About a year back I tried to build a few latency-sensitive products on Groq, and found the response times would vary so wildly that it just about erased any benefit in their throughput. Is that still this case?