Source: SemiAnalysis
Google's TPU strategy is no longer about hardware alone. It's embedding itself into the open-source inference stack through partnerships with PyTorch, vLLM, SGLang, and Ray. TPUv7 (Ironwood) delivers 50% better performance-per-dollar than NVIDIA's B200/B300 on realistic agentic workloads. Major inference buyers—OpenAI, Anthropic, Meta, Oracle—are now standardizing on TPU infrastructure rather than treating it as an alternative. Google is making it frictionless to build and deploy on TPU by owning the software layer that developers touch. That's the competitive threat, and it mirrors how NVIDIA built its decade-long lead.