Google's new Ironwood TPUs are set to challenge Nvidia's dominance in AI hardware. These accelerators offer 4.6 petaFLOPS of dense FP8 performance, slightly surpassing Nvidia's B200 GPU, and are just shy of the more powerful GB200 and GB300 models. Google has long relied on scaling its TPUs in large pods, which can contain as many as 9,216 chips. This approach contrasts with Nvidia's smaller, eight-way GPU boxes. Google's TPU v7 not only matches Nvidia’s performance but also brings a unique 3D torus topology for chip connectivity, reducing the need for high-performance switches that add latency. The TPU v7's architecture allows Google to build massive compute clusters, potentially supporting up to 400,000 accelerators. This capability has drawn interest from major AI model builders, including Anthropic, which plans to use a million TPUs for its Claude models. While Nvidia's CEO downplays the threat from Google's AI-specific chips, the increasing performance and scalability of alternatives from companies like Amazon and AMD suggest that the competition is intensifying. Google's switch to optical circuit switches for chip communication enhances fault tolerance and reduces latency, marking a significant evolution in TPU technology.