Etched Brand Guidelines: The Sohu Transformer Chip and a San Jose Inference Bet

Harvard dropouts and a single-purpose chip.

Etched is an American chip start-up that designs hardware for running AI models, the work known as inference, rather than for training them. It was founded by Gavin Uberti and Chris Zhu, two Harvard dropouts, TechCrunch reported in June 2024, with Robert Wachen and Mark Ross, a former chief technology officer of Cypress Semiconductor; Uberti had previously worked at OctoML and Xnor.ai. The company’s website today lists Uberti as co-founder and chief executive, Wachen as co-founder and president, Zhu as co-founder and Ross as chief technology officer.

Sources differ on the year. eeNews Europe described Etched in June 2024 as “a 2022 startup” based in Cupertino, California, and TechCrunch called it “only two years old,” quoting Uberti: “In 2022, we made a bet that transformers would take over the world.” SiliconANGLE wrote in August 2026 that Etched launched in 2021. eeNews Europe named the company Etched.ai Inc.; its website is now etched.com.

The bet was on one kind of model. The transformer, an architecture proposed by Google researchers in 2017, underlies most large language models, and Etched set out to build a chip that ran transformers and nothing else, leaving out the hardware a general-purpose processor needs for other kinds of model.

Sohu, then racks.

That chip was Sohu, announced in June 2024 alongside a Series A funding round. TechCrunch described it as an ASIC, an application-specific integrated circuit, made for running transformers and manufactured on TSMC’s 4-nanometre process. eeNews Europe reported the company’s own statement that Sohu could not run convolutional or recurrent neural networks, recommendation models or protein-folding models such as AlphaFold 2. Etched made large performance claims for it at the time, which both publications reported as the company’s claims.

The company was open about the risk of such a narrow design. eeNews Europe quoted its website at the time: “If transformers are replaced by SSMs, RWKV, or any new architecture, our chips will be useless.” Asked by TechCrunch what would happen if transformers fell out of favour, Uberti said Etched would design a new chip. The publication noted that other companies, among them Groq, were also tuning their hardware for transformers.

The company’s direction later widened. SiliconANGLE reported in August 2026 that Etched had pivoted to inference accelerators that support a variety of model architectures. On 30 June 2026 Etched described itself as building a new category of hardware, “frontier inference clusters,” in which it designs the chips, racks, software and manufacturing methods together; The Next Web reported that day that its first silicon, an A0 design made on TSMC’s N4P process, had come back working, and that systems were already running models including DeepSeek, Qwen, Mamba and Llama.

The company frames itself as a maker as much as a designer. Its website names two techniques it has developed, Low Voltage Inference and Cluster Scale Memory, and describes custom packages, circuit boards, cold plates and interconnects. The Next Web reported that the company says it has set up a factory in Taiwan and built a data centre, a test house and a prototyping lab at its San Jose headquarters, and quoted Wachen: “Production is the product.” The team had grown to more than 400 people, drawn, the company says, from Nvidia, Broadcom, Google’s TPU programme and SK Hynix, among others; Stripes had led its Series B round. A strategic investment from VentureTech Alliance, a fund linked to TSMC, was part of the funding the company announced in June 2026.

Design and documentation.

AI chip companies sell something most of their customers will never see. The processor sits inside a server, the server sits in a rack, and the rack sits in a data hall far from the people who use the models running on it. So the industry’s visual language has grown up around making that hidden hardware visible: a whole silicon wafer, as wide as a dinner plate, held up on a keynote stage; a single chip pinched between finger and thumb for the cameras; a die shot recoloured in neon false colour so that its blocks of memory and compute read like a city map at night.

The second set of images is the data centre. Press photographs tend to show long rows of racks in dark halls, lit by the glow of status lights and cold aisles, with cabling arranged as neatly as a diagram. These pictures stand in for scale, which is what buyers of AI compute are being asked to imagine, and they appear again and again in launch announcements, partner news and investor material.

Keynote slides do much of the remaining work. Model names, token counts and benchmark charts appear beside the chip’s own product name, and the companies’ identities are usually kept spare to let those numbers and images lead: black or white grounds, a single accent colour, a sans-serif wordmark and very little ornament. Several firms have adopted coined product categories, names for their processors that set them apart from the graphics processor.

The shop window is the technical conference. Hot Chips, NeurIPS and the AI Infra Summit are among the events where architectures are presented, demonstrations run and booths compete for engineers’ attention. This guide looks at how each company in the group presents itself, and keeps to what can be checked about who founded it, who owns it and what it makes.

A first rack, and a mark kept in two parts.

Etched shipped its first rack in 2026. On 18 August 2026 the company announced that it had delivered it to the trading firm Jane Street, which led a new funding round after testing the hardware; SiliconANGLE reported that Jane Street, Etched’s first customer, had installed the rack in its data centre in July, and that Etched had built a two-megawatt AI cluster at its San Jose headquarters where prospective customers test its chips. Uberti told the publication: “It took us three years to deliver our first rack from scratch.” Earlier that summer the company had announced a round led by Sequoia Capital. Etched remains privately held, with investors that include Sequoia Capital, Andreessen Horowitz, Kleiner Perkins and Peter Thiel, according to its website.

The brand itself is spare. The company’s media kit offers its logos in vector form and photographs of its hardware, and sets one rule: the logomark and the wordmark should always be displayed separately and never combined. Its website pairs plain statements of mission — “to run the world’s inference” — with long technical explanations and a footer line, “Etched is faster.” It has not published a colour standard or a typeface.

For a design guide, Etched shows how a young hardware company can build its identity on a single technical bet and then rewrite the product story while keeping the name. Compare TSMC, the foundry that makes its silicon, and Groq, another American start-up building chips for inference. For this guide we use black, a bone white and a silver grey with a sans, an unofficial interpretation; no product or logo is reproduced.

Keep Etched’s identity on hand — save it to a board in the Made Good inspiration library.