[moralestapia]

Just my two cents from experience, any sufficiently advanced LLM training or inference pipeline eventually figures out that the real bottleneck is the network!