News
IndustryTSMC’s grip on advanced chipmaking will face competitionIndustryCadence joins European Chip Design Platform to support semiconductor startups and SMEsMaking chipsMarc Hijink on ASML’s Culture, Supply Chain, and GeopoliticsProcessorsIntel's post-Nova Lake CPUs take shape as Razor Lake desktop and laptop variants surfaceAI chipsHow Google’s New AI Chip Software Helps It Battle NvidiaIndiaMore IC Packaging Plants Emerge In IndiaIndustryUS Tightens China AI Chip Access as Nvidia, TSMC and Semiconductor Stocks SlideDesignAxelera AI moves into HPC with TitaniaMaking chipsChinese envoy says chip equipment export curbs threaten EU-China trade thawIndiaSEMICON India 2026 Day 2 Sees 25 MoUs, Strategic Collaborations Across Semiconductor Value ChainIndustryTSMC’s grip on advanced chipmaking will face competitionIndustryCadence joins European Chip Design Platform to support semiconductor startups and SMEsMaking chipsMarc Hijink on ASML’s Culture, Supply Chain, and GeopoliticsProcessorsIntel's post-Nova Lake CPUs take shape as Razor Lake desktop and laptop variants surfaceAI chipsHow Google’s New AI Chip Software Helps It Battle NvidiaIndiaMore IC Packaging Plants Emerge In IndiaIndustryUS Tightens China AI Chip Access as Nvidia, TSMC and Semiconductor Stocks SlideDesignAxelera AI moves into HPC with TitaniaMaking chipsChinese envoy says chip equipment export curbs threaten EU-China trade thawIndiaSEMICON India 2026 Day 2 Sees 25 MoUs, Strategic Collaborations Across Semiconductor Value Chain
EveryChip
Sohu

Sohu

Transformer-specific AI accelerator chip by Etched.ai

Sohu is an application-specific integrated circuit designed by Etched.ai to accelerate transformer models used in large language models. Announced after a 2022 founding, it targets training and deployment with claims of higher efficiency than general-purpose graphics processors.

Share on WhatsApp

  • Built on TSMC 4nm process node
  • Designed exclusively for transformer models
  • Claims to replace 160 H100 GPUs per server
  • Raised $120 million in Series A funding
  • Supports models from Google, Meta, and OpenAI

What it is

Sohu is a specialised semiconductor chip developed by the startup Etched.ai. It functions as an application-specific integrated circuit, meaning it is built to perform one type of calculation extremely well rather than many different tasks. The chip focuses exclusively on transformer models, which are the neural network structures behind modern large language models. Etched.ai states that this narrow focus allows Sohu to operate with greater speed and lower cost than general-purpose graphics processing units when handling these specific workloads. The company positions the hardware for both training new models and deploying them for inference in data centres.

How it came about

Etched.ai was established in 2022 by Gavin Uberti and Chris Zhu, who left Harvard University to start the firm. Based in Cupertino, the company raised 120 million dollars in Series A funding to develop its hardware. Investors included Primary Venture Partners, Peter Thiel, and Amjad Masad. The founders bet early that transformer architectures would dominate artificial intelligence development. This strategic decision led to the creation of Sohu, a chip designed from the ground up to support this specific algorithmic approach rather than adapting existing general-purpose silicon.

Inside the chip

The architecture of Sohu is optimised for transformer models, excluding support for other common AI structures like recurrent neural networks or deep learning recommendation models. By hardwiring the logic for transformers, the chip avoids the overhead associated with general-purpose processors. Etched claims the design can accommodate current models from major developers such as Google, Meta, and OpenAI, as well as future variations. The system relies on high bandwidth memory to feed data quickly to the processing units. This specialisation aims to deliver an order of magnitude improvement in performance and cost efficiency compared to standard graphics cards.

How it was made

TSMC manufactures the Sohu chips using its four nanometre process node. This advanced fabrication technique allows for high transistor density and power efficiency. Etched.ai has also secured supplies of high bandwidth memory and server components from unnamed top-tier vendors. The combination of the four nanometre silicon and specialised memory architecture is intended to maximise throughput for language model tasks. The physical production is handled entirely by the foundry partner, while Etched manages the design and integration of the final server units.

Where it went

Sohu is intended for use in data centres that train and run large language models. The company highlights its ability to serve tokens for models like Llama 70B at high rates. An eight-chip server configuration is claimed to handle over 500,000 tokens per second. This performance target suggests the chip is suited for enterprise-level AI infrastructure where inference speed is critical. The hardware replaces traditional graphics processing unit clusters in these specific transformer-based workflows, offering a dedicated solution for organisations building or deploying generative AI applications.

Versions and family

The available information describes Sohu as a single product line focused on transformer acceleration. No distinct variants, siblings, or predecessor chips are mentioned in the provided material. The design is presented as a unified solution for all currently available transformer models. Etched.ai emphasises that the chip can handle tweaks to future models without requiring a new hardware generation immediately.

Why it matters

Etched.ai claims that Sohu represents a shift towards specialised silicon in the AI industry. The company asserts that one Sohu server can replace 160 Nvidia H100 graphics processing units for specific tasks. This claim challenges the dominance of general-purpose GPUs in the machine learning market. If accurate, such efficiency gains could lower the cost of running large language models significantly. The startup argues that technical decision-makers recognise the inevitability of specialised chips as AI models grow larger and more complex.

Problems and reception

The company bases its performance comparisons on extrapolations from published MLperf benchmarks for Nvidia H100 servers. Etched states that its eight-chip server outperforms an eight-GPU H100 server by a wide margin in token generation. These claims have been reported in technology news outlets but rely on the company's own projections rather than independent third-party testing results. The funding round led by Primary Venture Partners indicates investor confidence in the specialised approach, though broader market reception awaits real-world deployment data.

Questions people ask about Sohu

What type of AI models does Sohu support?

Sohu supports transformer models, which are used in large language models. It does not run other types like recurrent neural networks or deep learning recommendation models. The chip handles current models from major developers and adapts to future tweaks.

Who manufactures the Sohu chip?

TSMC manufactures the Sohu chip using its four nanometre process node. Etched.ai designs the chip and secures additional components like high bandwidth memory from other unnamed top vendors for the final server assembly.

How does Sohu compare to Nvidia GPUs?

Etched claims Sohu is an order of magnitude faster and cheaper than Nvidia Blackwell GPUs for transformer tasks. They state one eight-chip Sohu server can replace 160 Nvidia H100 GPUs, based on extrapolated benchmark data.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Sohu yet, so the facts on this page come from the sources listed above: geeky-gadgets.com/etched-sohu-ai-chip/, sdxcentral.com/news/ai-chip-startup-raises-120-million-to-take-on-nvid. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryChip wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    Photo credits on this page

    The small pictures above come from Wikimedia Commons under free licences. Some are cropped to fit.

    • Tensor Processing Unit: Norman P. Jouppi, George Kurian, Sheng Li, Peter Ma, Rahul Nagarajan,, CC BY 4.0, source

    The weekly digest

    Every Monday: the chip of the day, new pages on EveryChip, and the conferences and anniversaries of the week ahead.