategic conclusion mirrors the orbital one: quantum is a legitimate technology being offered as an answer to a question it does not address. A nation buying quantum capacity to relieve AI load is buying the wrong asset: a sovereignty instrument mis-sold as an infrastructure instrument.
Assembled, the evidence describes a single behaviour repeating at planetary scale: when a constraint binds, the industry relocates it to wherever measurement and regulation are weakest.
Grid congestion was answered with behind-the-meter generation; water limits with siting arbitrage; community opposition with jurisdiction shopping. Orbit is the terminal move in that sequence: the one environment with no permitting authority, no externality price, and no community to object. The attraction of space, examined honestly, is not its solar flux. It is its lawlessness.
Three properties make this relocation more dangerous than its terrestrial predecessors. First, the science lag. The atmospheric chemistry of reentry alumina is today where CFC science stood in the early 1970s (first measurements, early models, industry dismissal), but the deployment curve is ten times steeper, leaving a regulatory lag that history suggests closes only after a forcing event. Second, irreversibility. Carbon is at least theoretically removable; stratospheric particles disperse over years; cascaded orbital debris does not disperse on any human timescale. Reversible risks attract debate; irreversible risks attract silence, because they have no constituency until they occur. Third, the absence of the demand question. Paper I documented that less than 3% of enterprise AI compute serves climate-critical applications while the majority serves engagement and content. The orbital programme answers "where can we put more compute?" It never asks "what is the compute for?" Expansion has been chosen over allocation without the allocation debate ever being held.
This series does not argue prohibition; it argues sequencing. Extending the A+ Pathway of Paper I, orbital compute at scale should be considered viable only when four conditions are met:
None of these conditions is met in June 2026. Until they are, every orbital compute deployment is a bet placed with planetary assets against an unbounded variable: a design state no licensed engineer would accept in a baggage hall, let alone a stratosphere.
The orbital mirage is not a lie about space. It is a misdirection about Earth.
The evidence of this paper reduces to four sentences. Satellite disposal is a measured and growing source of stratospheric pollution capable of stalling the ozone recovery. Launch black carbon is an unplanned geoengineering experiment already underway. The thermal and electromagnetic physics of orbital compute at filed scale are unsolved and, in critical respects, unstudied. And quantum computing, the other promised exit, does not run the workload that created the problem.
What remains is the conclusion the industry has been organising itself to avoid: the constraints on artificial intelligence must be solved where they bind: on Earth, with engineering discipline and governance. The A+ Pathway of Paper I stands: appropriate siting, grid-aware architecture, power strategy as core competency, low-impact cooling, unified measurement, now extended with the four orbital viability conditions of Section 9. To these, this paper adds the demand-side instrument that expansion was meant to make unnecessary: allocation. A civilisation that directs less than 3% of its most powerful tool at its most dangerous problem has a prioritisation failure no constellation can launch its way out of.
None of this is an argument against ambition, and none of it is an argument against space. It is an argument about sequence, the oldest discipline in infrastructure. Foundations before floors. Science before scale. Rules before races. Every system the author has helped deliver in thirty years of mission-critical work exists because someone refused to skip steps. The planet deserves the same engineering standard as a baggage hall.
About the author. Helder Lira is Managing Director of HML Services Ltd, a Hong Kong-based advisory firm operating across APAC and Europe. He has spent more than thirty years delivering mission-critical infrastructure, including HKIA's Third Runway baggage systems, Changi T1 recovery, and Brisbane Airport's dual-terminal BHS, and has practised applied AI since 2020. He is the author of Still Standing, The Human Journey, and the white papers Surviving the Surge and Airport Operations Intelligence.
Ferreira, J. P., Nomura, K., & Wang, J. (2024). Potential ozone depletion from satellite demise during atmospheric reentry in the era of mega-constellations. Geophysical Research Letters, 51. · Typical 250 kg satellite ≈ 30 kg alumina nanoparticles; 2022 reentries +29.5% atmospheric aluminium; megaconstellation scenarios >360 t/yr.
