South Korea Will Offer Free AI on 512 B200 GPUs

What South Korea is actually launching

What stands out here is not the slogan about accessible AI, but the scale of the hardware: South Korea plans to distribute up to 512 NVIDIA B200 GPUs among consortia led by SK Telecom, KT and Kakao. Citizens are promised free AI services without usage limits.

The South Korean government describes the project as an AI for All/All for AI initiative. As of early September 2026, beta testing was expected within a month, with a broader launch planned for late 2026. Some reports also point to government support for operating costs through at least 2030.

The B200 is based on NVIDIA's Blackwell architecture and designed for tightly coupled computing systems using NVLink. Available specifications cite roughly 8 TB/s of memory bandwidth and around 1,000 watts per GPU. That means 512 accelerators are not simply a large procurement: they require serious cooling, scalable networking and a scheduler that does not turn a shared resource into a queue of stalled jobs.

There is a notable inconsistency around funding. I would not accept a claim of roughly $7 billion without separate confirmation: the available reports cite subsidies of around 6.76 billion won for groups with 256 GPUs and 3.38 billion won for groups with 128 GPUs. That is an entirely different order of magnitude.

There is also no public benchmark for this specific distributed system. The performance of individual B200 GPUs and tightly integrated servers is known, but that proves nothing about a national multi-tenant service operated by three providers.

The real test starts after launch

This is a genuine infrastructure bet, not merely an attractive government announcement. Free access to capable models could lower the barrier for developers, researchers and ordinary users who cannot afford their own accelerators.

But the word “unlimited” immediately raises architectural questions. How will interactive inference, batch workloads and training be separated? What tenant isolation will be used, how will one especially demanding workload be constrained, and who gets priority during peak demand?

Three providers add another layer of complexity: unified access rules must work across different cloud platforms and operational practices. Users benefit if interfaces and quotas are transparent. Otherwise, free AI may be technically available yet practically unpredictable.

The initiative has strong hardware and a clear public purpose. What will decide its outcome is the least glamorous part of the project: operating a shared resource under real-world load.

We previously examined Durov's Cocoon: its approach to confidential computing, inference costs and scaling AI access. That context helps assess the infrastructure South Korea will need to provide free AI to all citizens.