AGPL doesn't pre-empt future uses or require permission for any kind of re-use. You just have to share alike. It's pretty simple.

AGPL lets you take a bunch of data and AI-train on it. You just have to release the data and source code to anyone who uses the model. Pretty simple. You don't have to rent them a bunch of GPUs.

Actually it can be annoying because of the specific mechanism by which you have to share alike - the program has to have a link to its own source code - you can't just offer the source alongside the binary. But it's doable.