🚀 How to Win the Largest Market in AI

🚀 How to Win the Largest Market in AI
a16z partners Raghu Raghuram and Sarah Wang argue AI inference (not training) is becoming the largest and most stable computing workload in history, citing Google's jump to 3.2 quadrillion tokens processed per month by May 2026 (300x growth in two years). They contend general-purpose GPUs are increasingly mismatched to this workload since inference is memory-bound, making 'tokens per watt' the key economic metric, which favors purpose-built silicon over flexible chips. The piece profiles Etched, a startup founded by Harvard dropouts in 2022 that has built custom inference chips, hired 400+ engineers from Nvidia and Google's TPU team, and is now shipping production racks and standing up a 10-megawatt site. a16z frames the piece as its rationale for backing Etched, emphasizing that manufacturing at scale — not just chip design — is the real moat in AI infrastructure.
🔗 Read more