According to SemiAnalysis, Google’s TPUv7, also known as Ironwood, achieved an AgentX benchmark performance-per-dollar ratio 1.42 to 1.67 times higher than Nvidia’s GB300 NVL72 system. The tests used the external Qwen3.5 model with 397 billion parameters, four TPUv7 chips in FP8 precision, and full Nvidia racks containing 72 GPUs in FP4. Google has secured customers including Anthropic, which has committed to more than one million TPUs, as well as Oracle, Meta and, reportedly, OpenAI. Oracle also plans to resell TPUs through its cloud infrastructure starting in 2027 under a multi-gigawatt-scale contract. The findings show that TPUs are increasingly competing with Nvidia accelerators on the cost of agentic inference, although independent testing is still needed.
Source: Read the original article
