SK hynix’s $8B ASML bet reshapes memory for AI and PCs

SK hynix just placed the single largest order for ASML’s extreme ultraviolet (EUV) lithography machines—$8 billion for up to 30 machines over two years, according to its regulatory filing. The move isn’t just about scale; it’s a direct play to dominate the memory market’s most lucrative segment: high-bandwidth memory (HBM), the fast, stacked DRAM that powers AI accelerators like Nvidia’s H100 and AMD’s Instinct MI300. Every major AI chipmaker is scrambling for HBM supply, and SK hynix’s bet suggests it plans to outpace Samsung and Micron in both volume and performance.

The machines themselves, ASML’s Twinscan EXE series, are the only tools capable of printing the tiny, precise features required for 3nm and smaller nodes. That’s critical for HBM, where density and speed directly translate to AI training efficiency. But here’s the catch: ASML’s production capacity is finite. Even with this order, SK hynix won’t receive all 30 machines at once—deliveries will stretch over two years, and ASML’s backlog already stretches into 2026. For customers like Nvidia or Microsoft, that means HBM supply constraints aren’t disappearing anytime soon.

This isn’t just a memory story. The order reflects a broader shift in computing: AI workloads are now the primary driver of semiconductor innovation, and memory is the bottleneck. Traditional DRAM for PCs and servers is still important, but HBM is where the margins are. SK hynix’s investment signals confidence that AI demand will remain insatiable, even as the broader PC market stagnates.

For users, the implications are mixed. On one hand, more HBM supply could ease the chronic shortages that have inflated AI hardware costs. Nvidia’s H100 GPUs, for example, are still selling for over $30,000 on the secondary market, partly due to HBM scarcity. If SK hynix ramps production faster than competitors, we might see a gradual drop in prices for AI training hardware—a win for startups and researchers priced out of the current market.

On the other hand, this is a high-stakes gamble. The $8 billion price tag is nearly a third of SK hynix’s annual revenue, and EUV machines don’t come with guarantees. Each machine costs over $200 million, requires a cleanroom the size of a football field, and demands a team of specialized engineers to operate. If AI demand cools—or if a breakthrough in alternative memory technologies emerges—SK hynix could find itself with billions in underutilized equipment.

The real signal here isn’t just about SK hynix’s ambition; it’s about ASML’s stranglehold on the future of computing. With only one company capable of producing EUV machines, the entire semiconductor industry is now dependent on a single supplier. That’s a vulnerability regulators and chipmakers alike are starting to acknowledge, but for now, ASML holds all the cards. For SK hynix, the bet is simple: control the memory, control the AI market.

What happens if the AI bubble bursts? SK hynix’s $8 billion bet assumes demand for HBM will keep growing at its current pace, but what if the next generation of AI models doesn’t require exponentially more memory? The company is banking on a future where memory is the limiting factor—but what if the real bottleneck shifts back to compute power, or even software efficiency?