China's artificial intelligence (AI) startup Moonshot AI's open-weight model "Kimi K3" rattled Korea's stock market. By delivering performance close to U.S. frontier AI models at a much lower expense, it fueled doubts over whether big tech's astronomical AI capital expenditures (CAPEX) can continue to be justified.
Experts said the emergence of Chinese AI models could actually be a boon for infrastructure corporations such as chipmakers, and advised watching this week's big tech earnings. They pointed to cash flow, AI monetization, and CAPEX guidance as indicators to gauge AI's durability.
According to the Korea Exchange (KRX) on the 21st, the KOSPI index closed at 6,516.27, down 304.33 (4.46%) the previous day. After the release of "Kimi K3," AI-related stocks, including Samsung Electronics and SK hynix, broadly weakened.
◇ The essence of the "Kimi K3" shock… big tech's "economic moat" wobbles
At the center of this AI shock is the open-source large language model (LLM) "Kimi K3" unveiled by Chinese AI startup Moonshot AI. Kimi K3 is an ultra-large model with 2.8 trillion parameters, and is assessed to deliver performance comparable to global leading AI models ChatGPT and Claude.
The core concern raised by Kimi K3 is whether big tech AI models can truly build an "economic moat." An economic moat refers to a competitive edge that blocks rivals' entry and sustains high revenue over a long period, based on capabilities or network effects that are hard to replicate.
The AI industry's competitive landscape has recently shifted from model scale and benchmark races to how cheaply models can be used in real services and industrial settings. Accordingly, corporations are rapidly moving to a "multi-model" strategy, employing multiple AI models tailored to their purposes rather than relying on a single top-performance model.
In this context, the arrival of a high-performance open-weight model like Kimi K3 has put big tech's dominance to the test. Open-weight models disclose the model's weights so anyone can use and modify them, reducing reliance on a given corporation's closed model. On top of that, reports that Kimi K3 has significantly lower inference expense than leading U.S. models are fueling expectations that it could accelerate multimodal use.
As a result, cracks are forming in the prevailing notion that "the AI market will ultimately be winner-takes-all for big tech." As the premise that some big tech firms will dominate the market and reap massive revenue is shaken, questions are mounting over whether AI CAPEX, which is currently being poured in by the trillion-won scale, can truly be justified.
Kim Jae-seung, a researcher at Hyundai Motor Securities, said, "The emergence of high-performance open-weight AI models will accelerate the generalization of AI by prompting corporations that pay expense for AI models to selectively use multiple AI models," adding, "This reduces AI model corporations' monopoly power and limits future monetization."
◇ AI models and AI infrastructure should be distinguished
However, analysis is prevailing that AI model corporations should be viewed separately from AI infrastructure corporations such as chipmakers. Because Kimi K3 is also an ultra-large model, it requires massive memory and compute resources, and the drivers of AI investment are expanding from big tech-centered hyperscalers to general corporations.
Bloomberg said, "With 2.8 trillion parameters and a 1-million-token context window, K3 demands far more memory capacity than the previous generation," adding, "This suggests continued demand in a field dominated by a few players such as SK hynix and Samsung Electronics."
The shift in the center of AI CAPEX from hyperscalers to general corporations and AI cloud providers is also cited as a positive factor for AI infrastructure corporations.
Hwang Su-uk, a researcher at Meritz Securities, said, "The share of Nvidia's data center sales to hyperscalers is declining, while the shares to AI cloud and general corporations and enterprise customers are expanding," adding, "The center of AI CAPEX is shifting from hyperscalers to non-hyperscalers."
◇ AI super week… three things to check
Ultimately, whether Korea's stock market rebounds depends on this week's big tech earnings. Semiconductor corporations such as Samsung Electronics, SK hynix, and TSMC can keep improving results based on supplier advantage only if big tech's AI CAPEX continues.
Starting with Alphabet on the 22nd, Meta and Microsoft on the 29th, and Amazon on the 30th will announce earnings in succession. The market cites three key variables to check this earnings season: ▲ operating cash flow from existing businesses ▲ sales growth rate of AI cloud businesses ▲ next year's CAPEX guidance.
Kim said, "To gauge investment capacity, look at operating cash flow from existing businesses; to assess AI monetization, look at the sales growth rate of AI cloud businesses; and to see whether they intend to continue the AI investment cycle, check whether CAPEX guidance is being raised."
Experts especially stressed watching CAPEX guidance from Alphabet and Microsoft. While the two corporations continue to compete in AI, their CAPEX relative to operating cash flow is relatively conservative compared with Oracle, Amazon, and Meta. Analysts said if the two companies take an aggressive stance on AI investment, expectations for a sustained AI infrastructure investment cycle could grow.