At one time, extremely spiky usage: my primary use-case was short-term compute cluster deployments for specific projects, so it would fluctuate wildly. Typical sizes at any given time were 0 (no project), 16 (small cluster), 64 (moderate) and 128 (decent, at least for the particular domain I was in).

These days it's zero, because I work at an HPC-targeted IaaS startup that hosts our own hardware. (Because Infiniband is awesome when you're latency-bound.) But I used some decent-size ec2 deployments for a bit there.