Maybe the plan is to compete on latency? Say someone wants to regularly extract content from the same kind of document, and wants it fast, like 500 milliseconds.