The dashboard indicates it's based on 100 seconds, but my utility was generating quota exceeded errors with <200 queries / 100 seconds (well below the 1,000 requests / 100 seconds default quota), and this was resolved by increasing the quota, which would seem to indicate that the quota is actually per 1 second.
The quota team wrote that a 10 QPS quota is equivalent to a 1,000 queries per 100 seconds quota. That's only the case beyond 100 seconds; this has implications for short-lived batch utilities and exponential backoff, which isn't useful on quota timescales of 100 seconds, but would be for 1 second.
