[INTERVIEW] Korea's Sapeon looks to chip away at Nvidia's market share

Sapeon's Chief Technology Officer Michael Shebanow holds up the company's latest X330 NPU. [SK TELECOM]

Korean AI chip designer Sapeon is aiming to tackle Nvidia in the ever-expanding AI chip market, according to its Chief Technology Officer (CTO) Michael Shebanow.

And to do so, Sapeon plans to narrow its focus in on the AI inference market, rather than becoming an all-rounder — at least for now.

SK's Sapeon launches new AI chip for data centers

"We might [take market share from Nividia], taking some of the expanding market," the CTO said in a group interview with the Korea JoongAng Daily on Friday at the Sapeon's Korean office in Pangyo, Gyeonggi.

"There's a lot of people who just want to do inference. We are pretty narrowly focused. A small company can't take on Nvidia in every possible place that Nvidia is competing. We have to focus on something we can succeed in, and we are focused on AI inference."

Inference is applying information learned through training to actual work.

Sapeon is a Silicon Valley-based Korean AI chip designer that was founded in 2022 as a spinoff of SK Telecom. It had been invested in with a combined 80 billion won by the so-called SK ICT Alliance that comprises of SK Telecom, SK hynix and SK Square.

The company recently launched the latest X330 AI NPU (neural processing unit) that is used for data centers. The chip manufactured with TSMC's 7-nanometer technology has roughly twice the computational capability compared to its rival product launched this year and four times that of its own previous version X220, according to the company.

It is slated for mass production in the first half of next year.

Shebanow joined the Korean company in 2022 after working as a chip designer at Nvidia, AMD, Samsung Electronics and several other startups.

Shebanow saw an opportunity in Sapeon as its smaller size meant room for more innovation while still not having to worry as much about financial risks as it was backed by SK Group, the No. 2 conglomerate in Korea by market cap.

"I had some ideas on the type of processors I think we should build. And I'm going to try and convince people that it's a good idea here. And that would be something you probably couldn't do at a big company," Shebanow said.

"At big companies, they are kind of captured by their own success. Nvidia builds GPUs [graphics processing units], and it has a tremendous infrastructure in building GPUs. They have to build on GPUs and taking a break out of that, what they are doing, is a lot of work."

Sapeon's lineup is dedicated to NPUs at the moment, which are the type of chipset that emulate the neural networks of a human brain. They are said to be more suitable for processing AI compared to GPU because they are more purpose-built, enabling them to be more cost and energy efficient compared to GPUs, which serve a more general purpose and were originally designed for graphics and gaming.

Shebanow emphasizes the total cost of ownership of its chips as a key competitive point over its rivals.

"Our focus is on how much of the overall thing will cost the customer. Not just when they buy the card but when they operate the card. There are the card cost, the maintenance cost and the operational cost. We think if we'll be providing value in this thinking, we don't have to reduce the card cost. We are not going to compete on price, but we will be competing on value," the CTO said.

Sapeon has an ambitious goal set for the coming years as well: launching the X340, which is tailored for self-driving cars, next year and the X430 in 2026.

Sapeon is vying to go public, and the company said it will begin talks with potential investors from next year.

BY JIN EUN-SOO [jin.eunsoo@joongang.co.kr]