Google built the first TPU because it calculated that if every Android user spoke to their phone for three minutes a day, it would need to double the size of its data centres. Custom silicon was cheaper than concrete.
Seven generations later the TPU is the only accelerator outside NVIDIA with a credible claim to frontier-scale training, and Ironwood pods reach 9,216 chips. Anthropic, Apple and Google itself all train on them.
You still cannot buy one. TPUs have never been sold as parts — they exist inside Google data centres and are reached only by renting Google Cloud capacity. Google has also never published a TOPS figure for any Tensor mobile chip, from G1 through G6.