Etched Sohu Chip

ASIC chip purpose-built exclusively for transformer inference at extreme speed

Code & DevelopmentPaid

Etched's Sohu is an application-specific integrated circuit designed entirely for transformer model inference, discarding flexibility for raw performance. By burning the transformer architecture into silicon, Sohu delivers orders-of-magnitude greater token throughput per watt compared to GPU-based inference, enabling real-time applications that current GPU clusters cannot serve economically. Etched targets AI inference providers and hyperscalers with multi-year capacity reservations.

Key Features

• ✓Transformer-only ASIC
• ✓Extreme token throughput
• ✓Energy efficiency
• ✓H100 outperformance claim
• ✓Rack-scale deployment
• ✓Hyperscaler targeting