Samsung's HBM Chase Is an AI Supply-Chain Time Bomb

Samsung Won the Quarter. But the War Isn’t Over.

Samsung Electronics’ HBM market share jumped from 21% to 33% in the second quarter, according to Counterpoint Research. The gap with leader SK Hynix collapsed from 37 percentage points to 17 in a single quarter. At first glance, this looks like a comeback story—a Korean industrial heavyweight finally delivering after years of falling behind in high-bandwidth memory. But when you look past the headline number, what actually matters is the trajectory, and right now the trajectory points toward a fragile dependency that could choke the entire AI chip supply chain.

Samsung achieved this surge by being the first to ship HBM4 in February, a full four months ahead of SK Hynix. The company hit $1 billion in HBM revenue within four months of mass production, according to Counterpoint’s tracking. Meanwhile, SK Hynix held onto its 50% share by leaning on its established HBM3E product line, even as that share dropped 8 percentage points in the same period. Micron, stuck at 18%, slid another 3 points. Together, the two Korean companies now control 83% of the HBM market—a concentration level that should make every AI infrastructure buyer nervous, especially when you consider what happens next.

HBM4 Is Where the Real Stakes Live

Here is the detail most reports are glossing over: Counterpoint itself noted that the majority of current HBM revenue still comes from HBM3E. HBM4 shipments are only beginning to ramp. The revenue cliff for SK Hynix arrives when the market shifts—and that shift is coming faster than anyone outside the fab floor expects. When HBM4 becomes the dominant SKU, Samsung’s head start becomes a structural moat.

Samsung’s HBM4 is already flowing into NVIDIA’s next-generation Vera Rubin GPU, which is scheduled to ship later this year. Samsung started earning revenue from it in Q2. SK Hynix, which began HBM4 shipments later, has not yet matched that timeline. If NVIDIA’s demand for HBM4 exceeds what SK Hynix can deliver in the near term, the company will have no choice but to allocate more volume to Samsung—regardless of whether Samsung’s yield rates or pricing are competitive. NVIDIA has been transparent about this dynamic in recent investor calls, acknowledging that HBM supply remains a binding constraint on GPU production ramps.

This is not speculation. It is how NVIDIA’s supply chain works. The GPU maker does not integrate memory; it buys HBM stacks from whatever supplier can meet volume and quality demands on schedule. Right now, that means Samsung is gaining leverage it did not have six months ago. And leverage in this context is dangerous because it is concentrated in a single hand at a time when the market needs redundancy.

One Country, Two Companies, All the Risk

South Korea controls 83% of HBM. That is an extraordinary concentration for a component that sits at the center of the most expensive capital expenditure cycle in technology history. Every data center buying NVIDIA Blackwell and Rubin-class GPUs needs HBM. Every hyperscaler building AI infrastructure needs HBM. And nearly all of it is coming from fab locations in Gumi and Icheon—cities that together house the world’s most critical memory-production assets.

Geopolitical risk analysts have been tracking US-China chip restrictions for years. But the blind spot has been supply chain geography. If anything disrupts Korean semiconductor operations—a natural disaster like the 2022 drought that strained cooling water at Samsung’s factory, a labor dispute, a trade policy shift, or a military contingency on the Korean Peninsula—the entire global AI hardware pipeline faces immediate strain. There is no quick alternative. TSMC cannot simply start producing HBM; the process technology requires a different skill set and different equipment. Intel’s memory division is years away from meaningful HBM output. CXMT in China sits at roughly 10% of the broader DRAM market and has no HBM4 capability whatsoever.

Samsung’s rise in HBM4 actually increases this risk in the short term. The more NVIDIA relies on Samsung for early HBM4 volume, the more asymmetric the supply chain becomes. SK Hynix remains the safer long-term bet for HBM4—the company has deeper process expertise and higher yields on advanced nodes—but until its volumes materialize at scale, the ball is in Samsung’s court. That asymmetry is the definition of a single point of failure.

The DRAM Picture Above HBM

Samsung’s overall DRAM market share tells a slightly different story. It sits at 38%, well ahead of SK Hynix’s 25% and Micron’s 24%. Samsung pulled this lead by focusing on general-purpose DRAM, where prices rose sharply following industry-wide capacity discipline, rather than competing solely on HBM. That strategy worked for earnings in Q2. Samsung reported record quarterly profits driven by DRAM recovery. But general-purpose DRAM is a commodity business with thin margins and no strategic moat. HBM is where the power is. And in HBM, Samsung is still chasing.

The gap in general-purpose DRAM gives Samsung financial breathing room, but it also creates a tension inside the company. HBM margins are significantly wider than commodity DRAM—industry estimates place HBM gross margins above 50%, compared to 20–30% for standard DRAM. That means Samsung has a financial incentive to keep shifting capacity toward HBM, even as it races to close the technology gap. The danger is that over-committing to HBM before the process is fully stable could lead to yield disappointments at exactly the wrong moment.

Second-Order Effects: What Happens Downstream

The concentration risk does not stop at the fab gate. The consequences ripple outward. AI system integrators who rely on predictable HBM delivery face scheduling uncertainty. Cloud providers planning data center builds must account for potential memory shortages that could delay GPU installation. Even end customers—enterprises training large language models—are indirectly exposed, since HBM constraints can delay the availability of the chips they depend on for inference and training workloads.

There is also a pricing dimension. When two suppliers control 83% of a market and neither faces meaningful competition, there is limited downward pressure on prices. HBM3E prices have already climbed significantly since 2023. As HBM4 matures, the question is whether Samsung’s entry will moderate pricing through competition or whether the duopoly will simply find a new equilibrium at higher price points. Early signs suggest the latter: Samsung’s HBM4 pricing has been reported to be competitive with SK Hynix, not aggressively lower, which means the duopoly pricing power remains intact.

What Happens Next

The second half of this year will determine who actually wins the HBM4 era. Counterpoint expects HBM4 shipments to become visible then. If Samsung maintains its first-mover advantage and delivers reliable HBM4 volume to NVIDIA, its share could push past 40% and possibly threaten SK Hynix’s position as the primary AI memory supplier. If SK Hynix accelerates HBM4 production and stabilizes its customer base, the gap narrows again—or reverses.

For NVIDIA, the immediate concern is simple: secure enough HBM4 to meet Vera Rubin demand without over-concentrating risk on a single supplier. The company already diversifies across Samsung and SK Hynix. But diversification only works if both suppliers can scale independently. Right now, they cannot—not at the pace the market requires. NVIDIA has reportedly begun discussions with additional memory partners, including exploring whether Intel can serve as a tertiary supplier in the longer term, though those conversations are still early.

For investors, the takeaway is equally stark. Samsung’s Q2 numbers look impressive on the surface, but the underlying dynamic is a company racing to catch a leader who still controls the bigger market. The HBM4 transition is the inflection point. Until it resolves, the Korean memory duopoly remains the single most concentrated and most exposed link in the AI hardware supply chain—and that exposure benefits no one outside the two companies themselves.

The AI infrastructure build-out is supposed to be about scaling capacity and driving costs down through competition. Right now, it is about surviving concentration. The question is whether the market can tolerate that fragility long enough for a third supplier to emerge—or whether the duopoly will lock in for years, pricing power firmly in Seoul.