Etched, a two year-old startup, has raised $120 million to develop servers running on chips which can perform AI inferencing an order of magnitude faster than Nvidia’s upcoming Blackwell GB200 chip.
Etched’s ASIC, named Sohu and made on TSMC 4nm, is designed to run only one model – transformers – the neural net-type architecture which learns context and meaning by tracking sequences of words and can transform an input into an output.
“With over 500,000 tokens per second running Llama 70B, Sohu lets you build products that are 
impossible on GPUs,” says Etched, “one 8xSohu server replaces 160 H100s. Sobu is the first ASIC for transformer models. By specialising, we get way more performance: Sohu can’t run CNNs, LSTMs, SSMs, or any other AI models. Today, every major AI product (ChatGPT, Claude, Gemini, Sora) is powered by transformers. Within a few years, every large AI model will run on custom chips.”
Etched says its chip will be available in Q3 and that the company has “tens of millions of dollars” in reserved hardware sales.
The company will shortly be launching the Sohu Developer Cloud to let customers play with the tech in the expectation it will drive sales.
Backers include Peter Thiel, GitHub CEO Thomas Dohmke, Cruise co-founder Kyle Vogt, and Quora co-founder Charlie Cheever.
