Gavin Uberti

Gavin Uberti, 23, co-designed a chip called Sohu that he says can handle 10 times the throughput of a top Nvidia chip for a type of calculation used in generative AI applications like ChatGPT. Such an improvement would impact the performance of large language models being pushed by tech giants today and could spur advances in robotics, computer vision, molecular biology, and more.

Uberti’s approach is conceptually simple. Instead of using circuits able to handle many different tasks—as the H100 and other leading graphics processing units (GPUs) sold by current market leader Nvidia do—Sohu’s circuits can do only one thing, exceedingly well. As an application-specific integrated circuit, or ASIC, Sohu’s hardware specializes in matrix math calculations, which are central to an AI architecture known as a transformer.

“Back in 2022, we took a bet,” says Uberti, that the AI market would consolidate around transformers. “At the time, it was a risky bet,” he adds, since transformers were competing against many other architectures—like convolutional neural networks and long short-term memory—each being used for its own type of application. 

AI Chip Startup Etched Aims to Take On Nvidia

In the past two years, Uberti’s prediction has increasingly come to pass, as transformers have superseded other AI architectures. Now he’s doubling down, having dropped out of Harvard and cofounded a startup called Etched to produce the new chips. If the AI industry moves on from transformers, Etched will be toast. But if Uberti is right, Etched has the potential to become—or be purchased by—one of the largest businesses in the world. With $120 million in funds after a Series A round, and a partnership with chip giant TSMC to make his chips with their 4-nanometer process, Uberti is now focused on getting to market and scaling as fast as he can. 

A similar tale from the world of cryptocurrencies may prove apt: The early years of bitcoin calculations were dominated by generalist GPUs; today, narrowly tuned, energy-efficient ASICs have entirely taken over. Uberti is betting that story is about to repeat for AI.