What Etched does
Etched is a private AI hardware company building “frontier inference clusters” for serving frontier AI models, with a specialization in transformer-family inference (prefill and decode). Etched’s approach is to co-design silicon, racks, software, and manufacturing methods so that its systems deliver throughput, latency, cost, and power efficiency for production inference at scale. The company describes its cluster architecture around two core technical elements: Low Voltage Inference (LVI), intended to increase FLOPs density without thermal throttling by running key compute blocks at under half the voltage of typical AI chips, and Cluster Scale Memory (CSM), intended to reduce decode-time memory bottlenecks via a shared low-latency memory pool across chips. Etched states that it targets both prefill and decode workloads, including large MoEs, long-context, and agentic workloads.
Strategically, Etched positions itself as an inference-systems company (not a standalone chip vendor): the company discusses validating rack-scale products with customers, shipping first racks, opening manufacturing/supply-chain capabilities (including a Taiwan factory), and booking customer demand in excess of $1B. As of its latest public updates in 2026, Etched reports having raised large rounds at rapidly increasing valuations and having shipped its first rack to Jane Street after that customer tested the hardware.