SemiAnalysis, a semiconductor research firm, has released its first third-party inference test results for Google's TPUv7 Ironwood. The report indicates that, under comparable conditions, Ironwood leads NVIDIA's B200 and B300 by up to 50% in performance per dollar. At an interactive speed of 100 tokens per user per second, Ironwood's inference cost per million tokens is approximately $0.181, lower than the B200's $0.222 and the B300's $0.276. Google is accelerating the opening of TPUs to external customers and plans to open-source the new native PyTorch backend, TorchTPU, in October to lower the barrier for developers. Anthropic has committed to purchasing over a million TPUs, and its TPU usage is expected to surpass DeepMind's own by 2029, posing a direct challenge to NVIDIA's market position.