They weren't costing Amazon any money if they had excess capacity, but if those machines being in use prevented them from being sold to someone else then they did cost it.

this happened to me on ec2 this summer- dozen idle instances running for two months because of a bug in a script.

I didn't get any sort of break though. :( June was a lot of hotdogs and Budweiser till I could dig out of that hole.

This didn't happen to me, but that's only because I set up dozens of post-its and alarms reminding me to quit those tests I was running on EC2. That's perhaps the one web-service that has made me the most paranoid.