Etched is taking a narrow route into one of technology’s most crowded and expensive markets. Instead of building a general AI accelerator, the two-year-old startup is designing Sohu, an AI chip made to run only transformer models.
That choice is the company’s central bet. Transformers now sit behind major generative AI systems for text, image and video, and Etched believes a chip built around that architecture can outperform broader hardware used across the AI industry.
Why Etched Is Focusing on Transformers
The market for AI chips is dominated by Nvidia, which commands an estimated 70% to 95% of the market for AI chips. Cloud providers from Meta to Microsoft are spending billions of dollars on Nvidia GPUs as they race to keep pace with generative AI demand.
That dependence has made hardware a strategic concern for AI vendors. Their ability to deliver products can be shaped by chip availability, cost and power consumption. It has also created an opening for startups trying to offer alternatives to incumbent suppliers.
Etched was founded by Gavin Uberti, Chris Zhu, Robert Wachen and former Cypress Semiconductor CTO Mark Ross. Uberti, who is Etched’s CEO, previously worked at OctoML and Xnor.ai. Uberti, Zhu and Wachen are Thiel Fellowship alums.
Many companies are working on inferencing chips, including Meta with MTIA and Amazon with Graviton and Inferentia. Etched differs by making Sohu useful for one model family only: transformers.
That is a risky narrowing of scope, but the source of the company’s claimed advantage. By not supporting non-transformer models, Etched says it can remove hardware that is not needed and reduce the software overhead normally required to deploy broader model types.
What Sohu Is Designed to Do
Sohu is an ASIC, or application-specific integrated circuit, built for transformer inference. It is manufactured using TSMC’s 4nm process.
Etched says that specialization allows Sohu to deliver much better inferencing performance than GPUs and other general-purpose AI chips while using less energy. Uberti’s claim is direct: Sohu is positioned as a faster, cheaper and more efficient option for businesses that need specialized AI hardware.
The company’s comparison is ambitious. Uberti said Sohu is “an order of magnitude faster and cheaper than even Nvidia’s next generation of Blackwell GB200 GPUs when running text, image and video transformers.” He also said, “One Sohu server replaces 160 H100 GPUs.”
Those claims matter because transformer models are central to many high-profile generative AI products. The transformer architecture, proposed by a team of Google researchers back in 2017, underpins OpenAI’s video-generating model Sora, text-generating models such as Anthropic’s Claude and Google’s Gemini, and art generators including the newest version of Stable Diffusion.
Etched’s argument is simple: if the dominant generative AI architecture remains transformers, then hardware optimized for transformers could be more practical than general-purpose chips for a large class of workloads.
The Infrastructure Pressure Behind the Bet
The timing is important. Generative AI is pushing demand for data center hardware, and the cost of running models at scale is becoming a major business issue. The source article points to two pressure points: power use and water use.
Goldman Sachs predicts that AI is poised to drive a 160% increase in data center electricity demand by 2030, with a related rise in greenhouse gas emissions. Researchers at UC Riverside estimate that global AI usage could lead data centers to consume 1.1 trillion to 1.7 trillion gallons of fresh water by 2027, affecting local resources because many data centers use water to cool servers.
Etched is pitching Sohu as an answer to that consumption problem. The company’s logic is that if a chip can do the same transformer inference work with fewer servers or less energy, it could help lower operating costs and infrastructure strain.
For AI companies, that is not a side issue. Faster and cheaper inference can influence whether a product is financially practical, especially when models must respond to users repeatedly and at scale.
The Risks of a One-Architecture Chip
The same specialization that makes Sohu interesting also creates its largest strategic risk. Sohu is built around transformers, so a major shift away from that architecture would reduce the value of the chip’s core design.
Uberti’s answer is that Etched would design a new chip if transformers fell out of favor. That is a clear fallback, but it would be a significant move given the time and effort already involved in bringing Sohu toward market.
Competition is another open question. Etched does not currently have a direct competitor in the same narrow category, but Perceive recently previewed a processor with hardware acceleration for transformers. Groq has also invested heavily in transformer-specific optimizations for its ASIC.
The broader AI chip market is also difficult. The source article points to the high-profile near-failures of Mythic and Graphcore, along with declining investment in AI chip ventures in 2023, as reminders that strong technical claims do not automatically translate into durable businesses.
Funding Gives Etched a Larger Runway
Investors are still backing the company aggressively. Etched said it closed a $120 million Series A funding round, co-led by Primary Venture Partners and Positive Sum Ventures. That brings the company’s total raised to $125.36 million.
The round included angel backers such as Peter Thiel, GitHub CEO Thomas Dohmke, Cruise and the Bot Company co-founder Kyle Vogt, and Quora co-founder Charlie Cheever.
Etched is also preparing the Sohu Developer Cloud, an online interactive playground that will let customers preview Sohu. Uberti said unnamed customers have reserved “tens of millions of dollars” in hardware so far.
The startup has 35 people, a focused chip design and a large market problem to attack. What remains uncertain is whether that focus will become a lasting advantage or a constraint. Etched is betting that the future of generative AI will continue to run through transformers, and that the companies building AI products will need hardware designed with that reality in mind.