[Anchorwoman]

It is said that the cost of purchasing graphics cards and building data centers to develop high-performance AI models amounts to tens or hundreds of trillions of won.

However, if you do not stick to the 'giant model', there is a way to utilize AI without spending a huge amount of money.

Reporter Park Jin-jun reports.

[Reporter]

MWC, the world's largest mobile communications technology exhibition.

Questions from guests continue to arise at the exhibition hall of a domestic semiconductor development startup.

[Overseas Buyer]

"Does this product only work with the built-in processor?"

A 2cm wide x 2cm AI-specific semiconductor.

If you plug this chip into a home computer, it becomes a high-performance computer that can implement AI technology.

[Director Yoon Sang-hyun/AI semiconductor development company]

"You can think of it as a solution that has been optimized for AI and can now perform parallel operations well."

Large language models such as Chat GPT and Gemini require data to travel long distances to the server, which requires a lot of electricity and high-performance graphics cards to operate the server.

However, if smartphones, drones, robots, and electronic products are equipped with AI semiconductors, less electricity will be consumed and data will not need to be moved.

For small AI technologies that do not require large servers, dedicated semiconductors are used to increase economic efficiency.

Generative AI developed by a domestic startup.

When you command it to create a video of a teddy bear playing the guitar, a cute teddy bear is created in just 30 seconds.

This technology won an innovation award at CES, the world's largest IT exhibition where AI was a hot topic.

The AI ​​was made lightweight so that it can run on regular CPUs, not expensive graphics cards, or even on home PCs.

[CEO Seungjae Yoo/AI program development company]

"We quantify and quantify, remove unnecessary data, and utilize various AI techniques such as pruning to leave only the data that is really useful and necessary."

The value of AI increases as its utility in daily life increases.

DeepSec shock led the government to spend 2 trillion won to buy 20,000 high-performance graphic cards, but some point out that this is not the only way to secure AI competitiveness.

[Yoo Eung-jun/Professor, Korea Advanced Institute of Science and Technology]

"Whether it's chips for smartphones, cars, or drones, the market could be much larger than the existing training market."

Instead of focusing solely on developing large language models that require a lot of high-performance semiconductors, the idea is to target niche markets by developing AI-specific semiconductors or small language models.

This is Park Jin-jun from MBC News.

- Original article: https://imnews.imbc.com/replay/2025/nwdesk/article/6694075_36799.html