Sovereign Wealth Fund Managing 339 Trillion Won Relies on Chinese Core AI… Criticized for "Lax Security Awareness"

Desk

korocamia@naver.com | 2026-10-19 08:54:54


It has been revealed that the Korea Investment Corporation (KIC), a sovereign wealth fund managing 339 trillion won in assets, relies entirely on Chinese AI models for internal tasks ranging from document summarization and translation to internal data searches.

Criticism is mounting that the organization—which manages national assets—must tighten security reviews during its AI adoption process, as the deployed models include those developed by a research institute subject to U.S. government export controls.

According to data submitted by the KIC to Representative Kang Min-guk of the People Power Party and the National Assembly’s Strategy and Finance Committee on October 19, among the four AI models embedded in KIC's internal system, the language model was Alibaba's "Qwen3" from China, and the search embedding model was the Beijing Academy of Artificial Intelligence’s (BAAI) "BGE-M3."

The remaining two models are Meta’s "Llama3" and Google’s "Gemma3," which are responsible for blocking harmful content and processing images, respectively. KIC classified only Qwen3 as the "language model" that understands documents and generates responses.

Consequently, tasks using the internal AI—such as document summarization and translation, handling internal groupware regulations and weekly reports, and processing materials for parliamentary inquiries—all pass through Chinese models.

BAAI, the developer of the search model, was placed on the U.S. Department of Commerce's export control list on March 25 of last year due to its support for China's military modernization. KIC integrated BGE-M3 into its internal AI on December 15 of that year. KIC has been providing the internal AI service to all 350 employees since May 7 of this year.

Regarding the selection of these models, KIC explained that it considered accuracy, comprehension, reasoning, job suitability, and stability, with limited hardware conditions—specifically constraints of four Graphics Processing Units (GPUs)—serving as a major decision factor. Regarding the evaluation results, KIC stated that Qwen3 had the upper hand based on cost-effectiveness, VRAM availability, and suitability for multilingual (Asian region) use.

However, Representative Kang Min-guk pointed out that the submitted evaluation criteria did not include the developer's nationality or whether they were subject to U.S. export controls.

Controversy also surrounds the mismatch between internal AI usage performance and the scale of hardware expansion.

On May 7, the launch date of the AI service, KIC signed an infrastructure expansion contract worth 950.4 million won. It plans to complete the project next month to expand its GPUs from 4 to 12. KIC explained that the infrastructure expansion was for an internal AI platform upgrade project ordered in August of this year, and that the order was placed considering rising GPU and memory prices as well as delivery lead times.

Conversely, AI usage dropped sharply by 78.7% in just four months, from 3,755 cases in May, the first month of the service, to 798 cases in September. The number of users also fell by 56.6%, from 339 to 147 over the same period. Last month, KIC employees used ChatGPT 36,881 times—about 46 times the usage of the internal AI. However, KIC explained that its internal AI is used in a limited capacity to assist with nine specific types of tasks.

"While evaluating performance and cost-effectiveness, the developer's nationality and sanction status were excluded from the criteria, and equipment was tripled even as internal AI usage dropped to one-fifth," Representative Kang criticized. "This is evidence of both lax security awareness and poor budget management."

He urged, "An institution handling core national assets should naturally include security factors such as the developer's nationality and sanction status in its selection criteria. KIC must check whether the increased equipment is actually being used to meet real demand, and the Ministry of Economy and Finance, as the competent ministry, should inspect the AI model adoption status of its affiliated public institutions and establish clear nationality and security standards."

[ⓒ Global Economic Times. 무단전재-재배포 금지]

WEEKLY HOT