South Korea joins list of countries blocking DeepSeek access to ministries

South Korea’s finance ministry has revealed plans to block access to the Chinese AI service DeepSeek, citing concerns over its data collection practices, according to a report by Yonhap News Agency on Thursday.

 

“Due to multiple technical concerns raised about DeepSeek from home and abroad, we plan to block access for the service on PCs connected to external networks,” the official mentioned. This announcement comes a day after South Korea restricted access to DeepSeek on computers used by officials in the foreign, trade, and defence ministries.

 

The country’s unification ministry is also considering similar restrictions on AI services, including DeepSeek. A separate official from the ministry revealed, “The unification ministry has (since 2023) prohibited the input of undisclosed official data into all generative AI at the request of the National Intelligence Service and the interior ministry.”

 

 

When asked whether DeepSeek would be blocked, the official added, “We plan to take follow-up measures within the day, including blocking access,” without directly naming the service.

 

In addition to these steps, South Korea’s Ministry of Environment also joined the growing list of government bodies imposing restrictions on DeepSeek. On Thursday morning, the ministry blocked access to the AI service from its internet-connected PCs starting at 9 am local time. A ministry official noted, “The intelligence authorities have asked for caution in using DeepSeek, as its personal information collection system is not yet clearly known.”

 

Global trend of tightening DeepSeek controls

 

The South Korean government’s actions are part of a broader international trend of tightening controls on DeepSeek. On January 31, Taiwan’s Ministry of Digital Affairs (MODA) also prohibited government employees from using the service, citing concerns that it could expose sensitive data to Beijing. MODA classified DeepSeek as a Chinese information and communications technology (ICT) product, warning that data leaks could threaten national security.

 

Italy’s data protection authority has taken similar measures, blocking DeepSeek due to concerns about the transparency of its personal data practices. Both the US and Australia have also raised privacy alarms about the AI service.

 

In a related development, Howard Lutnick, the nominee for US Commerce Secretary, accused DeepSeek of stealing US technology and circumventing US export controls to acquire Nvidia chips. During his confirmation hearing before the US Senate on January 29, Lutnick alleged that DeepSeek had been able to develop its AI models at a significantly lower cost by purchasing large quantities of Nvidia chips and exploiting data from Meta’s open platform.

 

Lutnick emphasised, “I take a very jaundiced view of China. They only think about themselves and seek to harm us, and so we need to protect ourselves. We need to drive our innovation, and we need to stop helping them. Meta's open platform lets DeepSeek rely on it. Nvidia's chips - which they bought tons of, and they found their ways around [export controls] - drive their DeepSeek model. It’s got to end.”

 

DeepSeek unveils advanced AI models V3, R1

 

DeepSeek has launched two core AI models: DeepSeek-V3 and DeepSeek R1. The V3 model is a large language model (LLM) that utilizes a Mixture-of-Experts (MOE) architecture, which distributes tasks across smaller models to optimize performance.

 

With a total of 671 billion parameters, the V3 model activates 37 billion at any given time. It also integrates Multi-Head Latent Attention (MHLA) to reduce memory usage and uses mixed-precision training with FP8 computation for improved efficiency. The training of the V3 model involved 2,048 NVIDIA H800 GPUs over two months, requiring around 2.7 million GPU hours for pre-training and 2.8 million GPU hours overall.

 

While the estimated training cost stands at $5 million, based on a rental rate of $2 per GPU hour, Bernstein’s report suggests this estimate overlooks critical factors such as research, experimentation, and infrastructure costs.

 

[With inputs from agencies]