How to calm the AI bubble theory

For AI to thrive as an industry, more endeavors and achievements must be made in grooming data centers and AI applications.

Kim Yong-seokThe author is a professor at Gachon University and director of Gachon Semiconductor Education Institute. The arrival of an AI revolution through the November 2022 debut of Open AI’s generative AI model ChatGPT capable of processing human-like text and creative content has taken the world by storm. The euphoria also sent big-tech stocks grouped as the Magnificent Seven (M7) — Apple, Alphabet, Microsoft, Amazon, Meta, Tesla and Nvidia — to new heights.

But the M7’s shares have been skidding since the summer. Although they have recovered to some extent, scholars and investment banks alike have become level-headed and warn of an AI bubble crash as disastrous as the dot-com bubble crash. Tech behemoths like Amazon, Microsoft, Meta and Google have been trying to stay ahead in the game with data centers processing on Nvidia’s GPU chips to meet AI demand and supply. Their clients are data-based service providers like OpenAI which cannot foresee when it can turn into a for-profit company and make money out of astronomical expenses and investments. The sole beneficiaries of the AI craze are on-demand chipmakers Nvidia and TSMC responsible for the microchips enabling AI capabilities. SK hynix and Samsung Electronics, producers of high bandwidth memory (HBM) chips supporting complex algorithm tasks, are also enjoying some of the halo.

The talk of an AI bubble burst reminds me of my days with Samsung Electronics in the early 2000s when we were wrestling with a chipset to back mobile phone services. Telecom companies around the world have been working to standardize and commercialize the third-generation wireless communications network dubbed IMT-2000. The goal was to support data transmission rates of up to 2.4Mbps, or 2,400kbps, a leapfrog from the speed of 9.6kbps in voice communication via mobile network at the time. A transmission speed of over 2,000kbps would allow video calls.

The speed was achieved finally in 2006. The new technology was exalted to pave the way for science-fiction-like wireless service and serve as a goose laying golden eggs for the industry. But skeptics questioned the theory due to a lack of content services to make full use of the speed.

How many new subscribers of wireless operators could be lured through the new technology was also uncertain. The two options in the IMT standard of choosing either a synchronous or asynchronous transmission mode would burden both phone manufacturers and consumers.

The launch of IMT-2000, however, ended up adding impetus to the fledgling mobile internet services, giving birth to portal service providers like Naver in 1999, Facebook in 2004 and Twitter in 2006. The coming of smartphone age with Apple’s rollout of iPhone in 2007 gave the final push for the IMT-2000 wireless network to become mainstream.

It took at least five years for the so-called dream wireless technology to take root. New wireless technology and AI data centers are similar in the fact that they require heavy investments for their infrastructure. AI is still in the budding stage. It is premature to judge its value when the market is just at the onset of its development. Investment in AI therefore must not stop.

What needs to be done to create a mass market for AI? First, the machine learning model should shift from Large Language Model (LLM) to the compact version of small Large Language Models (sLLM). The latter can be crafted with fewer parameters and require less data for training, leading to reduced cost, capacity and power demand. Chips supporting AI tasks should move toward a neural processing unit (NPU) which uses less power than a CPU or a GPU.

Second, the industrial use of AI applications must speed up. That will be possible through expansion of on-device AI deployment. Such move is critical in creating an AI-based industry. AI must be embedded across the industry — personal computers, smartphones, home electronics and ground and aerial vehicles as well as factory and urban operations, medical and defense equipment. It should be employed in finance, law, education, marketing, sales, contents production, design and video games.

Third, AI must bring forth new services that consumers can easily access through mobile devices and computers. It would be too risky to create entirely novel services. Consumers will prefer AI-added services to what they are used to.

For AI to thrive as an industry, more endeavors and achievements must be made in cultivating data centers and AI applications. The two are essential to truly embrace an AI age.

Translation by the Korea JoongAng Daily staff.

For AI to thrive as an industry, more endeavors and achievements must be made in grooming datacenters and AI applications.