The Iron(wood) Throne: Google's TPU Stakes Its Claim on Inference
Independent benchmarks showing Google's TPUs outperforming NVIDIA GPUs in cost-efficiency for realistic agentic workloads—combined with rapid open-source software support and adoption by major labs like Meta, OpenAI, and Anthropic—signal a credible threat to NVIDIA's inference and software moat.
What happened
SemiAnalysis released AgentX benchmark results for long-context, multi-turn agentic inference, showing that Google's TPUv7 (Ironwood) achieves either competitive or superior performance per dollar compared to NVIDIA's Blackwell GPUs and AMD's MI355X across models like Qwen 3.5, GLM-5.2, Kimi K3, and Meta's Muse Spark Open. The report highlights that Google is expanding its software ecosystem by providing open-source maintainers at vLLM, SGLang, and PyTorch with direct hardware access, while NVIDIA faces growing internal bureaucracy.
Who it’s about
- NVIDIA
- TPUv7
- SemiAnalysis
- Meta