The triple cracks of the AI storage supercycle: profit structure misalignment, capacity mismatch, and warning of the 2027 cycle turning point

On April 30, 2026, Samsung Electronics released quarterly results that left the market holding its breath: total revenue of 133.9 trillion won, up 43% quarter-over-quarter, up sharply 69% year-over-year, and setting a new historical record for single-quarter revenue. Operating profit reached 57.2 trillion won, surging 185% quarter-over-quarter and jumping 756.1% year-over-year. Within that, the semiconductor equipment solutions division responsible for semiconductors reported revenue of 81.7 trillion won and operating profit of 53.7 trillion won—both hitting record highs for a single quarter in any year. The DX division posted revenue of 52.7 trillion won, but operating profit was only 3 trillion won—meaning semiconductors contributed 93.9% of the company’s total operating profit.

At nearly the same time, SK Hynix turned in a report showing operating profit soaring 405% year-over-year: revenue of 52.58 trillion won, up 198% year-over-year; operating profit of 37.61 trillion won, with an operating margin as high as 72%—exceeding TSMC’s 58% and Micron’s 67.6%. Since SanDisk, which was spun off from Western Digital a little over a year ago, listed at $38.50 in February 2025, its stock price has gained about 3,640% in total; it has also risen more than 460% year-to-date in 2026.

Yet beneath these astonishing figures, three cracks are quietly appearing—cracks that may not immediately change the direction of the cycle, but are enough to remind investors: even the grandest super cycle has its ridges and precipices.

Crack One: Samsung’s “Excess Profit” Dilemma — 48,000 Workers Were on the Brink of a Strike

On May 21, 2026, a large-scale strike that could have lasted 18 days and involved about 48,000 employees was paused at the last moment. Samsung Electronics’ labor and management, with the mediation of South Korea’s Minister of Employment and Labor, reached a preliminary wage agreement to allocate 10.5% of operating profit from the semiconductor business to a special bonus pool. Based on estimates, this year the average bonus for 28,000 employees in the memory chip division will be about 600 million won; employees in the mobile and TV divisions will receive only 6 million won—a gap of roughly 100 times.

Behind these numbers lies a deep internal split. Samsung’s semiconductor division contributes more than 93% of the company’s operating profit, while the profit margin of the DX division—which includes mobile and home appliances—has remained under pressure. Roh Tae-won, head of the mobile business, has internally warned that due to soaring storage chip procurement costs, in 2026 the mobile business may record the first full-year loss in its history. A comment from Song Jae-hyeok, a professor at Seoul National University, hits the issue squarely: “When PhD-level employees in the DX division receive only a 6 million won bonus, but high-school-educated production workers in the semiconductor division receive 600 million won, this structure is hard for employees to view as truly fair.”

On May 27, the union voted to approve the wage agreement with nearly 74% support, temporarily easing the strike crisis. But KB Securities analysts previously estimated that if the strike lasted 18 days, global DRAM supply could fall by 3% to 4%, and NAND supply by 2% to 3%—and even if it ultimately did not happen, the episode has already revealed a fact: when a super cycle creates unprecedented excess profits, the tension in internal distribution can become a black swan for the supply chain.

| Metric | Samsung Semiconductor Division | Samsung DX Division (Mobile/Appliances) | | --- | --- | --- | | 2026 Q1 Revenue | 81.7 trillion won | 52.7 trillion won | | 2026 Q1 Operating Profit | 53.7 trillion won | 3 trillion won | | Employee Annual Bonus (per person) | About 600 million won (memory division) | About 6 million won | | Key Momentum | HBM4 mass production + AI server demand | Galaxy S26 series flagship sales |

Crack Two: Traditional DRAM Profit Margins Overtake HBM — A Mismatch in Supply-and-Demand Logic

A signal that the market has severely underestimated comes from capacity decisions by Samsung and SK Hynix.

On April 14, SK Hynix cut its 2026 HBM4 shipment plan, reducing the supply to Nvidia by about 20% to 30% compared with the original plan. The company clearly stated that this is not demand deterioration, but a proactive optimization of product mix—capacity is being redirected toward HBM3E with higher profit margins and LPDDR products for server applications.

This aligns closely with the signal Samsung has sent. During Samsung’s Q1 earnings call, the company stated clearly that HBM4 achieved global first mass production in February, but the profit margin of traditional DRAM products is now overtaking HBM. According to industry data, in 2026 Q1, the contract price increase for standard DRAM was raised from the initial estimate of 55% to 60% at the start of the year to 90% to 95%; entering Q2, typical DRAM contract prices continued to rise quarter-over-quarter by 58% to 63%. The quarterly increase for consumer Mobile DRAM was even higher, reaching 80%. For example, DDR5 16Gb rose from 5.524 dollars to about 40 dollars within a year—an increase of 627%.

A typical comparison reveals a misalignment in profit ranking: because the chip area of HBM is far larger than that of traditional DRAM, the number of chips produced per wafer is lower—producing 1GB of HBM requires consuming 3 to 4 times the wafer area used for traditional DRAM. During a phase of soaring prices, unit profits for traditional DRAM may actually be higher. With profit maximization as the driver, Samsung and SK Hynix will naturally reallocate capacity—meaning the linear logic that “expanding HBM to ease supply” is being disrupted by a more complex profit ranking.

At a deeper level, this “crowding-out effect” is two-way: HBM consumes a large amount of wafer capacity (expected to account for 25% of total front-end capacity in the industry by the end of 2026, rising to 31% in 2027), which further shrinks traditional DRAM supply and pushes prices even higher; meanwhile, the high profits of traditional DRAM in turn incentivize original manufacturers to invest more into traditional capacity, slowing the pace of HBM expansion. When both categories are “tight” and “profitable” at the same time, the supply-side self-repair mechanism temporarily malfunctions.

