Google Achieves Strong Performance with New TPUv7 FP8 vLLM Results

2026-10-15

Google has reported its initial TPUv7 FP8 vLLM performance metrics for the Agentic Inference Benchmark, showing a 1.67x improvement in cost-effectiveness compared to GB300 NVL72 FP4. This marks a significant step in Google making its TPU technology accessible externally.