# Etched **Source:** https://geo.sig.ai/brands/etched **Vertical:** AI Infrastructure & Models **Subcategory:** AI Chips & Hardware **Tier:** Emerging **Website:** etched.com **Last Updated:** 2026-04-14 ## Summary Transformer-specific ASIC startup raised $500M at $5B valuation in Jan 2026; Sohu chip claims 20x Nvidia H100 inference speed for transformer workloads; fabricated on TSMC 4nm process alongside Apple and Nvidia silicon. ## Company Overview Etched is a semiconductor startup founded in 2022 that is building application-specific integrated circuits (ASICs) optimized exclusively for transformer-based neural network inference. Unlike general-purpose GPUs that must support a broad range of workloads, Etched's Sohu chip is hardwired at the silicon level to execute the transformer architecture — the mathematical backbone of virtually every major AI model including GPT, Gemini, and Claude. By eliminating the flexibility overhead of general-purpose hardware, Etched claims inference speeds up to 20x faster than Nvidia's H100 for transformer workloads, with corresponding reductions in cost per token.\n\nThe Sohu chip is fabricated on TSMC's 4nm process node, the same cutting-edge manufacturing technology used by Apple and Nvidia for their flagship chips. Etched targets large-scale inference deployments — hyperscalers, AI cloud providers, and enterprises running high-volume language model workloads where inference cost is the dominant operational expense. The chip is designed to slot into existing data center infrastructure and provide dramatic efficiency gains for organizations serving billions of AI queries daily.\n\nEtched raised $500M at a $5B valuation in January 2026, a financing round that placed it among the most highly valued AI chip startups globally. The raise reflects investor conviction that transformer inference will remain a dominant workload for years to come and that purpose-built silicon can capture significant market share from Nvidia in this specific segment. Etched is competing in the AI chip market alongside Google's TPUs, Amazon's Trainium/Inferentia, and startups like Groq and Cerebras. ## Frequently Asked Questions ### What is Etched's Sohu chip? Sohu is a custom ASIC fabricated on TSMC 4nm that hardwires transformer inference patterns, claiming 500K+ tokens/sec on Llama-70B vs 23K tokens/sec on an Nvidia H100 system. ### How much funding has Etched raised? Etched raised $500M in January 2026 at a $5B valuation, bringing total funding close to $1B. ### What is the tradeoff of Etched's approach? Sohu only runs transformer models, sacrificing flexibility for dramatic performance gains on the dominant AI architecture powering LLMs and image generators. ### What is Etched's approach to AI chip design? Etched is designing ASICs (application-specific integrated circuits) that are purpose-built exclusively for running transformer-based AI models. Unlike GPUs that handle many computational workloads, Etched's chips are hardwired to perform transformer operations, which the company claims enables dramatically higher performance and energy efficiency for inference. ### Why would a chip hardwired for transformers be advantageous? By eliminating the generality of GPUs and dedicating every transistor to transformer operations, Etched claims its Sohu chip can run large language model inference at speeds and efficiency levels that GPUs cannot match. If transformer architectures remain dominant in AI, this specialization could make Etched's chips the most cost-effective option for AI inference at scale. ### What is Etched's chip called and what are its claimed specs? Etched's chip is called Sohu. The company has claimed performance figures significantly exceeding NVIDIA H100 GPUs for transformer workloads, particularly for token generation in large language models. Sohu is designed for high-density server deployments for AI inference in data centers. ### What risk does Etched's architectural bet carry? Etched's chip is optimized exclusively for transformer architectures. If AI model architectures shift significantly away from transformers, Etched's hardware advantage could erode. The company is betting that transformers will remain the dominant paradigm for the foreseeable future, which most AI researchers currently believe but which is not guaranteed. ### Who are Etched's investors and target customers? Etched has raised venture funding from prominent AI and deep tech investors. Its target customers are hyperscalers, cloud providers, and enterprises running large-scale AI inference workloads where inference costs are a significant operational expense. The company positions its chips as a cost-effective alternative to GPU clusters for production AI serving. ## Tags ai-powered, b2b, infrastructure, saas --- *Data from geo.sig.ai Brand Intelligence Database. Updated 2026-04-14.*