Google TPUs on OCI: Capacity, Scheduling, and Abstraction Before 2027
Oracle plans to offer Google TPUs on OCI from 2027. That single sentence, reported on 28 October 2026, is the entire confirmed claim: no named TPU generation, no region list, no pricing, and no statement about whether buyers reserve capacity or request it on demand. This short works through what to do with a fact that thin. It covers why accelerator capacity planning is really a negotiation over allocation shapes, why a TPU slice is a topology rather than a device, how the job spec ends up naming a geometry instead of a chip, and where the checkpoint interval arithmetic lands once preemption is treated as a scheduled event rather than an incident. The position argued in the post: architect for abstraction at the orchestration layer now, keep the compiler layer vendor-specific, and treat the 2027 date as a reason not to foreclose an option rather than a reason to commit budget. Every performance number and topology sketch in the piece is an assumption to be replaced, not a specification. Full write-up: https://hackyjs.com/posts/google-tpus-on-oci-capacity-scheduling-and-abstraction-before-2027