High Bandwidth Memory has become the hard limit in AI infrastructure, and that has pushed SK Hynix, Samsung, Micron and TSMC into the part of the trade Nvidia used to own alone.
SK Hynix crossed $1 trillion in market value on May 27, after its shares closed 9.3% higher and took the South Korean memory maker into the same club as Nvidia, TSMC and Samsung. That is not a normal rerating for a company investors used to treat as a cyclical DRAM supplier. It tells you where the AI bottleneck has moved.
For most of 2023 and 2024, the story was simple enough: everyone wanted Nvidia GPUs and Nvidia couldn't ship enough of them. Now you have to look one layer deeper. The accelerator is still the headline product, but the server doesn't work without stacked memory, advanced packaging and a small circle of Asian suppliers that can actually deliver both at scale.
HBM, or High Bandwidth Memory, is not ordinary DRAM with a better label. It stacks memory dies vertically and connects them with thousands of tiny through-silicon vias, so data can move fast enough to keep an AI accelerator fed. If the memory can't keep up, the expensive GPU waits. Time is money, and idle AI hardware is very expensive time.
Only three companies make HBM at serious scale: SK Hynix, Samsung Electronics and Micron. SK Hynix supplied Nvidia early with HBM3 and HBM3E, which is why it has been treated as the purest public bet on AI memory. Samsung has since pushed hard into HBM4, with reports from TechRadar and other industry outlets saying its 11.7Gbps modules have been lined up for Nvidia's Vera Rubin platform. Micron said at GTC 2026 that it had started high-volume production of 36GB 12-high HBM4 for Vera Rubin, with more than 2.8 TB/s of bandwidth.
That is the part a casual AI investor misses. Nvidia designs the chips and captures huge margin, but the physical stack around those chips is becoming more valuable, not less. Tom's Hardware recently cited Bloomberg data showing Asian suppliers now account for roughly 90% of Nvidia's production costs, up from about 65% a year earlier. That includes TSMC fabrication, SK Hynix and Samsung memory, and server assembly from companies such as Foxconn and Quanta.
You can't wish that supply chain into existence in Arizona or Ohio by next quarter. TSMC's advanced nodes and CoWoS packaging capacity are already heavily booked. Jensen Huang warned during a Taiwan visit earlier this year, according to the South China Morning Post as cited by Tom's Hardware, that TSMC may need to more than double capacity over the next decade to meet Nvidia's demand alone. That is not a supplier politely asking for more orders. It is the customer admitting the whole machine depends on one island's manufacturing base.
The stock market has noticed, and then some. Business Insider reported that SK Hynix was up more than 230% for 2026 when it crossed the trillion-dollar mark, while South Korea's Kospi had more than tripled since the start of 2025. Samsung passed $1 trillion earlier in May. Micron also crossed that level in late May. If you owned the memory makers before the crowd understood the bottleneck, you didn't need a lecture on AI infrastructure. You got paid.
Frankly, this is where the easy story gets dangerous. Memory is still memory. It has a history of brutal cycles, and investors who forget that usually learn it with their own money. The same concentration that gives SK Hynix pricing power on the way up can hurt badly when customers pause orders, work through inventory, or shift to a newer generation of HBM. Business Insider reported that the Kospi plunged about 8% on Monday, June 8, before rebounding later in the week. That is what a crowded semiconductor trade looks like when rates or risk appetite move against it.
There is a real bull case underneath the volatility. MarketWatch, citing LSEG consensus estimates, recently reported that analysts expect SK Hynix sales to grow at a 25.6% compound annual rate from 2026 through 2028, with Samsung projected at 14.6%. SK Hynix has also told Reuters that the HBM market could expand about 30% a year through 2030 as Amazon, Microsoft, Google and other hyperscalers keep raising AI infrastructure spending. Those are concrete numbers, not mood music.
But don't confuse a real supply constraint with a one-way stock chart. Capital always follows margins. Samsung is expanding. Micron is no longer a distant third in the AI memory conversation. TSMC is adding capacity, even if the lead times are long. Scarcity is powerful right up to the point where the industry spends years trying to erase it.
The better conclusion is narrower and more useful: AI infrastructure was never only a GPU story. It is a systems story. If you want to understand who has leverage in the next phase, look at the companies that make the memory, packaging and wafers no model can run without. Right now, much of that leverage sits in Icheon, Hsinchu and Samsung's Korean fabs, not just in Santa Clara.
Also read: France drops Palantir for a local rival and hands every civil servant a Mistral AI assistant, OpenAI is asking public investors to fund a company that loses three dollars for every dollar it earns, STMicroelectronics raises $1.5 billion in convertible bonds as its AI photonics bet reshapes how the market sees the company
