Reminds me of when hedge funds started laying increasingly shorter fiber-optic cable lines to achieve the lowest possible latency for high-frequency trading.

I thought these TPUs were primarily used for inference?

TPU8t is for training. But even still, once you’ve trained, you need to run the model too. And these kinds of models already have a huge latency hit so there’s not much hurting running it away from the trading switches.

As the article states, there's both training and inference dedicated chips.