in a very report, Groq states its LPUs are scalable and will be joined jointly applying optical interconnect across 264 chips. it could possibly further be scaled using switches, but it will add to latency. based on the https://www.sincerefans.com/blog/groq-funding-and-products