Etched
Etched, founded in 2022 by Harvard dropouts Gavin Uberti, Chris Zhu, and Robert Wachen, designs and manufactures Sohu, a transformer-only ASIC for AI inference.
Profile
Etched builds a custom AI chip (Sohu) that runs only transformer-based neural networks, aiming to be faster and cheaper than Nvidia GPUs for inference.
Etched, founded in 2022 by Harvard dropouts Gavin Uberti, Chris Zhu, and Robert Wachen, designs and manufactures Sohu, a transformer-only ASIC for AI inference. The company raised $500 million in a January 2026 round led by Stripes and Peter Thiel, valuing it at $5 billion and bringing total disclosed funding to $620 million. Etched partners with TSMC’s Emerging Businesses Group for 4nm fabrication and has hired engineers from Broadcom and Cypress Semiconductor.
The company argues that AI scaling is limited by silicon capacity, not software, and positions Sohu as a specialized alternative to Nvidia’s general-purpose GPUs. Etched claims an eight-chip Sohu server can serve over 500,000 Llama 70B tokens per second, replacing 160 H100 GPUs. The company has not disclosed revenue, headcount, or named customers.
Its valuation and investor roster—including Thiel, Positive Sum, and Ribbit Capital—have fueled speculation about acquisition interest from AWS, Meta, xAI, Microsoft, Oracle, or OpenAI. Etched competes in a crowded field of over 100 private AI processor startups, many of which target Nvidia’s dominant market share. The company faces supply-chain risk given Nvidia’s priority access to HBM memory and TSMC capacity, and its single-architecture focus limits addressable workloads to transformer models only.
Products by Etched
Who buys this
- Large cloud service providers (e.g., AWS, Microsoft, Oracle) seeking specialized inference hardware
- AI model developers and labs (e.g., Meta, xAI, OpenAI) running large transformer-based LLMs
- Data center operators deploying high-throughput, low-latency inference infrastructure
- Enterprises with dedicated AI inference workloads that can be served by a single-architecture ASIC
Strengths and what to watch
Strengths
- Raised $500 million in January 2026 at a $5 billion valuation, with backing from Stripes, Peter Thiel, Positive Sum, and Ribbit Capital, signaling strong investor confidence.
- Sohu is a transformer-only ASIC fabricated on TSMC’s 4nm process, claiming an order-of-magnitude performance and cost advantage over Nvidia’s Blackwell GPUs for transformer inference.
- Partnership with TSMC’s Emerging Businesses Group and recruitment of senior engineers from Broadcom and Cypress provide credible manufacturing and design expertise.
Watch for
- Single-architecture risk: Sohu runs only transformer models, making it obsolete if AI model architectures shift away from transformers or if multi-model flexibility becomes a market requirement.
- Supply-chain vulnerability: Nvidia’s scale gives it priority access to HBM3E/HBM4 memory and TSMC capacity, potentially limiting Etched’s ability to secure enough components for volume production.
- No disclosed revenue or named customers: As of mid-2026, Etched has not announced any commercial deployments, raising questions about real-world adoption and revenue generation.
Recent moves
Key Information
- Industry
- AI Hardware
- Founded
- 2022
Frequently Asked Questions
What is Etched and what does the Sohu chip do?
Etched is an AI chip startup founded in 2022 by Harvard dropouts. It builds Sohu, a custom ASIC designed exclusively for transformer-based neural networks. Sohu aims to offer faster and cheaper inference than Nvidia GPUs by specializing in one model architecture.
How much funding has Etched raised and who invested?
Etched raised $500 million in a January 2026 round led by Stripes and Peter Thiel, valuing the company at $5 billion. Total disclosed funding is now $620 million. Other investors include Positive Sum and Ribbit Capital, signaling strong confidence in its specialized AI chip approach.
How does Etched's Sohu chip compare to Nvidia GPUs for inference?
Etched claims an eight-chip Sohu server can serve over 500,000 Llama 70B tokens per second, replacing 160 H100 GPUs. The company says its transformer-only ASIC offers an order-of-magnitude performance and cost advantage over Nvidia's Blackwell GPUs for inference workloads.
What are the risks of Etched's single-architecture chip design?
Sohu runs only transformer models, making Etched vulnerable if AI architectures shift away from transformers. This single-architecture focus limits addressable workloads and could render the chip obsolete if the market demands multi-model flexibility, posing a significant business risk.
Who are Etched's target customers for its AI chip?
Etched targets large cloud providers like AWS and Microsoft, AI labs such as Meta and OpenAI, data center operators, and enterprises with dedicated transformer inference workloads. These customers need high-throughput, low-latency hardware optimized for large language models.
Has Etched announced any customers or revenue yet?
As of mid-2026, Etched has not disclosed any revenue, headcount, or named customers. This lack of commercial deployments raises questions about real-world adoption, despite the company's $5 billion valuation and strong investor backing from prominent venture firms.
Sources
- www.etched.ai — Company homepage, product positioning, team photo, careers link
- www.jonpeddie.com — $500 million funding round led by Stripes and Peter Thiel, $5 billion valuation, total funding $620 million, Sohu ASIC details, TSMC partnership, potential acquirers
- finance.yahoo.com — $500 million raise, Stripes-led round, Peter Thiel participation, $5 billion valuation, total funding near $1 billion, TSMC partnership, engineers from Cypress and Broadcom
- www.datacenterdynamics.com — $500 million raise, $5 billion valuation, Stripes and Peter Thiel, Sohu chip details, TSMC 4nm process, claims of replacing 160 H100 GPUs, 500,000 Llama 70B tokens per second
- techcrunch.com — Etched listed among 17 US AI companies raising $100M+ in 2026