Etched has completed A0 stepping tape-out for its self-developed inference accelerator, with first rack-scale systems already built. Backed by **over $1 billion in customer orders** and $800M in Series B funding, the company plans summer 2026 delivery.

The industry's dirty secret: most AI chips throttle under heat, delivering barely half of peak theoretical throughput in real-world inference. Etched's chip, built on TSMC N4P, tackles this head-on with a low-voltage architecture – achieved through co-optimization of circuit design, packaging, and scheduling – cutting operating voltage by over 50% versus mainstream competitors.

The result: when running trillion-parameter sparse MoE models, the chip sustains >80% compute utilization – dramatically reducing thermal-induced performance loss.

On memory, Etched deploys a hybrid on-chip SRAM + external HBM solution with a proprietary high-bandwidth interconnect. SRAM delivers ultra-low latency, while HBM provides large capacity – balancing response speed and memory footprint to boost throughput and conversational fluidity for large-model inference.

ICgoodFind Takeaway:
80%+ real-world utilization on MoE models is a game-changer. Etched isn't just another ASIC – it's solving the thermal wall that plagues every hyperscaler's inference fleet. If summer delivery holds, incumbent GPU vendors will feel real pressure in the token economy.