• Sohu is an AI chip designed to train, deploy, and optimize transformer models, which is crucial for large language models like ChatGPT, Claude, and Gemini.
  • Sohu was built using Taiwan Semiconductor Manufacturing Corp.’s cutting-edge four-nanometer process.

An artificial intelligence chip startup, Etched.ai Inc., has raised USD 120 million in an early-stage funding round that aims to challenge Nvidia Corp. and lead the market in dedicated AI chips.

The two-year-old startup was founded by a pair of Harvard University dropouts who are confident they have the right solution to challenge Nvidia: a specialized chip designed to power AI models known as “transformers,” which are central to OpenAI’s ChatGPT and other chatbots.

Primary Venture Partners led the recent Series A round. It included contributions from several prominent individual investors, such as former PayPal Inc. CEO Peter Thiel and Replit Inc. CEO Amjad Masad. Although the company did not disclose a valuation following this round, it was previously valued at USD 34 million in March 2023 when it secured a USD 5.4 million seed investment.

Established in 2022 by CEO Gavin Uberti and Chief Technology Officer Chris Zhu, Etched has introduced its first product, Sohu. This dedicated AI chip is specifically engineered to train, deploy, and optimize transformer models. These models are pivotal in powering large language models like ChatGPT from OpenAI, Anthropic PBC’s Claude, Google LLC’s Gemini, and image generators such as DALL-E and Stable Diffusion.

An application-specific integrated circuit (ASIC) designed specifically to carry out a given task effectively is the Sohu chip. ASICs are designed to achieve optimal performance in terms of power consumption, speed, cost-effectiveness, or a blend of these criteria. By customizing ASICs for specific tasks, designers can maximize their performance capabilities and tailor them precisely to those functions.

For Sohu, its design focuses on optimizing the efficiency of running transformers. Utilizing Taiwan Semiconductor Manufacturing Corp.’s cutting-edge four-nanometer process, the chip aims to outperform Nvidia’s graphics processing units and other AI chips in inferencing performance. Additionally, the startup asserts that Sohu achieves this superior performance while consuming less energy.

Speaking with a leading media house, Uberti described Sohu as “order of magnitude faster and cheaper” for transformer-based text, image, and video generators compared to Nvidia’s forthcoming Blackwell GB200 GPUs, which are still in development.

“One Sohu server replaces 160 H100 GPUs. Sohu will be a more affordable, efficient and environmentally friendly option for business leaders that need specialized chips,” Uberti claimed.

Uberti stated that the company has collaborated with TSMC for chip manufacturing, emphasizing that the Series A funding is crucial for covering the substantial expenses involved in finalizing its chip designs with TSMC.

According to Uberti, the primary innovation of Sohu lies in its streamlined inferencing hardware and software pipeline. He elaborated that by focusing exclusively on Transformer models, his team could eliminate unnecessary hardware components. Additionally, they were able to reduce the software overhead associated with chips capable of deploying and running various workloads.

In an interview, Uberti mentioned that with the evolution and growth of the AI industry, there will be significant demand for specialized, energy-efficient AI chips. He noted that although Nvidia’s GPUs are highly powerful and versatile, their flexibility results in higher costs and increased energy consumption.

“We’re making the biggest bet in AI. If transformers go away, we’ll die. But if they stick around, we’re the biggest company of all time,” Uberti said.

Holger Mueller of Constellation Research Inc. remarked that considering the widespread adoption of ASICs in various computing domains, it was inevitable for someone to undertake the task of developing an ASIC specifically designed for transformers.

The analyst said, “Etched.ai’s decision could be seen as somewhat risky, as there is the danger that transformers may one day be replaced by a superior model architecture, but at present, there are more than enough transformer workloads around to justify the demand for something like Sohu.”

An Abundance of AI Chip Startups

Nvidia commands about 80% of the AI chip market share. Since the onset of the generative AI revolution in late 2022, Nvidia’s stock has soared, becoming one of Wall Street’s most coveted assets, with its value increasing more than ninefold in the past year. Recently, Nvidia briefly overtook Microsoft Corp. to claim the title of the world’s most valuable company.

Nvidia is a powerful competitor with a plethora of resources, but there are many upstarts eager to challenge it. Numerous semiconductor startups have gained attention as a result of investors’ belief that there is room for multiple major suppliers of AI chips. For example, it’s been reported that Cerebras Systems Inc., the company that developed the dinner plate-sized chip, is preparing for an IPO later this year. Tenstorrent Inc. is developing AI chips that make use of the RISC-V open-source architecture.

Other competitors include Taalas Inc., which uses chips designed to run particular neural networks to do something quite similar to Etched. It concluded a funding round of USD 50 million in March, and edge AI chipmaker SiMA.ai Inc. closed a USD 70 million round in April. Most recently, DEEPX Co. Ltd., a South Korean startup, announced last month that it had raised USD 80.5 million in an attempt to commercialize its AI chips. These chips combine a four-core central processing unit with a set of circuits optimized for machine learning workloads.

The AI chip market is poised for significant upcoming competition. Uberti asserts that Etched is well-positioned, having already secured orders from several undisclosed customers totaling “tens of millions of dollars” for its chips. Additionally, the startup plans to introduce the Sohu Developer Cloud platform in the coming weeks. This platform will allow potential customers to explore Etched’s hardware capabilities and experiment with customizations tailored to their AI models. Etched anticipates that the platform’s launch will further expand its customer base.

The abundance of startups vying for market share in the AI chip space believes the reality that breaking into the semiconductor market is among the most difficult. Making chips requires a substantial capital investment, lengthy development cycles, and agreements with a select group of highly sought-after chip manufacturers, like TSMC or Intel Corp.

Fortunately for these startups, investors are optimistic that Nvidia’s current dominant market position may not be permanent.