TheTechno Optimist.

SohuEtched—Etched receives first A0 silicon from TSMC for Sohu transformer ASICSohuEtched—Etched raises $500M Series B at $5B valuationSohuEtched—Etched ships first Sohu racks to Jane Street quant fundSohuEtched—Etched closes $300M Series C at $10.3B valuation with SK Hynix investmentSohuEtched—Etched raises $700M Series D at $21B valuation led by Jane Street

Etched

Transformer ASIC, AI inference chip, Sohu

US$21B privateest. 2022etched.com (opens etched.com)Checked 10 Oct

The brief

Analyst view
  • Sohu ASIC: 500,000 tokens/sec on Llama 70B; 20x throughput vs. 8x H100 GPUs.
  • Built on TSMC N4P; proprietary low-voltage (LVI) and cluster-scale memory (CSM) techniques.
  • August 2026: Shipped first production racks to Jane Street; $21B valuation after $700M Series D.

Technical approach

As reported
Single-architecture design

Optimized for transformer inference only

Built for decoding tokens only. Instruction sets and memory paths designed purely for this task.

Low-voltage inference (LVI)

Math at half the voltage of competing chips

Runs floating-point ops at <50% voltage vs. other AI chips. Cuts power and heat.

Cluster-scale memory (CSM)

Shared memory with SRAM latency

Multiple chips share memory via proprietary interconnects with SRAM-level latency.

Primer