Program Overview
Korea's Ministry of Science and ICT has allocated KRW 2.08 trillion (about $1.5B) for an initiative branded "AI Highway."
The program acquires 15,000 NVIDIA H200 and B200 GPUs and opens them to universities, research institutes, and businesses in stages. The cost split is sharp: universities and government-funded labs get free access, while small and mid-sized businesses pay only 5–10% of market rates.
Concrete numbers help. On commercial GPUaaS, a B200 rents for roughly $6/hour. SMBs accepted into the program pay around $0.45–$0.60/hour for the same hardware. Operating a 256-card cluster for four months yields savings in the hundreds of millions of KRW compared to market rates.
Two application channels exist: the NIPA (Korea National IT Industry Promotion Agency) "AI Computing Resource Strengthening Program" and the national AI Infra Hub portal (aiinfrahub.kr). Both gate selection on a proposal review.
Who Can Apply
Eligibility breaks into three groups.
Universities and national labs get free access by default — usually with disclosure conditions: published papers, open-source models, or post-program reports.
SMBs need a substantive proposal. The plan must specify the AI development goal, expected outputs, and resource allocation. "We want to try it out" does not pass review. NIPA also weighs the last 3 years of revenue and operating track record.
Startups can apply. B2B SaaS teams, AI agent builders, and groups working on RAG pipelines often have an edge. Joining a consortium with a GPUaaS operator or university increases approval odds.
Cluster size starts at 2 servers and scales up to 256 (2,048 GPUs). For a 70B-or-under LLM fine-tune, 8–16 cards are enough to begin.
H200 vs B200 — Which to Choose
A direct comparison simplifies the decision.
The H200 is the final Hopper-generation card: 141GB HBM3e, 4.8 TB/s bandwidth. It is well-suited to 7B–70B model inference and fine-tuning with a mature software stack.
The B200 is one generation ahead, built on Blackwell: 192GB HBM3e, 8 TB/s bandwidth. On GPT-3 175B pre-training, B200 finishes jobs roughly 2x faster than H200; on 70B LoRA fine-tuning, the gap is 2.2x. It also supports FP4 precision, lowering large-scale inference cost.
The practical rule is simple. For models under 141GB (about 70B parameters), H200 is the right choice. For larger models or multi-modal workloads, choose B200. Although B200's hourly rate is higher, faster throughput often lowers total cost.
Most SMBs that ARC Group advises start with an 8-card H200 cluster. Internal AI assistants like Slack/Discord bots, document RAG, and recommendation engines for B2B SaaS run well on H200.
Pitfalls — Confirm Before Applying
"Free" comes with conditions.
First, usage is time-boxed. One application typically grants 4 months. That is tight for end-to-end model training, tuning, evaluation, and report preparation. Have your training pipeline and dataset ready before approval.
Second, GPU ownership stays with the government. After the program, the hardware is not yours. If continued compute is needed, plan a parallel commercial cloud contract or self-hosted setup.
Third, disclosure conditions apply. Academic tracks almost always require open output (papers, weights). Even on the business track, some artifacts must be submitted to NIPA. If model weights or training data must remain proprietary, confirm conditions before applying.
Fourth, define responsibilities clearly when forming a consortium. Multi-party applications fare better in review, but cluster operations, IP rights, and reporting ownership should be settled upfront.
ARC Group's Recommended Strategy
Treating government GPU access purely as cost savings limits its value.
ARC Group recommends a three-phase approach to mid-market clients.
Phase 1 is PoC. Use an 8-card H200 cluster for 4 months to complete an LLM fine-tune, a RAG pipeline, or an internal AI assistant prototype. Cut compute spend by roughly 80% versus commercial cloud while measuring real model quality.
Phase 2 is operational transition. If PoC results justify it, move the same model to commercial cloud (AWS SageMaker, GCP Vertex AI) or a self-hosted setup. The model weights you trained during the program are yours.
Phase 3 is downstream funding. Your program report becomes a track record for future applications — to IITP, KISA, or other government R&D initiatives.
The point is not "free GPUs" — it is "a verified starting point for workloads worth scaling." ARC Group supports clients across proposal writing, model selection, 4-month execution planning, and the eventual cloud transition.