EtchedTranslation site

Company Overview

Etched is an American AI chip startup founded in 2022, headquartered in San Jose, California, and co-founded by three Harvard dropouts (Gavin Uberti, Chris Zhu, and Robert Wachen). The company's core team brings togetherchip design, top talent in AI algorithms and hardware engineering, including former Cypress Semiconductor CTO Mark Ross, 22-year NVIDIA veteran Brian Loiler, and others. As of January 2026, Etched has completed a cumulative total of nearly $1 billion in financing at a valuation of $5 billion, with investors including Peter Thiel, Stripes Group, and other well-known organizations.

Products & Services

Core Product: Sohu Chip

Sohu is the world's first ASIC (Application Specific Integrated Circuit) chip optimized for the Transformer architecture, launched by Etched, which is mainly aimed atAI macromodelReasoning Scenarios. Its core parameters are as follows:

• Processes: TSMC 4nm
• arithmetic power: 2000 TFLOPS at FP16 precision (single card performance equivalent to 20 NVIDIA H100 GPUs)
• Memory bandwidth: 144GB HBM3e video memory to support high-throughput data transfer
• efficiency ratio: Power consumption of only 10 watts, 20 times the throughput of the H100 when running models such as GPT-3, with only 351 TP4T more power consumption
• application scenario: AI tasks requiring low latency such as natural language processing, intelligent customer service, code generation, real-time video generation, etc.

core technology

1. Hardware-level Transformer optimization
   • Sohu hardens Transformer's attention mechanism, matrix multiplication, and other core operations into the chip through a customized architecture, avoiding the parallel computing bottlenecks that traditional GPUs have when dealing with these tasks.
   • The chip supports only Transformer models, shedding support for traditional AI models such as Graphics Rendering Units (GPUs) and CNN/RNN, thus dramatically simplifying the design and boosting hardware utilization to 901 TP4T (the average general-purpose GPU is only 301 TP4T).
2. Innovative Memory and Computing Architecture
   • A layered memory design combining HBM3e graphics and local cache reduces data handling latency.
   • With an innovative memory bandwidth allocation scheme, latency is reduced to 1/8th of H100 when batch processing 128 concurrent requests.
3. software ecosystem synergy
   • Self-developed programming framework supports open model formats such as ONNX, reducing developer migration costs.
   • Cooperating with cloud service providers to provide arithmetic leasing service, users can use Sohu chips directly through the cloud.

Market Competitiveness

1. Performance Crushes General Purpose GPUs
   • Real-world data shows that Sohu can process more than 500,000 tokens per second when running the Llama 70B model, which is 20 times that of the NVIDIA H100 and 10 times that of the Blackwell B200.
   • A server with 8 Sohu chips can replace 160 H100 GPUs, reducing hardware procurement costs by 60% and improving energy efficiency by an order of magnitude.
2. Pinpointing Vertical Markets
   • Focus on Transformer model inference to avoid head-to-head competition with NVIDIA in the general-purpose GPU space and create a differentiated advantage.
   • Target customers include AI big modeling companies, cloud service providers, autonomous driving companies, and other scenarios that are sensitive to inference performance and cost.
3. Ecological challenges and responses
   • CUDA Ecological Barriers: The current 95% AI development marketplace is based on the NVIDIA CUDA framework, and Etched needs to rebuild its developer community.
   • Supply Chain Risks: Dependence on TSMC's 4nm process, need to deal with global chip capacity allocation and geopolitical risks.
   • response strategy: Reduce user migration costs by opening hardware interfaces, supporting ONNX format, and providing cloud services.

development prospect

1. Industry Trend Dividend
   • AI model parameter scale continues to expand (e.g., GPT-4 reaches 1.8 trillion parameters), inference costs will account for more than 70% of the total cost of AI, and the demand for specialized chips is surging.
   • The development of multimodal large models (e.g., text, image, and video fusion) will drive the evolution of AI chips towards scenario specialization, and Sohu's technology path is in line with this trend.
2. Accelerated commercialization on the ground
   • It has received “tens of millions of dollars” in hardware pre-orders, and has partnered with Decart to launch the AI-generated game Oasis, which validates the chip's utility for real-time rendering of scenarios.
   • Plans to expand revenue streams by offering online arithmetic services through Sohu Developer Cloud.
3. Long-term strategic layout
   • In the future, we plan to expand into areas such as image generation and protein folding simulation to develop specialized chips for different AI models.
   • The goal is to become a “vertical giant” in the field of AI hardware and form a complementary market pattern with NVIDIA's general-purpose GPUs.

Etched has demonstrated strong competitiveness in the AI chip space with its specialized chip design and significant performance and cost advantages. Despite the challenges of ecological construction and mass production, the company is expected to become a leader in the dedicated AI chip market and drive the industry to evolve towards scenario specialization against the backdrop of the convergence of AI model architectures and surging demand for computing power. Its development path provides a new paradigm for AI hardware innovation, i.e., realizing complementary coexistence with general-purpose GPUs through vertical breakthroughs.