TPU Inference Externalization Full Steam Ahead - InferenceX

SemiAnalysis published the first independent benchmarks of Google's TPUv7 Ironwood pitted against Nvidia's B200 and B300, and the numbers favor Google by a wide margin: up to 50% better performance-per-dollar on inference workloads. The report goes further than a raw speed comparison, breaking down what Google's internal total cost of ownership looks like against what external customers actually pay to rent TPU capacity, a gap that matters for anyone deciding whether to build on Google's cloud silicon or stick with Nvidia GPUs. The timing lines up with Anthropic's reported push toward using over a million TPUs, which reads less like a hedge and more like a bet that TPU economics now clear the bar for a frontier lab's primary training and inference stack. If the benchmarks hold up under further scrutiny, they weaken the assumption that Nvidia hardware is the default choice for large-scale inference, and give every other lab and cloud buyer a concrete number to push back with in their own hardware negotiations.