Skip to main content

Etched

Category: AI Infrastructure

A specialized semiconductor company developing the world's first Transformer-specific ASIC (Application-Specific Integrated Circuit) designed to run large language models with unprecedented speed and efficiency. Etched was founded in 2022. The company is led by Gavin Uberti. Based in Cupertino, USA. Team size: 35. Total funding raised: ~$620M+ total (all rounds). Latest round: Series A ($500M, Jan 2026 at $5B valuation led by Stripes). Key investors include ["Primary Venture Partners","Positive Sum","Peter Thiel","Stanley Druckenmiller","David Bonderman","Thomas Dohmke","Sky9 Capital"].

Founded
2022
Headquarters
Cupertino, USA
Team size
35
Total funding
~$620M+ total (all rounds)

Value proposition

Delivers up to 20x higher throughput and significantly lower latency compared to general-purpose GPUs (like Nvidia's H100/B200) by hard-coding the Transformer architecture directly into the silicon.

Products and solutions

["Sohu: The flagship Transformer-specific ASIC chip","Etched Software Stack: Specialized compilers and SDKs for deploying LLMs on Sohu hardware","Sohu Developer Cloud: Remote access for developers to test and optimize models"]

Unique value

Unlike general-purpose GPUs that are designed to handle various mathematical tasks, Etched 'burns' the Transformer architecture into the hardware logic, eliminating the overhead of programmability to maximize compute density for LLMs.

Target customer

AI labs (OpenAI, Anthropic, Google), hyperscale cloud providers, and enterprises deploying massive-scale Transformer models.

Industries served

["Artificial Intelligence","Semiconductors","Cloud Infrastructure","Enterprise Software"]

Technology advantage

By focusing exclusively on Transformers, the Sohu chip can dedicate nearly all its transistors to the specific matrix multiplications and attention mechanisms used in modern AI, achieving performance-per-watt and cost-efficiencies that general-purpose chips cannot match.

How they differentiate

Etched develops the world's first Transformer-specific ASIC (Sohu), which 'burns' the Transformer architecture directly into the silicon. Unlike Nvidia's general-purpose GPUs or Groq's LPU, Etched's hardware is non-programmable for other architectures, allowing it to achieve 20x higher throughput for LLMs by dedicating nearly all transistors to Transformer-specific matrix multiplications.

Main competitors

["Nvidia","Groq","Cerebras Systems","SambaNova Systems"]

Key partnerships

["TSMC (Manufacturing partner for advanced process nodes)","Primary Venture Partners","Positive Sum","Strategic individual investors including Peter Thiel, Stanley Druckenmiller, and Thomas Dohmke (GitHub CEO)"]

Notable customers

["Major AI Labs (OpenAI, Anthropic, Google)","Hyperscale Cloud Providers"]

Major milestones

["Successful tape-out of the Sohu Transformer ASIC","Raised $120M Series A led by Primary Venture Partners in June 2024","Secured $500M in Series B funding in January 2026 to scale mass production and challenge Nvidia's Blackwell dominance"]

Growth metrics

Scaled from a founding team of Harvard dropouts to over 35 senior engineers from Apple, Intel, and Nvidia; moved from architectural design to mass production readiness following the 2026 capital injection.

Market positioning

High-performance AI infrastructure challenger targeting the LLM inference market.

Geographic focus

Global (Headquartered in Cupertino, USA; targeting global hyperscalers and AI labs).

Patents and IP

Proprietary ASIC architecture and specialized matmul kernel designs; specific patent filings are generally kept confidential during the pre-launch phase but involve custom interconnects and memory management for Transformer weights.

About Gavin Uberti

Gavin Uberti is the Co-founder and CEO of Etched. He was previously a compiler engineer at OctoML, where he specialized in high-performance computing and matmul kernels for the Apache TVM project. He is a 2024 Thiel Fellow and a former student at Harvard University, where he studied Mathematics and Computer Science before dropping out to focus on building specialized AI hardware.

Official website: