AI chip startup raises $120 million to take on Nvidia

Company claims its algorithm-specific chips are significantly faster than Nvidia’s upcoming Blackwell offering

Artificial intelligence (AI) chip startup Etched.ai has raised $120 million in a Series A funding round.

The round was led by Primary Venture Partners and saw participation from former PayPal CEO Peter Thiel and Amjad Masad, the CEO of AI software development platform Replit.

Founded in 2022 by Harvard University dropouts Gavin Uberti and Chris Zhu, the Cupertino-based company has developed an AI chip dubbed Sohu to train, deploy, and optimize transformer models – a type of neural network architecture that underpins large language models (LLMs) such as ChatGPT.

The chips will be produced by TSMC using the manufacturer’s 4nm process and the company says it has also secured high bandwidth memory and server supply from “top vendors”, although those companies were not named.

According to Etched, while its algorithm-specific ASIC chips won’t be able to run most traditional AI models such as deep learning recommendation models (DLRMS) or recurrent neural networks (RNNs), Sohu will be more efficient than traditional GPUs as they execute only the specific transformer models they were designed to perform.

The company says Sohu supports all currently available models, including Google, Meta, Microsoft, OpenAI, and Anthropic, and can handle tweaks to future models.

Etched also claims that Sohu is an “order of magnitude faster and cheaper” than Nvidia’s upcoming Blackwell GPUs, with an eight-chip Sohu server serving more than 500,000 Llama 70B tokens per second. The company has made this determination by extrapolating data from published MLperf benchmarks for Nvidia H100 servers, which showed an eight GPU server can serve 23,000 Llama 70B tokens per second.

In an interview with TechCrunch, Etched’s CEO Uberti said that one Sohu server would replace 160 H100 GPUs.

“In 2022, we made a bet that transformers would take over the world,” he told the news outlet. “We’ve hit a point in the evolution of AI where specialized chips that can perform better than general-purpose GPUs are inevitable — and the technical decision-makers of the world know this.”