Copper-based GPU interconnects have a scale ceiling for LLM inference prefill, and this paper puts a number on it. At 1152-GPU scale, electrical systems cannot deliver time-to-first-token targets for large MoE models at production batch sizes. 3D-integrated photonic interconnects clear that ceiling: 2.8-5.8x latency improvement in communication-limited configurations, 2.1-3.2x in the stressed high-batch regime, 2.2-4.5x speedup across production-grade platforms once copper crosses its scale-up-pod limit.
The paper benchmarks three MoE model classes -- short context (1K-8K tokens), medium (128K), and long (1M) -- across existing copper GPU clusters and a 3D photonic interconnect system. The photonic advantage is not uniform: it is largest precisely where it matters most for SLA-bound serving, in the communication-constrained configurations that dominate real deployments once per-request context grows. The crossover is not about bandwidth marketing; it is about the physical limit on how many copper lanes you can aggregate before signal integrity closes the door.
The implication for infrastructure buyers is direct. If your prefill SLA requires scaling past the copper pod limit, switching to photonic interconnects is now a quantified decision, not an architecture debate. The 1152-GPU threshold is the trigger. Below it, copper is still the cheaper and simpler choice. Above it, 3D photonics is not a research option; it is the only way to close the gap. NVLink- and InfiniBand-based vendors selling into frontier MoE serving deployments should take note: the next pod-size decision for top-tier customers will have a crossover analysis attached, and this paper is the template.