Tensor Processing Unit (TPU)

Google's custom-built chips, a type of application-specific integrated circuit, designed to speed up machine learning work. TPUs are built for the large matrix operations common in machine learning, carry high-bandwidth memory on the chip, and can be linked in groups called slices to scale a workload up. Google's guidance is that they suit large models dominated by matrix calculations and training runs lasting weeks or months, while models with many custom operations may run better on GPUs or CPUs, and work needing high-precision arithmetic is not a good fit. Businesses use TPUs through Google Cloud services such as Compute Engine and Google Kubernetes Engine. Frontier labs rent them at very large scale: Reuters reported in October 2026 that Anthropic has committed $125.2bn to a five-year lease of TPU capacity, chips Google has developed with Broadcom over several generations.