What Etched does
Etched is a private semiconductor company building “frontier inference clusters” for AI inference workloads. Instead of selling only general-purpose AI accelerators, Etched designs and co-optimizes chips together with rack-scale systems and the supporting software + manufacturing approach, targeting high throughput/low latency and inference cost/power efficiency for prefill and decode phases of large-model inference.
More
The company says its architecture is built around two differentiators: Low Voltage Inference (LVI) for sustaining high FLOPs utilization without thermal throttling, and Cluster Scale Memory (CSM), a shared low-latency memory subsystem intended to reduce memory/interconnect bottlenecks for decode. Etched’s public product framing emphasizes complete rack-scale deployments (co-designing “chips, racks, software, and manufacturing methods”), and it highlights customer-driven validation—most recently shipping its first rack to Jane Street and describing first-pass silicon success and active production ramping toward gigawatt-scale throughput. Etched is currently positioned as a rapidly scaling, post-stealth inference-hardware entrant competing against GPU-centric inference systems by attempting to deliver better tokens-per-watt and lower sustained-cost performance via custom inference-centric hardware and cluster-level memory/power design.