The Tensor Processing Unit is Google's family of custom AI accelerators, first deployed internally in 2015 and announced publicly in 2016. TPUs are designed specifically for the matrix multiplication workloads that dominate neural network training and inference, and are optimized for Google's ML software stack — initially TensorFlow, now increasingly JAX.
TPUs are the only credible alternative to NVIDIA GPUs at frontier training scale. Google's internal models (Gemini) train primarily on TPU. Google Cloud offers TPU access to external customers, including Anthropic, whose April 2026 Broadcom deal secured multi- gigawatt TPU capacity for Claude training.
TPUs illustrate a strategic hedge: by maintaining a viable non- NVIDIA compute supply chain, Google and its cloud customers retain optionality that pure NVIDIA buyers lack. The tradeoff is ecosystem friction — most open-source ML code is CUDA-first — which limits TPU's appeal to organizations with the engineering capacity to adopt Google's ML stack.