Harvard dropouts raise $120 million to take on Nvidia’s AI chips

Nvidia has been the hottest story on Wall Street of late, increasing eightfold in value since the end of 2022 and soaring past $3 trillion in market cap this month.

A 2-year-old startup founded by Harvard dropouts has just raised $120 million in venture funding to try and build a competitive chip and take on Nvidia in artificial intelligence.

Headquartered in Cupertino, California — home to Apple — Etched is developing a chip called Sohu, which the company says will be used to train and deploy AI models using “transformers,” the core architecture underpinning advancements like OpenAI’s ChatGPT.

Co-founder and CEO Gavin Uberti said that as AI develops, most of the technology’s power-hungry computing requirements will be filled by customized, hard-wired chips called ASICs. Their efficiency comes in executing only the AI model they were designed to perform, in contrast to general purpose graphics processing units (GPUs) from Nvidia that are more capable but are also much costlier.

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

Uberti and his co-founders are aware that it’s a high-risk wager, and that they’re going up against some of the most richly capitalized and competitive companies on the planet. While $120 million is a lot of money to raise in a series A, it’s about how much Nvidia generates in revenue in half a day. Nvidia’s sales have more than tripled annually for three consecutive quarters, topping $26 billion in the latest period.

Nvidia has more than 80% of the market share for AI chips, according to estimates. Etched is among a group of startups attracting capital to go after the burgeoning opportunity. Primary Venture Partners and Positive Sum Ventures led the round. Peter Thiel, Stanley Druckenmiller and Cruise founder Kyle Vogt are also backers.

Despite Nvidia’s head start and, what some developers describe a moat, new chipmakers are plowing ahead anyway, mainly because the opportunity is so big. Other chip startups taking on Nvidia include Cerebras Systems, which is building a physically larger AI chip, and Tenstorrent, which is using a trendy technology called RISC-V to build AI chips.

“The reason we were so excited about what we’re doing, why we dropped out of school and we’ve convinced so many people to leave these chip projects — this is the most important thing to be working on,” said Robert Wachen, Etched’s operating chief. “The entire future of technology is going to be shaped by whether the infrastructure can handle the scale.”

Semiconductors have traditionally been one of the hardest industries for startups given the long development cycles, the significant capital required to build a chip and the need to engage with a limited number of manufacturing partners, such as Taiwan Semiconductor Manufacturing Co., which is building Etched’s chip.

Venture capitalists invested $6 billion in AI semiconductor companies in 2023, up slightly from $5.7 billion in 2022, according to data from PitchBook.

Uberti and Chris Zhu started to work on a chip company after Uberti did a summer internship working on compliers. That put him in contact with the low-level hardware ideas that led to Etched.

The pair dropped out of Harvard in 2022, and added Uberti’s college roommate, Wachen. They quickly started hiring chip industry veterans. The company settled in Cupertino, and now has 35 employees. It offers housing stipends to new hires.

“When ChatGPT came out, and Nvidia stock exploded, and especially when every other model coming out would be a transformer too, we found ourselves in the right place at the right time,” Uberti said.

Etched is preparing to bring Sohu to market, and the founders say they’ll be ready to show something later this year. The startup is also working to secure customers and says that technology companies are eager to check out new AI chips.

For the business to work, companies that are spending billions of dollars on GPUs will need to see some significant savings in building custom chips designed for their specific AI model and be willing to make trade-offs in flexibility.

By specializing on transformers, which moves data in predictable ways from the chip to memory, Etched’s Sohu chip can dedicate less space to memory and more to the kinds of transistors that define a chip’s raw computing power, Uberti said.

Another aspect of Eteched’s efficiency is that the chip has one large core. That results in fewer inefficient calculations done by a part called a streaming multiprocessor to coordinate calculations by different cores.

Uberti says the impact of specialized AI chips could be similar to how custom chips called ASICs, first introduced in 2013 specifically for mining bitcoin or ether, reduced the demand for Nvidia GPUs.

The Etched founders expect the need for chips to run these models will increase, especially once they’re used to serve AI software millions of times per minute.

They also say that by hard coding the AI architecture into the chip, their device can reduce the latency of returning answers, unlocking new use cases, such as AI agents or real-time voice conversations. Etched says its chips are more than 10 times faster than Nvidia’s GPUs, thanks to its simpler architecture and single use case.

But Etched is taking on some of the most valuable companies in the world, including Nvidia, which have massive development teams and access to the capital needed to secure production and improve their chips on an annual basis.

Etched, which has a countdown timer displayed in its headquarters, has to move faster.

“The way that we’re going to win this specialized AI chip market, and the ones after this, is by being the first product to market,” Uberti said.