SK Hynix Just Hit a 72% Operating Margin, and the HBM Bottleneck Now Shapes Where Startups Can Build

What Happened

SK Hynix’s Q1 results reset the company’s records. Revenue reached 52.58 trillion won (about $35.5 billion), the first time quarterly revenue cleared 50 trillion won, up roughly 198% year over year. Operating profit was 37.61 trillion won (about $25.4 billion) and net profit 40.35 trillion won. The number that stands out is a 72% operating margin. The company itself noted that revenue topped 50 trillion won for the first time on a quarterly basis while operating profit and margin both hit record highs. That 72% sits above Nvidia’s roughly 65%, TSMC’s 58%, and Samsung Electronics’ 43% in the same window. A component supplier out-marginning the finished-product makers is an unusual picture.

The engine is HBM, the high-bandwidth memory stacked onto Nvidia’s AI accelerators. Only three companies mass-produce it at scale: SK Hynix, Micron, and Samsung. SK Hynix held about 59% of the market in 2025, and Q2 2026 tallies widened the gap to 62% for SK Hynix, 21% for Micron, and 17% for Samsung, with SK Hynix reported to have secured roughly two-thirds of Nvidia’s next-generation HBM4 orders. The company said customer demand already exceeds its supply capacity. Its seventh-generation HBM4E is slated for samples in the second half of the year and mass production in 2027.

What This Means for Founders

This is more than one company’s blowout quarter. The component the entire AI boom rests on sits with three firms, more than half of it with one, and the demand for it has outrun what those firms can make. The narrowest chokepoint in AI runs through a single supply chain. For founders that splits two ways.

One direction is opportunity. HBM sold out years deep, and SK Hynix pouring money into fab expansion and back-end packaging, means demand rises underneath it too: materials, equipment, inspection, thermal management, advanced packaging. The narrow processes the memory giants do not run in-house, HBM stack inspection, cooling for high-heat parts, localized materials, fab automation, are exactly the seams a deep-tech startup can work. A giant’s expansion becomes a supplier’s order book.

The other direction is constraint. For founders building AI products, memory is no longer a priced input but a scarce resource allocated to the biggest customers first. While HBM goes to Nvidia and the hyperscalers ahead of you, a small team in line pays more for the same performance or cannot get it at all. That is also why betting a business on inference getting steadily cheaper is shaky. Running smaller models, quantizing, pushing on-device, and cutting calls with caching stops being an optimization and becomes a survival condition.

What You Can Do Now

If you are building in the semiconductor ecosystem, track the expansion roadmaps and supplier tenders from SK Hynix and Samsung the way you would track commodity prices; attaching to a process where orders land first beats trying to manufacture demand. If you are on the AI-product side, measure your actual per-task memory and inference cost now, and simulate whether the business survives that cost doubling. The team that builds margin at expensive-memory prices, instead of waiting for supply to loosen, takes the next cycle.