Leibniz Institute for Atmospheric Physics (2026). First direct measurement of a reentry pollution plume. Communications Earth & Environment (February 2026).
Maloney, C., et al. / NOAA Chemical Sciences Laboratory (2025). Reentry aerosol impacts on the stratosphere and mesosphere. JGR: Atmospheres. · 10 Gg/yr alumina scenario by 2040; mesospheric heating to 1.5°C; polar vortex modification.
Plane, J. M. C., et al. (2021); Molina, M. J., et al. (1997); Hanning-Lee, M. A., et al. (1996). Chlorine activation on alumina surfaces · reaction probability ~2%.
NASA TM-20240013276 (2024). Impact of spaceflight on Earth's atmosphere · alumina and black carbon modelling.
University College London, Marais, E. A., et al. (2026). Megaconstellation pollution accumulating in the upper atmosphere. Earth's Future. · Launch black carbon ~500× surface radiative impact; megaconstellations 42% of space-sector climate impact by 2029.
Ryan, R. G., Marais, E. A., Balhatchet, C. J., & Eastham, S. D. (2022). Impact of rocket launch and space debris air pollutant emissions on stratospheric ozone and global climate. Earth's Future, 10. · BC radiative forcing 8 mW m⁻² after three years of routine launch growth.
Marais, E. A., et al. (2024). Global 3D rocket launch and re-entry emissions inventory 2020–2022. Scientific Data, 11. · SMC missions 37–41% of BC, CO and CO₂ emissions by 2022.
SatNews (2026). The "Physics Wall": orbiting data centres face a massive cooling challenge. · Stefan-Boltzmann analysis; ~1,200 m² radiator per MW at 20°C; Starcloud-1 H100 (Nov 2025); Axiom ODC nodes (Jan 2026).
EE Times (2026). The hidden physics of running data centers in orbit.
PRIF (2026). Escaping the heat? The politics and problems of orbital data centers. · Launch dynamics as persistent atmospheric perturbation; orbital environment as dependency, not receptacle.
HML Services research series (May 2026). Orbital horizon analysis · conjunction safety margin 164 days (2018) → 3.8 days (2026); ~29,800 tracked objects; >1.2 M lethal debris objects >1 cm.
FCC (2026). SpaceX constellation filing, up to 1,000,000 computing satellites (filed January, accepted February 2026); Google Project Suncatcher disclosures (2025–26).
ESA ACT (2004). Environmental impact of high-power-density microwave beams on the ionosphere. · ~2% atmospheric absorption at 2.45 GHz; ~100 MW absorbed per 5 GW GEO station; MINIX experimental basis.
Bassa, C. G., et al. (2024). Unintended electromagnetic radiation from Starlink V2-mini satellites. Astronomy & Astrophysics, 689, L10. · UEMR up to 32× first generation; exceeds EMC and ITU-R thresholds.
Grigg, D., et al. (2023). Detection of intended and unintended emissions from Starlink at the SKA-Low site. arXiv:2309.15672.
US Court of Appeals, D.C. Circuit (2021). EHT v. FCC · FCC RF exposure reaffirmation ruled "arbitrary and capricious."
EMF field measurements (2024–26). Ambient ground-level satellite power density 10⁻⁶–10⁻⁵ µW/cm² · below cosmic microwave background.
McCollum, D., et al., Oak Ridge National Laboratory (2026). Resource demands of large-scale fault-tolerant quantum computing. Renewable and Sustainable Energy Transition; Nature Reviews Clean Technology. · Electricity, water, helium-3 constraints; national strategies overlook resources.
CNBC / UBS (2026). Quantum's big leap puts data centers in the spotlight. · Hybrid accelerator consensus; 200 s vs 10,000 yr problem class; commercial value ~2029; quantum-accelerated DCs late 2030s.
Lira, H. (2026). Surviving the Surge: AI Infrastructure at Scale Without Breaking the Planet, V3.1. HML Services Ltd. · Terrestrial constraint set; A+ Pathway; misallocation analysis (<3% climate-critical compute).
© 2026 HML Services Ltd. All rights reserved. · hml-services.com · This paper may be cited with attribution.