Etched

Etched is an AI chip startup building “frontier inference clusters” designed to accelerate inference (both prefill and decode) for frontier AI models. Instead of selling a general-purpose accelerator alone, Etched emphasizes end-to-end co-design across chips, packaging/racks, software, and manufacturing methods, aiming to improve throughput, latency, cost, and power efficiency for production deployments. Etched states that its cluster-scale approach is driven by two core technical ideas it developed: Low Voltage Inference (LVI), which Etched says increases FLOPs density without thermal throttling, and Cluster Scale Memory (CSM), which Etched says forms a lower-latency shared memory pool across chips to reduce decode bottlenecks. Etched also positions its hardware as “architecture-agnostic” at the system level, saying its inference clusters are intended to run models of different architectures (including transformers and other designs) and arbitrarily large parameter sizes. Recent company disclosures emphasize movement from silicon bring-up to rack-scale product validation and deployment with customers, including its first customer delivery to Jane Street.