Is SK hynix the Power Behind NVIDIA’s Throne?

For most of the computer era, memory chips were the semiconductor industry’s supporting cast. Necessary, certainly. Occasionally scarce. Often expensive at precisely the wrong moment. But they were not what made people line up for product launches or drove investors into a frenzy. Processors were where the action was. Memory was the mundane stuff sitting beside them, quietly holding data until the important chip needed it.

That is not how Wall Street treated SK hynix last week.

The South Korean memory manufacturer raised $26.5 billion through a U.S. offering of American depositary shares, reportedly attracting more than seven times as many orders as there were shares available. Its Nasdaq debut became the largest U.S. share sale by a foreign company, passing Alibaba’s 2014 offering. SK hynix already trades in Seoul, so this was not an initial public offering in the conventional sense, but investors did not seem troubled by the distinction. They were buying access to one of the most strategically important companies in the AI supply chain.

The shares opened at $170, well above the offering price of $149, and finished their first day up roughly 13%. That reception was Wall Street’s acknowledgment that memory is no longer just something attached to the star of the show. In AI, memory has become part of the performance, part of the scarcity and increasingly part of the power.

NVIDIA may sit on the AI throne, but SK hynix helps make the throne work.

When DRAM Became a Star

Dynamic random-access memory, better known as DRAM, has been around for decades. It stores the working data processors need while operating. Every laptop, server, smartphone and increasingly every car contains some form of it. For much of its history, DRAM has been a commodity business governed by a familiar and often painful cycle. Demand rises, prices climb and manufacturers add capacity. Too much capacity comes online, prices collapse and the same manufacturers slash spending. Eventually, supply tightens and the cycle begins again

Investors learned not to become too attached to memory-company profits. The good years often carried the seeds of the bad ones

High-bandwidth memory, or HBM, has complicated that history. HBM is built from DRAM, but instead of arranging memory chips in the more conventional ways found in PCs and servers, manufacturers thin the individual dies, stack them vertically and connect them through microscopic pathways. The completed stacks sit close to the GPU or other accelerator, allowing enormous quantities of data to move quickly while consuming less energy than would otherwise be required

That speed matters because an AI accelerator is only useful when it has data to process. NVIDIA can design a GPU capable of staggering computational performance, but if the processor spends its time waiting for data to arrive from memory, much of that expensive capacity sits idle. This is the so-called memory wall, and it has become one of the defining constraints in advanced computing.

HBM helps break through that wall. It also makes memory far more difficult to dismiss as a commodity.

Building HBM requires more than producing ordinary DRAM dies. The stacking, through-silicon connections, thermal management, testing and advanced packaging all add complexity and introduce opportunities for something to go wrong. Producing a usable gigabyte of HBM is commonly estimated to consume roughly three times the wafer capacity required for an equivalent amount of conventional DDR5 memory. Qualification is demanding, production yields matter enormously, and customers cannot casually switch suppliers when a product is already designed around a particular memory configuration.

This is how DRAM took on superstar status. It did not stop being memory. It became memory that could determine whether the world’s most valuable AI systems shipped on time and performed as promised.

NVIDIA’s Memory Dependency

NVIDIA is frequently discussed as if it manufactures the AI revolution itself. What NVIDIA actually does exceptionally well is design the architecture, software and systems that have become the dominant platform for AI computing. Manufacturing those systems depends on an intricate network of outside companies.

TSMC manufactures NVIDIA’s leading processors and plays a central role in advanced packaging. SK hynix, Samsung and Micron provide the HBM. Other suppliers provide substrates, networking components, power systems and cooling. A problem anywhere along that chain can prevent NVIDIA from shipping a completed system.

SK hynix has emerged as the most important memory supplier in that network. It held approximately 58% of the HBM market during the first quarter of 2026, according to figures cited in coverage of the offering. Samsung and Micron each held approximately 21%. Market-share estimates vary by quarter and methodology, but the hierarchy is unmistakable: SK hynix is the HBM leader.

The connection with NVIDIA now extends well beyond filling purchase orders. In June, the two companies announced a multiyear technology partnership covering next-generation memory for NVIDIA’s AI factories, Vera Rubin systems, Vera CPUs, personal AI computers and Jetson robotics platforms.

Jensen Huang described SK hynix as NVIDIA’s largest memory partner and said it would continue in that role. He also said NVIDIA already purchases billions of dollars of products from the company annually and expects that amount to grow substantially. Even SK hynix’s plan to double its memory-wafer capacity by 2030, Huang said, would not be sufficient to meet anticipated demand. Reuters reported on the expanded partnership.

This is closer to co-development than conventional procurement. HBM4 introduces opportunities for customized logic in the base die beneath the memory stack, allowing the memory to be more closely tailored to a particular accelerator and system architecture. The closer SK hynix gets to NVIDIA’s product roadmap, the harder it becomes to describe its product as an interchangeable component.

Daniel Newman, CEO of The Futurum Group, captured the competitive picture succinctly in comments to Reuters: “SK Hynix leads on share and Nvidia proximity, Micron competes on power efficiency, U.S. positioning, and momentum from third place.”

The NVIDIA proximity may be the most valuable part of that equation. SK hynix sees where the leading AI architecture is going, participates in the development process and prepares its manufacturing around NVIDIA’s future requirements. That position gives it greater visibility into demand while raising the barriers facing Samsung and Micron.

It also creates risk. Newman subsequently described SK hynix as the purer AI investment, noting that its greater concentration in HBM and NVIDIA “cuts both ways.” If AI infrastructure spending continues growing, SK hynix has perhaps the best seat in the house. If spending contracts sharply, its customer concentration becomes a liability.

For now, shortage rather than contraction is dominating the conversation.

