You use a Google TPU as a service. Google's documentation says TPUs are available through Google's own compute, Kubernetes and AI platform services, and we found no option to buy a TPU chip described in it.
What the documentation says
Google's TPU introduction lists the ways to use TPUs: through Google's virtual machine, Kubernetes and AI agent platform services. It points to pricing and a contact-sales page. For Ironwood, Google's TPU7x page lists reservations in All Capacity mode and says that for slices above 2,048 chips you should "contact your account team."
This is why TPUs are discussed as a vendor choice, not a part you can slot into your own servers. A workload written for them also has to go through XLA, so moving between TPU and GPU means a software change.
What to read next
See what a TPU is and whether PyTorch runs on TPUs. The competitor landscape is in NVIDIA's main competitors in AI, and the cost and architecture comparison is in TPU vs GPU.
