[amelius]

> In order to efficiently use TPUs, your code should build on the high-level Estimator abstraction.

Does this mean it's inference-only? (I only quickly scanned the article)

[jlebar]

No, this whole blog post is about training models.