South Korea Controls AI Memory: HBM4 Race and What It Means

SK Hynix completed the world's first HBM4 development in January 2026 and is targeting 70% of supply for Nvidia's next Rubin GPU platform. Samsung is close behind, nearing Nvidia qualification approval. South Korea controls the memory that powers all frontier AI training.

What Is HBM and Why Does It Matter for AI?

HBM (High Bandwidth Memory) is a type of stacked DRAM that sits directly on the same package as an AI accelerator, connected by thousands of tiny wires called through-silicon vias. It delivers memory bandwidth an order of magnitude higher than regular DRAM, which is the bottleneck that determines how fast a GPU can feed data to its processing cores during AI training.

Every Nvidia GPU used for AI training — H100, H200, B200, and the upcoming Rubin — uses HBM. There is no substitute at the required performance level. SK Hynix and Samsung between them produce essentially all HBM in the world. The two Korean companies are, functionally, a duopoly over the memory supply chain for global AI infrastructure.

The HBM4 Milestone

SK Hynix demonstrated 16-layer 48GB HBM4 running at 11.7 gigabytes per second per pin at CES 2026 in January 2026. UBS projects SK Hynix will capture approximately 70% of HBM4 supply allocated to Nvidia's Rubin GPU platform, which is scheduled for commercial release in Q3 2026.

Samsung reached a different milestone: it began shipping HBM3E chips to Nvidia in Q3 2025 and has sold out its entire 2026 HBM supply as a result. Samsung is now in the qualification and approval process for HBM4 at Nvidia, which it is expected to complete in early 2026.

Both companies raised HBM3E prices by nearly 20% heading into 2026, reflecting the supply constraint.

HBM Generation Comparison

South Korea's 250,000 GPU Deployment

At a Seoul summit in October 2025, Nvidia announced it would deploy over 250,000 GPUs across South Korea. The deployments span four major organisations:

• Samsung: 50,000 Nvidia GPUs to automate chip manufacturing (computational lithography), claiming 20x performance improvement
• SK Group: 50,000+ GPUs including Asia's first industrial AI cloud, using RTX PRO 6000 Blackwell cards
• Hyundai Motor: 50,000 Blackwell GPUs for robotics AI and factory automation
• NAVER Cloud: Scaling to 60,000+ GPUs for enterprise and physical AI applications

South Korea now ranks among the most GPU-dense countries on Earth relative to its population. More importantly, every GPU deployed there runs Nvidia silicon — which requires HBM supplied by SK Hynix or Samsung. South Korea sits on both sides of the AI hardware supply chain.

The KSMC Question

South Korea's government is considering creating KSMC (Korea Semiconductor Manufacturing Company), a state-backed contract chip manufacturer. The proposed investment is KRW 20 trillion ($13.9 billion). The target process node is 40nm, aimed at automotive and industrial semiconductor markets that TSMC currently dominates.

This is a direct response to Samsung Foundry struggling to compete with TSMC at advanced nodes. TSMC holds roughly 70% of global foundry market share; Samsung holds around 7%. Samsung has had wins — it is manufacturing the Nintendo Switch 2 chip at 8nm, and undercut TSMC's 2nm wafer pricing by 33% ($20,000 vs $30,000+). But at advanced yield parity, Samsung is not there yet.

A state-backed KSMC would give South Korea a foundry that does not depend on Samsung's commercial decisions, providing a third option for the US-led supply chain diversification away from Taiwan.

What Developers Should Know

HBM supply is the immediate constraint on AI GPU production. If you are trying to get H100 or B200 allocations from cloud providers and finding them scarce, this is why. Nvidia can manufacture GPU dies faster than SK Hynix and Samsung can produce HBM to stack on them.

The HBM4 transition to Rubin in Q3 2026 will reset the supply constraint at a higher tier. Rubin GPUs require HBM4, which is not interchangeable with HBM3E. That means HBM4 supply in 2026 determines how many Rubin GPUs enter the market in 2026 and 2027.

For anyone building AI infrastructure cost models: watch Korean memory company earnings calls. SK Hynix and Samsung HBM revenue is the leading indicator for GPU availability and pricing 6 to 12 months forward.

Key Takeaways

• 70% — SK Hynix projected market share of HBM4 supply for Nvidia Rubin GPUs
• 11.7 GB/s per pin — SK Hynix HBM4 performance at January 2026 CES demonstration
• 250,000+ — Nvidia GPUs being deployed across South Korea in 4 major deployments
• 20% — price increase on HBM3E memory heading into 2026
• For developers: GPU scarcity is a memory supply problem, not a compute die problem. HBM4 production rate determines how quickly Rubin GPUs reach cloud providers in late 2026.
• What to watch: Nvidia's Rubin GPU commercial launch in Q3 2026 — the first real-world test of whether HBM4 supply is sufficient to meet demand