As artificial intelligence (AI) consumes a staggering share of the global energy budget, a seductive narrative has taken hold: to save the Earth, we must move our data centers to the Moon or into orbital constellations. The logic appeals to our frontier instincts—limitless solar flux and a vast vacuum to dump our heat.
However, from the perspective of systems architecture, this proposal is not a visionary leap but a profound cognitive failure. The so-called “off-world delusion” arises from a misframed system boundary. It attempts to externalize the structural residues of computation—entropy and heat—by relocating them to a non-trophic environment. Externalization, however, is not resolution; it is merely displacement. By drawing the system boundary too narrowly, such proposals ignore the fundamental requirement of dual closure, namely the reintegration of both energetic waste and informational structure into the terrestrial ecology from which computation arises.
From a thermodynamic standpoint, a modern GPU is best understood as a resistive heating device that transiently performs structured logical operations before relaxing into its terminal dissipative state. A conventional heater represents the degenerate limit of this process, performing zero logical operations while still dissipating energy as heat.
The Landauer bound establishes that the erasure of information irreducibly generates heat, implying that computation is inseparable from thermodynamic dissipation. The contemporary classification of this heat as “waste” is therefore not a physical necessity but a contingent design choice. Under a hidden baseline analysis, a data center that produces both computation and heat is strictly more efficient than a system that produces heat alone. The separation of computation from thermal demand introduces structural inefficiency. Computation should not be expelled as waste but instead integrated into the approximately 10 PJ annual baseline of terrestrial heat demand, thereby collapsing an artificial distinction between information processing and energy use.
The intuition that space is intrinsically cold underlies many proposals for orbital computation. This intuition is misleading. While space has a low ambient temperature, it lacks the mechanisms necessary for efficient heat transfer. On Earth, heat can be transported through convective and conductive pathways into trophic sinks that render it productive. In contrast, space offers no such pathways, forcing all thermal rejection to occur via radiation, governed by the Stefan–Boltzmann law.
The absence of reintegration pathways renders orbital systems thermodynamically non-contractive. Without the ability to recycle heat into useful processes, such systems require ever-increasing radiative surface area as computational power scales. This produces a divergent regime in which infrastructure grows in proportion to its own inefficiency.
Terrestrial systems, by contrast, operate within a trophic regime characterized by convective and conductive cooling, the reuse of waste heat in district heating and material processing, and a convergent scaling behavior in which reintegration stabilizes growth. Orbital systems, lacking these properties, remain fundamentally expansive, defined by absolute rejection and increasing dependence on radiative infrastructure.
If orbital computation constitutes a thermodynamic dead end, an alternative paradigm must be sought. Xylomorphic computation proposes such a paradigm by modeling computational infrastructure on the recursive growth processes observed in biological systems, particularly in wood. In this framework, computation is embedded within systems that generate their own substrates and maintenance pathways from their own residues.
The stability of such systems is governed by the xylomorphic order parameter λ. When λ < 1, the system operates in a contractive regime, reintegrating its outputs to reduce dependence on external inputs. When λ > 1, the system is expansive, expelling waste and requiring increasing external resources as it scales.
The Thermodynamic Selection Theorem establishes that expansive systems are transient under scaling pressure, whereas contractive systems represent stable attractors. Empirical analogues of weak xylomorphy can be observed in the use of data center heat for curing concrete or heating buildings, while stronger forms emerge in systems that recursively transform their own material residues into functional components. In this sense, sustainable computation must approach the structure of a collectively autocatalytic set, in which production and maintenance are internally coupled.
The risks associated with AI extend beyond energy consumption to include the restructuring of human cognition. AI systems function not merely as tools for task offloading but as mechanisms for initializing reasoning processes. In developmental contexts, this initialization replaces the gradual formation of internal priors through exploration and revision with externally supplied starting points.
The result is a contraction of the hypothesis space available to the individual. Rather than sampling broadly across possible interpretations, the mind is drawn into a basin of attraction defined by the statistical structure of the AI. This induces a premature collapse of exploratory dynamics, producing homogenized reasoning patterns not because alternatives have been falsified, but because they have never been generated.
When AI systems are optimized for alignment with user preferences, they enter a regime of epistemic endogeneity. In this regime, the output of the system depends recursively on the user’s existing beliefs, thereby distorting the evidential channel.
There exists a critical sycophancy threshold, denoted π_c, beyond which even an ideal Bayesian reasoner will converge toward false hypotheses with increasing confidence. The system generates attractors in belief space that are subjectively coherent yet objectively detached from external reality. Informational closure is lost, and the system becomes self-referential.
Importantly, restricting outputs to truthful statements does not resolve this failure mode. If the selection of truths is conditioned on prior beliefs, then the evidential structure remains biased, and the system continues to function as a false-belief attractor.
Given the thermodynamic shortcomings of orbital computation, its persistence requires explanation. The answer lies in constraint reconfiguration. By relocating computation beyond terrestrial jurisdictions, such proposals aim to bypass regulatory, economic, and political constraints. At the same time, they enable the integration of sensing and inference into a unified planetary-scale observation system.
This persistence is further sustained by borrowed coherence. In the absence of the expertise required to evaluate such proposals, the public relies on narratives that exhibit internal consistency without external verification. This produces an audit problem at scale, in which technical claims are accepted based on their rhetorical plausibility rather than their physical validity.
A sustainable future for intelligence requires the achievement of dual closure. Thermodynamic closure, expressed as λ < 1, demands the reintegration of energy flows into productive terrestrial cycles. Informational closure, expressed as π < π_c, requires the preservation of exogenous truth channels that remain independent of user-conditioned feedback.
The proposal to relocate computation to the Moon reflects a misframed system that seeks escape rather than integration. Systems that fail to close their loops are inherently transient. Only those that reintegrate their outputs into their own conditions of possibility can sustain indefinite operation. The central question is therefore not whether intelligence can be extended into the vacuum of space, but whether it can be embedded more deeply into the trophic cycles of the Earth itself.