As genetic AI touches a growing number of industries, the companies that make the chips to run the models stand to benefit immensely. Nvidia, in particular, wields enormous influence, running one is appreciated 70% to 95% of the market for AI chips. Cloud providers from After to Microsoft are spending billions of dollars on Nvidia GPUs, wary of being left behind in the AI race.
It’s understandable, then, that the AI vendor producers aren’t happy with the status quo. A large part of their success depends on the whims of the dominant chip makers. And so, along with opportunistic VCs, they are looking for promising startups to challenge the AI chip incumbents.
Engraved is among the many, many alternative chip companies vying for a seat at the table — but it’s also among the most interesting. Just two years old, Etched was founded by a pair of Harvard dropouts, Gavin Uberti (ex-OctoML and ex-Xnor.ai) and Chris Zhu, who along with Robert Wachen and former Cypress Semiconductor CTO Mark Ross, tried to to create a chip. which could do one thing: run AI models.
This is not unusual. Many startups and tech giants are developing chips that run dedicated artificial intelligence models, also known as inference chips. Meta has MTIA, Amazon has Graviton and Inferentia, and so on. But Etched chips is unique in that they only work one type of model: Transformers.
Transformer, proposed by a team of Google researchers in 2017, has become the dominant AI model architecture by a wide margin.
Transformers supports OpenAI’s Sora model that generates video. They are at the heart of text-generating models such as Anthropic’s Claude and Google’s Gemini. And they power art generators, like the latest version of Stable Diffusion.
“In 2022, we bet Transformers would take over the world,” Uberti, CEO of Etched, told TechCrunch. “We’ve reached a point in the evolution of artificial intelligence where specialized chips that can outperform general-purpose GPUs are inevitable — and the world’s technical decision makers know it.”
Etched’s chip, called Sohu, is an ASIC (application-specific integrated circuit) — a chip tailored for a specific application — built for transformer operation. It is made with use TSMC’s 4nm processSohu can deliver dramatically better inference performance than GPUs and other general-purpose AI chips while drawing less power, Uberti claims.
“Sohu is an order of magnitude faster and cheaper than even Nvidia’s next-generation Blackwell GB200 GPU when running text, image and video transforms,” Uberti said. “One Sohu server replaces 160 H100 GPUs. … Sohu will be a more affordable, efficient and environmentally friendly option for business leaders who need specialized chips.”
How does Sohu manage all this? In a few ways, but the most obvious (and intuitive) is an improved hardware and software inference pipeline. Because Sohu doesn’t run transformerless models, Etched’s team could eliminate non-transformer-related hardware components and cut the burden on software traditionally used to develop and operate non-transformers.
Etched arrives on the scene at an inflection point in the race for productive AI infrastructure. Cost concerns aside, GPUs and other hardware components necessary to run models at scale today are dangerously power-hungry.
Goldman Sachs predict that AI is poised to drive a 160% increase in data center electricity demand by 2030, contributing to a significant increase in greenhouse gas emissions. Researchers at UC Riverside, meanwhile, assessment that global use of artificial intelligence could cause data centers to absorb 1.1 trillion to 1.7 trillion gallons of fresh water by 2027; affect local resources. (Many data centers use water to cool servers.)
Uberti is optimistically—or bombastically, depending on how you interpret it—touting Sohu as the solution to the industry’s consumption problem.
“In short, our future customers will not be able to afford not to switch to Sohu,” Uberti said. “Companies are willing to bet on Etched because speed and cost are existential for the AI products they’re trying to build.”
But can Etched, assuming it fulfills its goal of bringing Sohu to the mass market in the coming months, succeed when so many others are following close behind?
The company doesn’t have a direct competitor at the moment, but recently AI chip startup Perceive previewed an editor with hardware acceleration for transformers. Groq has also invested heavily in transformer-specific optimizations for its ASIC.
Competition aside, what if one day the transformers fail? Uberti says that in this case, Etched will do the obvious: Design a new chip. Fair enough, but that’s a pretty drastic fallback considering how long it took Sohu to pull off.
However, none of these concerns have deterred investors from pouring a huge amount of money into Etched.
Today, Etched said it closed a $120 million Series A funding round, led by Primary Venture Partners and Positive Sum Ventures. Bringing Etched’s total raised to $125.36 million, the round was joined by heavyweight angel backers including Peter Thiel (Uberti, Zhu and Wachen are Thiel Fellowship seeds), GitHub CEO Thomas Dohmke, the co-founder of Cruise (and the Bot Company) and Kyle Vogt and Quora co-founder Charlie Cheever.
These investors apparently believe that Etched has a reasonable chance of successfully scaling its server business. Maybe so – Uberti claims unnamed customers have held “tens of millions of dollars” in hardware so far. The upcoming launch of the Sohu Developer Cloud, which will allow customers to preview Sohu through an online interactive playground, should boost sales, Uberti suggested.
But it seems too early to tell whether that will be enough to propel Etched and its 35-person team into the future its co-founders envision. The AI chip sector can be unforgiving at the best of times — see the high-profile near-failures of AI chip startups like Mythic and Graphcoreand decrease in investment in AI chip ventures in 2023;.
Uberti makes a strong sales pitch, however: “Video production, audio-to-audio functions, robotics and other future AI use cases will only be possible with a faster chip like Sohu. The entire future of AI technology will be shaped by whether the infrastructure can scale.”
