General Compute, an inference cloud for alternative chips, is buying Cerebras systems to run beside its Nvidia GPUs. Every request gets split in two, with the GPUs reading the prompt and Cerebras writing the answer. Putting two rival chipmakers on a single request sounds backwards. Here's the mechanism behind the split. Prefill reads your whole prompt in one pass, a parallel math job GPUs were built for. Decode writes the answer one token at a time, and that part is limited by how fast the…