Been reading more about TPUs lately and I’m curious how many people are genuinely using them outside of demos or research projects.
Most AI infrastructure conversations still end up centered around GPUs, but it feels like Google Cloud keeps pushing TPUs further into real-world workloads now, especially for larger inference and training setups.
I haven’t had the chance to work with them deeply yet, so I’m wondering:
-
what kind of workloads are you using TPUs for?
-
was setup/migration harder than expected?
-
any noticeable cost or performance advantages?
-
anything you miss from regular GPU environments?
Would love to hear practical experiences rather than benchmark numbers 