“AI chips from Groq and Cerebras, which use SRAM for memory, offer superior latency of around one millisecond per token.”