Samsung and SK Hynix Are Still Climbing — The AI Memory Boom Has Farther to Run

The Number That Should Make Wall Street Sit Up

Nineteen trillion won. That is what Samsung Electronics and SK Hynix are expected to generate in combined operating profit for the third quarter of 2025 — a figure that shatters every existing benchmark and, more importantly, undermines the growing consensus that the memory supercycle has already peaked.

Samsung alone is projected to post 111.38 trillion won in quarterly operating profit, marking the first time any South Korean company has breached the 100-trillion-won threshold in a single quarter. It would also break its own Q2 record of 89.49 trillion won, extending a streak of four consecutive quarters of all-time highs that began in Q4 2023. SK Hynix is expected to contribute 78.13 trillion won in operating profit on revenue of 99.54 trillion won.

Together, the two companies are posting year-over-year operating profit growth rates of 815% and 586% respectively. The memory cycle is not cooling. It is accelerating.

What the Numbers Actually Mean

The conventional narrative among some analysts has been that memory prices have run too far, too fast, and that a correction is inevitable. Intel CEO Pat Gelsinger recently pushed back hard against that framing, predicting at the AI Infrastructure Summit in Santa Clara that the memory shortage will worsen next year as artificial intelligence deployment drives further demand surges.

The data supports him.

General-purpose DRAM prices are projected to rise another 13% to 18% in Q3 alone, according to TrendForce, with NAND flash prices climbing 10% to 15%. But those headline figures obscure the deeper structural distortion happening inside the fab floor.

Here is the critical mechanism: Samsung and SK Hynix are redirecting massive portions of their DRAM production capacity toward HBM — high-bandwidth memory — which is essential for AI training and inference chips. Samsung’s HBM share of total DRAM production capacity is expected to climb from 27% last year to 33% this year and 40% by 2027. SK Hynix is running a similar shift.

Every watt of capacity devoted to HBM is a watt not available for commodity DRAM. The result is a structural shortage in general-purpose memory that extends far beyond any typical cyclical peak. DDR5 lead times for server-grade memory have already stretched to 52 weeks — nearly a year — compared to the normal six-week cycle, according to KB Securities researcher Kim Dong-won. That is an eightfold increase from baseline and represents an abnormal supply constraint that would normally signal a market on the verge of collapse. Instead, it is being absorbed by AI data center buildouts that show no sign of slowing.

HBM4 Is the Real Story

The newest generation of high-bandwidth memory, HBM4, is where the profit trajectory gets most interesting. Samsung’s HBM4 revenue is expected to more than triple from Q2, and it should account for more than 60% of the company’s total HBM sales in the second half of the year. By 2027, HBM4 is projected to represent 80% of HBM mix, up from an estimated 40% in 2026, with HBM4 revenue itself doubling year over year.

SK Hynix is tracking a similar curve. Analysts note that HBM adoption is expanding beyond GPUs into ASICs — application-specific integrated circuits designed for particular AI workloads. Customer diversification is accelerating, which reduces reliance on any single buyer and strengthens pricing power for the memory makers.

This creates a feedback loop that most models are not capturing. Every new HBM generation requires more complex stacking, more advanced packaging, and fewer dies per wafer compared to standard DRAM. The effect is compounding: as HBM becomes a larger share of the mix, the effective supply of commodity DRAM contracts further, pushing prices higher even as overall memory revenue grows. The companies are making more money on less physical product — a margin expansion story that is difficult to model linearly.

Why the Top Isn’t in Sight

The “top in sight” argument rests on a simple premise: prices rose fast, so they must fall soon. But that logic assumes the demand side is static. It is not. The global AI infrastructure buildout is still in its early innings, with cloud providers and hyperscalers committing capital that extends well into 2026 and 2027.

Yuanta Securities analyst Baek Kil-hyeon acknowledged that the pace of memory price increases may decelerate slightly, but projected that tight supply-demand conditions will persist through 2027 and possibly 2028. That is a significantly longer tail than most earnings models factor in.

What is also being underestimated is the pricing leverage shift. When a manufacturer controls 40% of its output with a product that has a year-long lead time and only two credible suppliers in the world, the commercial dynamics change fundamentally. Buyers cannot simply walk away and wait for prices to drop — there is no “wait” option when your lead time is 52 weeks. This structural advantage is likely to persist longer than the cyclical intuition suggests.

The Global Implications

For the broader semiconductor market, the Korean memory numbers are a leading indicator. They signal that demand for compute-intensive workloads is not saturating; it is expanding. Any model of the global chip market that treats memory as a separate, decoupled segment is missing the point. Memory is now the bottleneck that determines how fast AI training pipelines can run, and that bottleneck is getting tighter, not looser.

For investors and industry watchers, the takeaway is straightforward: the memory cycle has a longer runway than the surface numbers suggest. The 190 trillion won projection is not a peak — it is a milestone in what appears to be a multi-year supercycle driven by structural demand shifts, not temporary inventory buildup. The companies with HBM capacity and the pricing power to manage the commodity-drain effect will be the ones capturing the most value, and Samsung and SK Hynix are positioned to deliver for at least the next two years.