Etched is building production-ready custom silicon and cluster-scale systems for the rapidly growing AI inference market

Etched is developing inference-focused hardware designed to maximize throughput and latency across frontier AI models, with custom silicon expected to ship in 2026. The company says its architecture incorporates low voltage inference and cluster-scale memory to support dense transformers, sparse MoEs, and emerging model designs like Mamba.

The announcement reflects a broader shift from general-purpose AI compute toward specialized infrastructure optimized for serving models at massive scale. If inference becomes the dominant AI workload, custom inference systems could become a major competitive frontier alongside model development itself.