Tesla just taped out an AI chip that matches Nvidia's H100 on one specific workload. It can't train a language model. It can't run general ML research. It runs INT4, INT2, and FP8 integer math, and almost nothing else. That's the entire strategy. Nvidia's H100 is built to do everything. FP16 training, FP32 science, transformer attention, sparse operations. Every one of those capabilities costs die area, power, and cooling. The chip ends up drawing 700 watts and 80GB of HBM because it has to…