In this case, I’d say it’s accumulating RAM or CPU-time - instances are kept alive during multiple requests. @KentonVarda said they would change CPU-time to not fail after accumulating (ex with multiple requests over time) and instead fail directly, but that might not have been changed yet.
For example, I saw 10K requests (10K users/unique IP) over 40 seconds work fine, but requests after that started to fail and ever subsequent request after that would also fail. It was due to CPU-time. Unfortunately, tests like that are expensive to create, which is why I only tested 10K at first and could have missed it.