Google pushes in-house TPUs as alternative to Nvidia GPUs in AI data centers

Google has spent over a decade developing its own Tensor Processing Units while remaining a major Nvidia customer: Nvidia's data center segment brought in $193.7 billion of its $215.9 billion fiscal-year revenue. Each new TPU generation, including Ironwood, lets Google shift more AI workloads off Nvidia GPUs onto its own platform.

• Nvidia fiscal-year revenue hit $215.9 billion, with $193.7 billion from data centers
• Google has developed TPUs for over a decade; Ironwood is its seventh generation
• Google Cloud sells both Nvidia GB300 systems and its own TPUs
• Inference is moving to custom chips faster than model training