Worldwide, demand is growing to make artificial intelligence (AI) lighter and more efficient. This trend is also evident in global acquisition battles for AI lightweighting platforms corporations. Nvidia acquired OmniML in 2023 and last year bought Deci AI for $300 million (about 442.5 billion won). Qualcomm acquired Edge Impulse in March. All the acquired corporations focus on AI lightweighting and optimization.
In Korea, there is a corporations competing with them. The main player is Nota AI, led by CEO Chae Myeong-su, 36. Meeting with ChosunBiz at the Nota AI office in Samseong-dong, Seoul, on the 12th, Chae said, "There was an acquisition offer from a big corporations, but we turned it down because our goal is to have AI run on every electronic device in the world, not to work for a single company."
In 2022, Nota AI developed and launched NetsPresso, an AI lightweighting and optimization platform that shrinks AI models to as little as one-tenth their size and delivers fast inference speeds. That is why NetsPresso is nicknamed an AI "diet pill." Nota AI's business is largely divided into two parts based on NetsPresso. The platform business provides software development kits (SDKs) to global semiconductor companies to optimize and lightweight AI models and semiconductor chips. Last month, Nota AI signed a contract to provide optimization technology for Samsung Electronics' latest application processor (AP), Exynos 2500.
The other pillar, the solutions business, aims to supply customized AI solutions to existing B2B (business-to-business) markets such as transportation, security and healthcare. Nota AI's April contract with the Dubai Roads and Transport Authority to supply a Generative AI-based intelligent transportation system (ITS) solution is an example. With the video monitoring solution provided by Nota AI, when a traffic accident occurs, the authority can automatically and easily receive an accident investigation report via CCTV.
A month after listing on the KOSDAQ market, Chae said, "These days it's hard to handle all the proposals coming into the company," adding, "The institutional sector platform is drawing more attention after the contract with Samsung Electronics, and the solutions institutional sector is also getting many inquiries from Southeast Asia and the United States as well as local governments after the contract with the Dubai Roads and Transport Authority."
Nota AI was founded in April 2015 by four people including Chief Technology Officer (CTO) Kim Tae-ho. It started as a company providing a solution to reduce keyboard typos. The company name also came from "no typos." But while developing a solution to reduce typos, privacy problems arose at first, followed by heat generation and processing speed issues. Chae recognized that there would be similar demand for optimization and lightweighting technologies when running AI on smartphones. So in 2018 it pivoted to an AI optimization corporations. The staff, which numbered four at founding, has now grown to about 140. The following is a Q&A with Chae.
◇ "Slim down AI models by 50%–60% with NetsPresso"
─Why didn't you start your own company from the beginning?
"I met Kim, the CTO, in 2017 after graduating from the Department of Industrial and Information Systems Engineering at Soongsil University and earning a master's in knowledge service engineering at KAIST. While serving alternative military duty at the KAIST AI Research Institute, my service overlapped with Kim's for a year. Our research chemistry was good then. We are also both friends born in 1989. In September 2018, I was preparing to study for a Ph.D. in Germany. I had a three-month gap, and Kim said the Nota AI business wasn't going well and three founders had left, and asked me to take over Nota AI. When I joined Nota AI, I was the only employee, and the office was a space of about 132 square feet on the KAIST Daejeon campus. I connected the solution to reduce typos to AI lightweighting and optimization as a way to solve its problems and thought demand would be big. Frankly, I had no work experience, so I never thought about starting my own company. In research, my role was to stand at the whiteboard and propose ideas, and Kim, the CTO, quickly implemented my ideas on the computer. That's still how we work today."
─The name "NetsPresso" is intriguing.
"Launched in 2022, NetsPresso reduces AI model size and computation so that high-performance AI can run even on low-spec devices. Named through an in-house contest, NetsPresso means compressing AI models like extracting only the essence from coffee beans for a neural network. With NetsPresso, corporations can reduce AI model size to under 10% at maximum. In that case, power consumption drops 30%–40%. Typically, AI models are lightweighted by about 50%–60%."
─There are many AI lightweighting corporations. What sets Nota AI apart?
"In AI optimization and lightweighting, 'know-how' matters more than a 'golden rule' about how you write code and make it work. Like master craftsmen whose work rookies cannot replicate at once, experience must accumulate. We are artisans who sculpt AI models. Nota AI has experience with hundreds of models across 37 semiconductors. Our key clients include Samsung Electronics, Qualcomm, ARM, Nvidia, Sony and Renesas."
◇ Responsibility grows after listing… aiming to turn profitable in 2027
─The stock that started at an offer price of 9,100 won is nearing 40,000 won.
"Listing on the KOSDAQ has made me more responsible and more cautious in decision-making. Fortunately, the stock has performed well since listing. When Nvidia CEO Jensen Huang visited Korea this year, CTO Kim Tae-ho brought an Nvidia DGX Spark and THOR to a chicken restaurant in Samseong-dong, Seoul, and, as a partner, offered congratulations on the listing and got the products signed."
─Sales are growing, but operating income is still in the red.
"Sales have been increasing since late 2023 as we expanded the business in earnest. Costs inevitably rose to develop products, but now, aside from labor costs, there are no major expenses. I think operating losses will improve significantly next year, and we will turn profitable in 2027."
─What are your short- and long-term plans?
"Right now, we discuss AI models individually, such as large language models (LLMs), lightweighting and optimization. But we believe there will soon be demand for lightweighting and optimizing AI systems, and we plan to develop related technologies. For example, an AI agent would operate with multiple AI models combined, such as voice and video. For On-device AI, lightweighting is needed because of device capacity limits, while the cloud will need lightweighting to cut costs. We plan to focus on this area."