OP’s linked blog post mentioned they got hit with a large spike in bandwidth charges. Sending them garbage information costs money.

Yeah you have a point, hmmm, wish there were a way to somehow generate those garbages with minimum bandwidth. Something like, I can send you a very compressed 256 bytes of data which expands to something like 1 mega bytes.

Good ol' zip bomb

https://furry.engineer/@niko/113728467796605323

there is -- but instead of garbage expanding data, add in several delays within the response so that the data takes extraordinarily long

Depending on the number of simultaneous requesting connections, you may be able to do this without a significant change to your infrastructure. There are ways to do it that don't exhaust your number of (IP, port) available too, if that is an issue.

Then the hard part is deciding which connections to slow, but you can start with a proportional delay based on the number of bytes per source IP block or do it based on certain user agents. Might turn into a small arms race but it's a start.

It does not even have to be dynamically generated. Just pre-generate a few thousand static pages of AI slop and serve that. Probably cheaper than dynamic generation.

Tarpit instead? Trickle out a dead end response (no links) at bytes-per-second speeds until the bot times out.

https://en.wikipedia.org/wiki/Tarpit_(networking)