$1B in Transformer ASIC Orders Drives Etched to $5B Valuation

On June 30, 2026, silicon startup Etched emerged from stealth with $1 billion in contracted revenue for its custom AI hardware. The company reached a $5 billion post-money valuation after closing an unannounced $500 million round led by Stripes in December 2025. Etched is delivering first-pass (A0) silicon on TSMC’s N4P process, positioning its rack-scale systems as high-efficiency alternatives to Nvidia’s general-purpose GPUs.

Sohu Architecture Constraints

The company’s flagship Sohu chip is an Application-Specific Integrated Circuit (ASIC) hardcoded exclusively for transformer models. By stripping out the flexibility required to run other neural network architectures, Etched optimizes die space strictly for the matrix multiplications that power models like Llama, DeepSeek, and Qwen.

The company reports that a single 8-chip Sohu server processes 500,000 tokens per second when running Llama 70B. At this throughput volume, Etched calculates that one Sohu-equipped server effectively matches the output of 160 Nvidia H100 GPUs.

Power and Memory Infrastructure

Heat dissipation is a primary bottleneck in AI inference deployments. Etched mitigates thermal throttling through its Low-Voltage Inference (LVI) architecture. This proprietary design runs the chip’s math blocks at less than half the voltage of traditional AI hardware. The system maintains over 80 percent peak FLOPs during continuous operation, even when routing traffic through trillion-parameter Mixture-of-Experts (MoE) architectures.

For memory bandwidth, the Sohu chip implements Cluster-Scale Memory (CSM). This subsystem pairs HBM3E with SRAM to accelerate token decoding speeds across an entire cluster. To validate these components at scale, the company established a 2MW data center and a prototype manufacturing lab in San Jose, backed by a production factory in Taiwan. A strategic investment from the TSMC-linked VentureTech Alliance supports this manufacturing pipeline.

Production Scale

Etched has raised $800 million across four rounds and expanded to over 400 employees. The engineering team draws heavily from Nvidia, Broadcom, SK Hynix, and Google’s TPU division. Investors include Jane Street, Two Sigma, Jump Trading, and AI figures like Andrej Karpathy and Geoffrey Hinton.

The company’s $1 billion in booked contracts for its Frontier Inference Clusters reflects a broader hardware transition as production serving costs begin to outpace training expenditures. This shift previously drove the Cerebras Systems IPO and Nvidia’s $20 billion acquisition of Groq in late 2025.

Hardware specialization forces a strict architectural commitment. If you are building infrastructure for massive production scale, transformer ASICs alter your unit economics by orders of magnitude. You must weigh these cost reductions against the inability to run alternative architectures like state space models or emerging diffusion networks on the same cluster hardware.