The AI trade is moving deeper into the supply chain, and high-bandwidth memory is becoming one of its most important pressure points.
Investors are starting to treat SK hynix less like a cyclical memory stock and more like a toll collector on the AI buildout. That is the real story behind the latest rush into high-bandwidth memory, or HBM, the stacked DRAM that sits beside advanced AI processors and feeds them data fast enough to keep expensive GPUs from sitting idle.
For the last two years, the cleanest public market AI trade was Nvidia. That was understandable. Nvidia controlled the accelerator layer, set the software standard and became the default supplier for hyperscalers building large model infrastructure. But every accelerator needs a memory system around it. As training clusters and inference farms grow larger, the bottleneck is no longer only how much compute a buyer can secure. It is also how quickly those chips can access data.
That is why SK hynix matters. The South Korean company has built an early lead in HBM and remains closely tied to Nvidia’s GPU roadmap. Global X research published this spring put SK hynix at 57% of global HBM revenue and 62% of shipment volume, with more than two-thirds of Nvidia’s HBM orders for its 2026 Rubin platform already secured by the company. Those numbers explain why memory has suddenly become a serious institutional theme rather than a niche semiconductor detail.
HBM is expensive because it solves a very specific problem. Large AI systems need enormous amounts of information to move between processors and memory at high speed. Conventional memory cannot do that efficiently enough at the scale Nvidia, AMD, Google and other AI hardware buyers now require. HBM stacks memory vertically and places it close to the processor, improving bandwidth and power efficiency in a way ordinary DRAM cannot match.
The economics are changing with the technology. HBM is harder to manufacture, capacity is limited and customers are increasingly willing to lock in supply early. Global X estimates that global HBM spending could rise 58% in 2026 to $54.6 billion, then approach $100 billion by 2028. That is not a small side market. It is becoming a core piece of AI infrastructure spending.
There is also a practical reason investors are watching this closely. A single Nvidia GB200 NVL72 system can require up to 13.4 terabytes of HBM, alongside dozens of Blackwell GPUs, CPUs, networking equipment and cooling systems. When AI clusters move from thousands of GPUs to tens of thousands, memory demand scales with them. The chip may get the headline, but the memory determines how much useful work the system can actually do.
SK hynix has leaned into that shift. Nvidia announced a new SK Group AI factory project featuring more than 50,000 Nvidia GPUs, with the first phase planned for completion by late 2027. The project is not only about buying GPUs. It also deepens work on SK hynix HBM and future memory solutions for Nvidia platforms, chip manufacturing and telecommunications. That makes the partnership look less transactional and more structural.
Samsung is pushing back, but SK hynix still has the lead
The race is not standing still. According to Reuters, Samsung Electronics said on May 29 that it had started shipping samples of its 12-layer HBM4E chip to global customers, calling it the industry’s first shipment of that product class. Samsung said the chip offers more than 20% higher speed performance than its previous HBM4 products and uses its sixth-generation 10-nanometer-class DRAM process with a 4-nanometer logic base die.
That matters because Samsung cannot be counted out. Its customers include major AI players such as Nvidia, AMD and Google, and the company is working hard to regain ground after SK hynix built the stronger position in recent HBM cycles. Samsung’s HBM4E sample shipment gives buyers another path and gives investors a reason to watch qualification milestones closely.
Still, samples are not the same as dominance. AI hardware buyers care about yield, reliability, thermal behavior, supply volume and roadmap alignment. SK hynix has benefited from being early where it counted, especially with Nvidia. It has also continued to push the packaging side of the problem, including recent work on iHBM thermal architecture that targets heat buildup inside dense memory packages. That is a reminder that the next phase of AI hardware competition will be fought across many small engineering details, not only headline bandwidth numbers.
For startups, this has a blunt implication. AI hardware companies building accelerators, inference boxes or specialized servers cannot plan around logic chips alone. Memory access, supply commitments and power behavior may determine whether a product can ship at cost and at scale. A clever processor design is not enough if the surrounding memory stack is unavailable, too expensive or already reserved by larger customers.
For investors, the lesson is just as direct. The AI trade is maturing from a single-stock story into a supply chain story. Nvidia is still central, but the market is now asking who supplies the parts that make Nvidia-class systems usable. SK hynix, Samsung and Micron sit close to that question. The next signal to watch is not only who announces the fastest HBM sample, but who can deliver qualified volume when hyperscalers start converting ambitious AI budgets into real machines.
Also read: Jiducloud becomes a unicorn as China’s AI money moves into 3D tools • MiniMax M3 gives Chinese AI labs a new frontier coding test • AI detectors are turning ordinary student writing into evidence.
