LAS, NEXUS enable bit-exact neuromorphic LLMs

Neuromorphic computing has finally solved the conversion loss bottleneck, transforming it from an academic curiosity into a credible threat to the GPU status quo. LAS and NEXUS prove that spiking neural networks can match the precision of floating-point ANNs at 70B parameter scales, while energy efficiency gains on Intel's Loihi 2 could make high-end LLM inference sustainable for edge devices. IBM's NorthPole provides a non-spiking alternative that still delivers a 72x efficiency boost over H100s, showing multiple viable paths to the post-von Neumann era. The inherent defect tolerance of distributed neuron architectures could also lead to higher manufacturing yields and lower consumer prices.