The South Korean government plans to support the 'Cybersecurity Specialized AI Foundation Model Project' and the 'Everyone's AI Project' with approximately 256 graphics processing units (GPUs) each. This is significantly lower than the minimum of 1,000 GPUs provided to elite teams in last year's independent AI foundation model project, raising concerns about the feasibility of building a free AI chatbot for the entire population.
According to the IT industry on August 6, the Ministry of Science and ICT and the National IT Industry Promotion Agency (NIPA) published revised announcements for the 'Universal Utilization Support for National AI Services (Everyone's AI)' project and the 'Cybersecurity Specialized AI Foundation Model Development' project on July 29.
The Everyone's AI project aims to provide unlimited and free AI chatbots to all citizens, utilizing 512 B200 GPUs owned by the government. If two companies are selected, they will each receive 256 GPUs; if three companies are chosen, the top company will get 256 GPUs, while the second and third will receive 128 GPUs each.
Based on the market rental price for B200 GPUs (approximately 6.6 million won per month), the annual value of 256 GPUs is estimated at about 20.2 billion won, meaning the total government investment in the Everyone's AI project amounts to around 40.5 billion won annually.
The Cybersecurity Specialized AI Foundation Model Development project will provide one selected team with 256 NVIDIA B200 GPUs (32 nodes) for ten months, with a total budget of up to 30 billion won, covering GPU procurement, labor costs, data construction, and service validation expenses.
With 256 B200 GPUs (32B class) deployed over ten months, the theoretical token processing capacity is estimated at around 62 trillion tokens. Government officials explain that this level of support allows for initial training and service validation.
However, challenges arise with the Everyone's AI project. The theoretical computational performance of 256 B200 GPUs is approximately 1,152 petaflops (PFLOPS) based on FP8 standards, which is about 2.3 times lower than that of 256 H100 GPUs.
Industry analysis indicates that when converted to the actual usage benchmark of domestic AI startup Allganize, the maximum simultaneous processing capacity for 256 B200 GPUs is only about 14,000 processes. While optimization may increase this limit, providing a free AI chatbot service to the entire population is deemed practically difficult.
Particularly, if specific users excessively utilize the government-provided AI chatbot, bottlenecks are likely to occur.
Cho Sung-bae, a professor in the Department of Computer Science at Yonsei University, stated, "While it is possible to make the model lightweight to accommodate as many users as possible for Everyone's AI, there are realistic limitations. It is certainly not at a level suitable for development or research purposes. Nevertheless, the attempt to pursue this with limited resources is commendable."
Concerns have also been raised about the insufficient national investment in AI compared to international examples. The Indian government has allocated 1.7 trillion won to its IndiaAI mission, securing over 38,000 GPUs and operating them at lower rental prices for startups and research institutions.
Saudi Arabia's Public Investment Fund (PIF) has announced plans to deploy 600,000 GPUs through its state-owned AI company, Humein, over the coming years. These initiatives aim to go beyond providing free chatbots to the public, focusing on computing rentals for startups and research institutions, as well as data center exports.
The government plans to allocate a significant portion of this year's national AI budget of 9.9 trillion won to expanding data centers and securing GPUs. However, as new projects increase, the number of GPUs allocated to individual projects is relatively decreasing. The Everyone's AI project aims to select operators this month and launch a beta service by the end of September, with a full official service expected later this year.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.