While the GPU compute cycle dominates the conversation here, what does the next layer of optimisation actually look like? GPUs are unbeatable at matrix multiplication - we know this. But combinatorial explosion problems (logistics routing, drug discovery, financial risk modelling) are still brutal e