The development

What happened

A November working paper from the Innovation for Cool Earth Forum (ICEF), covered by Axios on 15 November, calls for caution before deploying large constellations of orbital data centers to train and run AI models. It identifies potential harm to night skies, astronomy and the atmosphere, alongside collision risks. The authors recommend phased deployment and better evidence before scaling; their paper is not peer reviewed, and proposed fleets are not the same as satellites already operating.

Key facts

  • David Sandalow's preface describes initial analysis by ten coauthors, mainly based on documentary research from summer 2026. It calls the report a working paper, says it has not been peer reviewed and invites corrections.

  • The findings distinguish roughly a dozen satellites launched to demonstrate AI training and inference in space from proposals for much larger constellations. Those proposals do not establish deployment at that scale.

  • Cooling remains a constraint: in a vacuum, waste heat must be radiated. The authors say radiator size and mass, operating temperature and satellite weight per unit of compute shape feasibility and cost.

  • The report identifies night-sky changes, interference with astronomy, orbital debris and atmospheric effects as risks of large constellations. It says some mitigation would rely on unproven technologies and significant uncertainties remain.

  • Recommendations include slow, phased deployment, performance-based annual license renewals, shared operational data and independent research. These are the authors' proposals, not a description of rules already adopted.

The context

Why it matters

TideReach assessment: moving AI compute into orbit changes the infrastructure trade-offs rather than simply removing them. The relevant test is whether operators can demonstrate cooling, collision avoidance and environmental controls at increasing scale. This working paper makes a case for evidence-gated expansion; it does not establish the outcome of every proposal or independently prove that orbital AI compute is commercially viable.