Cerebras has significant multi-year upside because its wafer-scale chip does fast inference better than anything else, and latency-sensitive inference is growing about 2x faster than training. Holding the whole model in on-chip SRAM gives up to 15x faster inference than leading GPU solutions; it took OpenAI's fast-inference business from NVIDIA, and management has already raised prices on excess demand. A $20B take-or-pay OpenAI deal for 750MW through 2028 anchors $24.6B of RPO. x.com x.comÂ
