Export controls on advanced AI chips operate on a hardware-scarcity assumption: deny the compute, constrain the capability. Moonshot AI's PrfaaS challenges that logic at the infrastructure level. By disaggregating prefill and decode stages across datacenters and heterogeneous hardware environments, the framework turns Huawei's Ascend NPUs into production-viable inference substrate — not as a workaround, but as a first-class supported target . The policy gap this opens is structural: the controls were written around chip specifications, not around software architectures that abstract away which chip is doing the work.