“Dylan Patel believes that if future OpenAI models reach 10+ trillion parameters, they will not be able to fit on Cerebras's SRAM-based architecture, especially with long context lengths.”