AI is Tightening the Entire Memory Market

The rush toward HBM not only affects companies building AI clusters. HBM and conventional DRAM compete for portions of the same manufacturing resources. When memory manufacturers allocate more wafer capacity to highly profitable HBM, fewer wafers remain for the memory used in PCs, smartphones, automobiles, networking equipment and traditional servers.

That helps explain why an AI data center can influence the cost of a laptop or car built thousands of miles away.

TrendForce projected that conventional DRAM contract prices would rise between 90% and 95% during the first quarter of 2026 compared with the previous quarter. That was a forecast rather than a final market accounting, but the magnitude shows how severe the imbalance had become. Reuters reported the revised TrendForce projection.

Manufacturers cannot simply fix this by running their existing factories harder. Modern semiconductor fabs cost tens of billions of dollars and take years to design, permit, construct and qualify. HBM also requires specialized packaging and testing capacity that must expand alongside wafer production. Customers are reserving supply years in advance, while each new generation of AI accelerator appears to require more memory than the one before it.

The three companies that control nearly all advanced DRAM have another reason to exercise caution. Samsung, SK hynix and Micron remember what happened when the industry previously mistook a shortage for permanent demand and built too much capacity. Memory gluts destroy pricing power remarkably quickly.

SK hynix now says the worst may still be ahead. CEO Kwak Noh-jung told Reuters that 2027 could be the worst supply year in the industry’s history and that customer demand may exceed the company’s capacity beyond 2030. The warning came on the day of the Nasdaq debut.

Rolf Bulk, head of semiconductors and infrastructure at Futurum Equities, sees the market expanding accordingly. He told Reuters that Futurum expects HBM to grow from approximately $65 billion this year to $120 billion next year and around $290 billion by 2030.

Forecasts reaching four years into the semiconductor future should always be treated as forecasts, but Bulk’s numbers show the scale of the opportunity investors believe they are buying. A memory category that was once a specialized corner of the market is threatening to become one of its largest profit pools.

Micron and the American Counterweight

There is a U.S. industrial-policy story sitting just beneath the SK hynix celebration.

NVIDIA is an American company, as are AMD, Broadcom and the hyperscalers designing their own AI accelerators. Yet the United States does not presently manufacture enough of the leading-edge memory those companies require. According to the CHIPS for America program, the country accounts for less than 2% of global advanced-memory production.

Micron is the only U.S.-headquartered company competing at the leading edge of DRAM and HBM. That makes it more than SK hynix’s smaller rival. It is Washington’s most credible opportunity to create a domestically controlled source of the memory required for AI infrastructure.

Here again, headquarters and manufacturing location should not be confused. Much of Micron’s current advanced manufacturing remains in Taiwan and Japan. American ownership does not automatically make the finished supply chain American.

Micron is trying to change that. The company increased its planned U.S. investment to more than $250 billion through 2035, including leading-edge fabs in Idaho, as many as four fabs in New York, modernization in Virginia, advanced HBM packaging and approximately $50 billion in research and development. Its stated goal is eventually to produce 40% of its DRAM in the United States. The projects have received commitments for as much as $6.4 billion in direct CHIPS Act funding. Micron describes the expansion here, while CHIPS for America details the federal award and project scope.

The timing was striking. Micron announced the latest increase in its American investment on July 9, immediately before SK hynix began trading on Nasdaq. It also committed as much as $3 billion to strengthen the U.S. semiconductor ecosystem, including $500 million supporting GlobalWafers’ expansion of 300mm silicon-wafer manufacturing in Texas.

Wall Street is giving SK hynix $26.5 billion to extend its advantage while Washington and Micron assemble an American alternative.

That alternative will not solve the current shortage. Micron’s first new Idaho production is expected sooner, but initial production from the enormous New York campus is not anticipated until around 2030. Semiconductor industrial policy operates across decades. NVIDIA’s roadmap advances in annual or biennial generations.

The gap between those two clocks is the vulnerability.

South Korea is a close American ally, and none of this makes SK hynix an unreliable partner. The global semiconductor ecosystem has created extraordinary specialization and innovation. But concentration creates risk even among allies. An interruption affecting South Korea or Taiwan could constrain American AI development, cloud expansion, defense systems and large portions of the broader economy.

The United States has produced the world’s leading AI accelerator company without retaining enough domestic capacity to manufacture all the critical ingredients inside its products. Micron’s expansion, combined with TSMC’s investment in Arizona and new domestic packaging capacity, begins addressing that problem. It will take years before those pieces resemble an end-to-end American supply chain.

Who Really Holds the Power?

There is still a cyclical-memory bear case hiding beneath all this enthusiasm. SK hynix, Micron and Samsung are committing extraordinary amounts of capital to new production. If hyperscaler spending slows just as those fabs begin delivering capacity, today’s shortage could become tomorrow’s glut. HBM is more differentiated than conventional DRAM, but it has not repealed the economics of supply and demand.

The larger change is harder to dismiss. Memory is moving closer to the architecture. Suppliers are collaborating with accelerator designers earlier in the development process. Customers are making longer commitments. The product is becoming more customized, qualification is becoming harder and the cost of switching suppliers is rising.

For years, the semiconductor business treated memory as the mundane chip beside the processor that did the real work. AI has exposed the weakness in that assumption. An accelerator without enough fast memory is not an AI supercomputer. It is expensive silicon waiting for data.

NVIDIA owns the architecture, the CUDA ecosystem and much of the market power defining the AI era. It deserves the throne it has built. But power is not the same thing as self-sufficiency.

Behind that throne sits a South Korean company that spent more than a decade betting on a memory technology many people considered too specialized to matter. Wall Street just made a $26.5 billion wager that the supporting actor has become indispensable.