Samsung’s chip division just posted a record quarter, with memory revenue hitting $25.9 billion in Q4 2025 alone. Profits tripled year on year. Meanwhile, SK Hynix overtook Samsung in annual profits for the first time in the company’s history, driven almost entirely by AI memory demand. Memory now accounts for around 40% of Samsung’s total revenue. A company most people associate with phones and televisions is quietly becoming one of the most important players in the AI era.
The AI supply chain story is usually told through three countries: American chip design, Dutch lithography, Taiwanese fabrication. But there is a fourth pillar that gets far less attention. South Korea produces the overwhelming majority of the world’s High Bandwidth Memory (HBM), the specialist memory chips that sit at the heart of every AI server. Samsung and SK Hynix between them dominate HBM production. Both of them have signed supply deals with OpenAI’s Stargate project, for example, that could consume up to 40% of global DRAM output. South Korean semiconductor exports surged 102% in January. If Taiwan is the world’s chipmaker, South Korea is fast becoming the world’s memory bank.
But that memory bank is running low. DRAM prices rose 90-100% in Q1 2026, the largest quarterly increase on record. The cause is a structural supply shortage that analysts believe could persist until 2028. Unlike previous memory cycles, the physics of DRAM scaling have hit a wall. Over the past decade, DRAM density has only doubled in total, compared to roughly 100x per decade during its peak scaling era. New capacity cannot be engineered quickly; it requires years of construction and billions in investment.
The consequences are already spreading. Nvidia reportedly plans no new gaming GPU in 2026, the first time in nearly 30 years, as it prioritises memory allocation for data centre chips. RTX 5090 prices have hit $4,000. Memory could add $96 to the cost of a basic PC this year. AI is effectively levying a tax on consumer electronics.
For the hyperscalers, who have committed to $600+ billion on capex in 2026, the shortage creates a kind of accelerating paradox. Scarcity may not slow the great datacentre race. It will most likely intensify it. No company can afford to relinquish its memory orders and allocation for fear that a competitor will take them. The result is another arms race dynamic where constraint breeds even greater commitment.
Takeaways: Memory is the hidden bottleneck and hidden accelerant of the AI boom. South Korea’s role as the world’s memory producer places it alongside Taiwan as a critical, and under-discussed, node in the global AI supply chain. The current shortage, the biggest in four decades, will not slow AI deployment. It will make it more expensive, more competitive, and more urgent.