Distributed Mixture-of-Experts (MoE) inference is an attractive architecture for on-orbit large- language-model serving, because sparse activation decouples model capacity from per-token compute and lets a model be spread across the limited memory and power budgets of multiple satellites. A natural hypothesis—implicit in much terrestrial expert-placement work and in the recent SpaceMoE agenda—is that intelligently placing experts to minimize scarce inter- satellite-link (ISL) bandwidth, while respecting onboard thermal limits, yields a meaningful advantage over naive placement. We test this hypothesis directly. Building an absolute-power thermal model (Stefan–Boltzmann radiator capacity sized to balanced per-node load) and a bandwidth-minimizing placement optimizer, we evaluate placement strategies across a 9,000- scenario parameter sweep and an analytical boundary-math decomposition. Our central result is negative, and we argue it is useful. Once placement is required to respect an absolute radiator constraint sized for balanced load, a thermally balanced placement is already near-optimal on ISL bandwidth: a thermal-first strategy improves on a thermal-spread baseline by a median of 0.00% and a mean of 0.67%, while a trivial round-robin baseline outperforms our optimizer on a majority of feasible scenarios. We further find that the regime in which placement matters at all is narrow and contingent: it vanishes under perfect router load balancing and depends on a lean radiator-sizing premise more aggressive than standard aerospace practice. We conclude that in the thermally binding regime, thermal structure dominates expert placement, leaving little headroom for bandwidth-oriented optimization. This sharpens SpaceMoE’s qualitative claim that thermal is a “first-class constraint” into a quantitative one, and reframes the practical design question from “how to optimize placement” to “how to size radiators and when an optimizer is unnecessary.” We distill the results into five concrete design rules for orbital data-center operators (Section 7). This is an implementation study within an established problem space, not a claim of novelty over it.
Vincent Zhang (Thu,) studied this question.