South Korea's $246.6 Billion Wealth Fund Faces Scrutiny Over Chinese AI and Hardware Spending

Chinese models handle sensitive work at KIC

South Korea's sovereign wealth fund is facing scrutiny over its use of Chinese artificial intelligence models to process internal documents, including weekly reports, internal rules and material used to answer questions from lawmakers. The Korea Investment Corporation (KIC) managed $246.6 billion at the end of August 2026, equivalent to about 339 trillion won at the exchange rate it used.

Documents KIC submitted to National Assembly member Kang Min-kook show that Alibaba's Qwen3 handles language tasks, while BGE-M3, developed by the Beijing Academy of Artificial Intelligence (BAAI), supports document searches. The United States Commerce Department added BAAI to its Entity List on March 25, 2025, citing support for China's military modernization. KIC installed BGE-M3 on December 15 that year.

Kang, a member of the People Power Party serving on the assembly committee responsible for economic and fiscal affairs, said the selection criteria omitted the developers' nationality and their status under American export controls. KIC said it considered performance, operational suitability, stability, cost and the limits of its computing equipment.

The dispute also covers spending. KIC signed a contract worth 950.4 million won to triple its graphics processing units, or GPUs, from four to 12. Meanwhile, monthly use of its internal AI fell from 3,755 requests in May 2026 to 798 in September. The documents bring together two questions for the institution: how it protects sensitive information and how it matches technology investment to actual demand.

What the four models do

KIC's internal AI combines four models with different jobs. Qwen3 is the language model, the component that interprets written instructions and produces answers. BGE-M3 is an embedding model, which turns text into numerical representations so the system can find documents or passages related to a question.

Meta's Llama3 filters harmful content, while Google's Gemma3 processes images. Qwen3 is the only component KIC classified as the language model responsible for understanding documents and generating responses. The American models therefore perform supporting tasks, while the Chinese models handle language generation and information retrieval.

For an employee searching an internal rule, the retrieval model helps locate relevant material and the language model helps turn that material into an answer. KIC uses the system for document summaries, translation, draft preparation, internal rule searches, weekly reports and material concerning lawmakers' questions.

Those tasks place the system close to information about the institution's operations and its responses to parliamentary oversight. The location of the servers, any external connections and the controls governing access to documents were not disclosed in the details made public.

Why the export control listing matters

BAAI's inclusion on the American Entity List is central to Kang's criticism. The list places restrictions on exports, reexports and transfers of specified goods and technology to named organizations. For KIC's model selection process, it provides information about a developer that can be considered alongside technical performance.

KIC adopted BGE-M3 almost nine months after BAAI was listed. Kang said the documents submitted to him contained no separate selection criterion covering either the developer's nationality or its export control status. His objection concerns how KIC assessed a supplier connected to a system handling internal information.

Kang called for KIC to include those security factors in model selection and urged the ministry overseeing the fund to examine AI adoption across public institutions. His proposal would broaden procurement reviews beyond whether a model produces useful answers at an acceptable cost.

The practical security review would need to address how information moves through the system, who can access it and how software changes are checked. Those operational details matter because the same model can be deployed under different technical arrangements, with different connections and permissions.

How KIC chose Qwen3

KIC said it selected the models together with the company building its system. Its assessment used LMArena, which ranks models through user evaluations, and MMLU-Pro, a benchmark testing knowledge and reasoning across specialist subjects. It also considered accuracy, comprehension, reasoning, suitability for its work and stability.

Hardware capacity was another major factor. KIC initially had four GPUs available to run the system. GPUs are processors that carry out large numbers of calculations at once, making them useful for AI. Their memory limits the size of models and the amount of work they can handle at a given time.

According to KIC, Qwen3 performed better on value for money, available GPU memory and multilingual work involving Asian languages. For an institution translating and searching documents, those are practical considerations: a model must handle the relevant languages while fitting the equipment available.

Kang's criticism focuses on what was missing from that assessment. Performance tests and equipment requirements received attention, while developer nationality and export control listings were absent from the submitted criteria. The alternative considered for the BGE-M3 search model was also an Alibaba Qwen model.

The timeline from listing to expansion

KIC's decisions span more than a year, beginning with the American action against BAAI and continuing through the rollout and expansion of the fund's internal platform. The dates help explain why lawmakers are examining both model selection and procurement.

