Memory stocks are posting record AI profits, but their valuations remain surprisingly low

The artificial intelligence boom has minted fortunes across the technology sector, but one critical corner of the industry is posting record-shattering profits while trading at bargain-basement prices. Memory semiconductor manufacturers—the companies that build the essential data-storage chips powering everything from smartphones to AI supercomputers—are having their best year in history. Yet, despite triple-digit earnings growth, Wall Street is treating them like a ticking time bomb.[1]

The disconnect between fundamental performance and market valuation is striking. During its most recent quarter, Micron Technology posted a staggering 756 percent year-over-year growth in earnings per share, while Samsung Electronics saw its earnings surge by nearly 500 percent.

Despite this massive influx of cash, shares of Micron, Samsung, and SK Hynix are trading at price-to-earnings ratios in the single digits. Micron trades at roughly 9 times forward earnings, while SK Hynix and Samsung sit at around 6.5 times—a fraction of Nvidia’s 23 times multiple and the broader S&P 500’s 20.3 average.[1]

To understand why investors are so terrified of these record profits, one has to look at the brutal history of the memory chip business. For three decades, the industry has followed a predictable and devastating pattern known as the "hog cycle." A surge in demand leads to skyrocketing prices, which prompts manufacturers to aggressively over-invest in new fabrication plants. When those massive facilities finally come online years later, they flood the market with supply, crashing prices and wiping out profit margins.[1][3]

This boom-and-bust rhythm has bankrupted companies and destroyed shareholder value so reliably that investors have learned to never trust the phrase "this time is different." And according to the historical clock, the current boom is living on borrowed time. We are currently 30 months removed from the industry's mid-2023 trough, which means this upcycle has already matched the duration of the longest historical peaks on record. By traditional metrics, the market should be bracing for an imminent crash.[3]

However, a growing chorus of industry analysts and hyperscale cloud providers argue that the artificial intelligence revolution has fundamentally rewritten the economics of memory. They contend that the industry is entering a "supercycle"—a structural shift where the insatiable demand for AI infrastructure creates a permanent demand floor, replacing the lumpy, consumer-driven cycles of the past.[2]

At the center of this disruption is a specialized technology called High Bandwidth Memory, or HBM. Unlike conventional DRAM, which arranges memory dies side-by-side, HBM stacks multiple layers of memory vertically—typically 12 to 16 layers deep—and connects them through microscopic through-silicon vias. This complex 3D architecture allows data to move at the unprecedented speeds required by advanced AI accelerators, making it the undisputed bottleneck in modern data centers.[2]

The manufacturing reality of HBM is what makes this cycle entirely different from previous booms. Producing HBM is incredibly complex and capacity-intensive. Industry data reveals a staggering "cannibalization ratio": manufacturing one bit of HBM requires approximately three times the wafer capacity of standard DDR5 memory. As manufacturers rush to meet AI demand, they are effectively deleting massive amounts of overall production capacity from the global supply chain.[3]

The scale of this reallocation is historic. By the end of 2026, HBM is projected to consume roughly 25 percent of total DRAM wafer production globally, up from 19 percent just a year prior. Because the industry operates as an effective oligopoly—with SK Hynix, Samsung, and Micron controlling over 95 percent of the market—this coordinated pivot toward high-margin AI chips is fundamentally altering the global supply balance.[3]

The immediate consequence of this shift is a cascading shortage that is actively starving the conventional memory ecosystem. By dedicating their most advanced fabrication lines to HBM, chipmakers have created a persistent, structural undersupply of the standard DDR4 and DDR5 memory used in personal computers, smartphones, and traditional enterprise servers. This dislocation has driven contract prices for commodity DRAM up by as much as 40 percent quarter-over-quarter.[3]

Furthermore, the financial mechanics of the AI boom look nothing like the consumer electronics cycles of the past. Hyperscalers like Amazon, Microsoft, and Google are not buying memory on the spot market; they are securing massive portions of future DRAM wafer output under multi-year forward agreements. This means that a significant percentage of the industry's 2026 and 2027 production capacity is already sold, effectively locking in revenue and preventing the sudden inventory gluts that triggered past crashes.

Despite these structural changes, cyclical skeptics remain unconvinced. They point out that while HBM is currently constrained, manufacturing yields will inevitably improve. Samsung, for instance, is rapidly ramping up its HBM capacity, and as production processes mature, the supply response could arrive several quarters earlier than consensus projections. When that happens, skeptics argue, the immutable laws of semiconductor overcapacity will reassert themselves.[2][3]

There are also geopolitical wildcards that could disrupt the oligopoly's pricing power. Chinese manufacturers like CXMT are aggressively pursuing HBM capabilities, backed by massive state subsidies. While export controls currently restrict their access to the most advanced lithography equipment, any breakthrough in domestic Chinese production could introduce a flood of cheap supply into the lower tiers of the memory market, undermining the current margin structure.

Meanwhile, the reality of this supercycle is creating a nightmare for hardware procurement teams. With memory costs soaring, original equipment manufacturers are being forced to adapt. Instead of building buffer inventories—a classic late-cycle behavior that usually precedes a crash—PC and smartphone makers are actively reducing the memory content in their devices to protect their own profit margins.[3]

Yet, even some of the most hardened bears are beginning to capitulate. Major investment banks that predicted a "DRAM winter" just a year ago have recently pivoted to supercycle calls, projecting continued price increases and raising price targets for companies like Micron. Historically, this kind of bear capitulation is a mid-cycle signal, suggesting that the peak may still be quarters, or even years, away.[3]

Ultimately, the memory sector in 2026 represents one of the most fascinating standoffs in modern finance. Wall Street is pricing these companies as if a devastating crash is just around the corner, anchored by decades of boom-and-bust trauma. But if the structural demands of artificial intelligence have truly broken the cycle, the market is currently offering the foundational companies of the AI era at the valuation of a dying industry.[4]