• March 25, 2025: The United States Commerce Department added BAAI to the Entity List.
• December 15, 2025: KIC installed BAAI's BGE-M3 model in its internal AI.
• March 16, 2026: An infrastructure expansion request for proposals cited increasing work demand.
• May 7, 2026: KIC made the internal AI available to all 350 employees and signed the 950.4 million won expansion contract.
• August 2026: KIC commissioned an internal AI platform improvement project that it later cited as a reason for the hardware expansion.
• September 2026: Internal AI usage totaled 798 requests from 147 users.
• November 6, 2026: The scheduled contract period for the infrastructure expansion ends.

KIC said the additional equipment supports the platform improvement project commissioned in August. It also cited rising GPU and memory prices and delivery times as reasons for ordering the equipment. The March procurement document had described growing work demand as the background for expansion.

Use fell after the staff rollout

KIC's internal AI recorded 3,755 requests in May, its first month of availability to all employees, and 798 in September. That is a decline of 78.7% over four months. The number of users fell from 339 to 147, a decrease of 56.6%.

Against the 350 employees covered by the launch, the May user count represented about 97% of staff. September's count represented 42%. The decline therefore involved both fewer people using the system and fewer requests from those who continued to use it.

The figures also show average monthly requests per user falling from about 11.1 to 5.4. That calculation uses the reported request totals divided by the user counts in each month. It gives a more detailed picture of the slowdown than either total alone.

KIC said its internal AI has a limited role supporting nine types of work. That narrower remit is part of its explanation for usage levels. For managers assessing demand, the next useful detail would be which of those tasks employees continue to use and which account for the decline.

ChatGPT use was about 46 times higher

KIC employees recorded 36,881 uses of ChatGPT in September, compared with 798 uses of the internal AI. The external service's total was about 46 times higher, making it a much more frequently used tool by that measure.

KIC's explanation points to differences in scope. Its internal platform supports a defined set of nine work categories, while ChatGPT serves a broader range of tasks. The figures show where employees directed their requests, while KIC's account explains why the services may attract different volumes.

For the fund, the combination creates a practical management issue. It is investing in an internal platform while employees make much heavier use of another service. Understanding the tasks behind those requests would help KIC decide where its own platform needs improvement and where additional computing capacity would be useful.

Why triple the equipment?

The expansion contract adds eight Nvidia H200 GPUs to KIC's existing four, bringing the total to 12. The contract value is 950.4 million won, with the work scheduled to run until November 6, 2026. KIC said it planned to execute the related budget after the project was completed in November.

The procurement documents list a budget of 950 million won. KIC linked the purchase to infrastructure needed for its platform improvement project, along with equipment prices and delivery schedules. Those explanations describe preparation for a future system rather than spending tied solely to September's request count.

Kang nevertheless questioned the contrast between falling usage and growing capacity. He called on KIC to check whether the expanded equipment would be used in line with actual demand, making utilization a central issue alongside security.

The figures set out a concrete test for the investment: GPU numbers are increasing by 200%, while monthly requests fell by 78.7% between May and September. KIC's planned platform changes will need to explain what work requires the additional capacity and how the institution will measure its use.

Government plans a move toward domestic models

The KIC dispute comes as South Korea's government promotes the replacement of foreign AI models with models developed domestically. Bae Kyung-hoon, deputy prime minister and minister of science and ICT, addressed the use of Chinese models during a parliamentary audit.

Bae acknowledged that many businesses and public bodies currently use Chinese AI. He said domestic models already installed in some platforms came from the first stage of the government's independent AI foundation model project. Foundation models are general AI systems that can be adapted to many tasks.

Bae said four domestic models from the project's second evaluation had undergone sufficient performance testing and would be introduced more widely in the public sector. He also described plans to expand use further when results from a third stage become available.

For KIC, any replacement would have to cover the work now divided between language generation, document retrieval, harmful content filtering and image processing. The fund's selection process already considers language performance, memory and cost. Kang is asking for security and developer related checks to become explicit parts of that decision as well.

Key Points

• KIC managed $246.6 billion, or about 339 trillion won, at the end of August 2026.
• Its internal AI uses Alibaba's Qwen3 for language tasks and BAAI's BGE-M3 for document retrieval.
• KIC installed BGE-M3 in December 2025, almost nine months after BAAI entered the American Entity List.
• Lawmaker Kang Min-kook said developer nationality and export control status were absent from the submitted model selection criteria.
• Internal AI requests fell from 3,755 in May 2026 to 798 in September, while users declined from 339 to 147.
• A 950.4 million won contract will increase KIC's H200 GPUs from four to 12, with the contract period ending November 6, 2026.
• KIC cited performance, hardware limits and future platform requirements in explaining its choices.