Crack Three: Capital Expenditures Surge from $83 Billion to $144 Billion — Why Isn’t Capacity Release Keeping Up?

The paradox that most confuses the market is this: capital expenditures are expanding substantially, but supply tightness keeps getting worse rather than improving.

A research report released by JPMorgan on April 22 showed that the bank had raised its forecast for 2027 global storage industry capex from 83 billion dollars in September 2025 to 144 billion dollars, an increase of 74%. Samsung’s 2026 capital expenditure plan exceeds 110 trillion won (about 733 million dollars), setting a historical high and up 21.7% from 2025. SK Hynix is expected to spend about 20.5 billion dollars, up 17% year over year.

However, over the same period, the forecast for bit supply growth for DRAM and NAND was increased only by about 12% to 14% each. The surge in capex has not translated into capacity release—because of three layers of damping:

First, the die-area loss of HBM chips. Producing 1GB of HBM consumes 3 to 4 times the wafer area of traditional DRAM, and yields are only 50% to 60%. This structural loss cannot be fully offset by increasing equipment investment. Second, longer equipment lead times. Samsung ordered about 20 EUV lithography machines from ASML for a total amount exceeding 10 trillion won, but deliveries of these devices take 18 to 24 months; from construction to full ramp-up, any new production lines require at least five years. Third, original manufacturers’ “strategic restraint.” After experiencing huge losses in the previous cycle caused by misjudging demand, Samsung and SK Hynix have shown clear caution in expanding general DRAM capacity.

| Supply Constraint Dimension | Specifics | Impact on the Cycle | | --- | --- | --- | | HBM die-area loss | Under the same process, HBM occupies 3 to 4 times the wafer area of traditional DRAM | Reduces effective bit output by about 7% to 10% | | Extended equipment lead times | EUV lithography equipment delivery cycle of 18 to 24 months | New capacity release delayed until after the end of 2027 | | Strategic restraint by original manufacturers | New factories take at least five years from construction to reach full production | Limited incremental growth in traditional capacity | | Uncertainty in China’s capacity | Changxin Memory’s storage plan for a new factory to begin production in 2027 | Could become the biggest variable on the supply side |

At the same time, data from research firm Counterpoint is even more aggressive: between 2026 and 2027, DRAM annual capacity would need to grow at a 12% rate to ease shortages, but the actual growth rate is only about 7.5%. By the end of 2027, global DRAM supply can meet only about 60% of market demand.

The Triple Narrative of Bulls, Bears, and the Middle Faction

Regarding the sustainability of the AI storage super cycle, the market has formed three distinct narratives.

Bull narrative: structural shortages will last at least until the end of 2027. UBS expects the DRAM super cycle will persist through at least the end of 2027, with the share of HBM-dedicated capacity rising from 25% to 31%, “because the three oligopolists lack competitive expansion motivation.” JPMorgan believes this super cycle could extend into 2027 or 2028.

Bear narrative: a cycle reversal may arrive in the second half of 2027. On May 18, former Samsung Semiconductor president Kyung Gyu-hyun issued a major warning at a forum of the Korea National Academy of Engineering: Chinese companies are expanding capacity at an aggressive pace, and the global supply structure could see a major turn in the second half of 2027, with prices potentially reversing and falling. He predicts that, at the latest by the first half of 2028, the effect of price declines will become fully visible. If AI investment returns fall short of expectations, large technology companies may cut capital expenditures; at that time, not only prices would face pressure, but demand could also weaken.

Middle-faction narrative: the cycle will lengthen but not disappear; only the form of volatility will change. The “new business model” pioneered by SanDisk—locking in revenue for many years through long-term contracts worth at least $42 billion, with more than $11 billion in financial guarantees as downside protection—is changing the memory industry’s traditional quarterly quoting model. CEO David Goeckeler said, “The pain in this industry has always been the cyclical cycle of boom and bust.” The middle-faction view is representative: the memory market remains cyclical. Although this cycle may be longer and stronger, valuations have already priced in too many optimistic expectations—JPMorgan sets its target price for SK Hynix at 1 million won, based on a 2.7x price-to-book ratio, which is a 30% premium over the historical peak.

Scrutiny of Narrative Authenticity: Which Risks Are Being Underestimated?

Beneath the mainstream narratives above, three risks that are often overlooked deserve attention.

First, the risk of “hardware and software substitution.” In March 2026, the market suffered a wave of negative narrative after concerns that SRAM optimization at the hardware level and data compression at the software level might reduce the total amount of memory consumed by AI systems. The controversy itself shows that the story on the demand side is not a single, solid sheet.

Second, the ceiling of downstream resilience. In 2026 Q2, the server DRAM market has already shown clear signs of a standoff: upstream manufacturers insist on raising prices, but downstream customers refuse to accept high prices and instead purchase second-hand components and pulled-from-equipment parts. When such substitution behavior reaches a tipping point, it will form a real suppressive force on contract prices.

Third, the issue of valuation bubbles. SanDisk’s current price-to-sales ratio is about 16 times, up from only 4.5 times at the start of the year; its stock price once reached about $1,410, and its gain so far in 2026 exceeds 460%. Even if the company’s fundamentals are strong, such valuation expansion means that any news of order slowdowns could trigger sharp pullbacks. Samsung Electronics’ cumulative stock gain for the year is about 100%, and SK Hynix’s share price has also climbed sharply—market pricing already includes sustained high prosperity for the next two to three years, and that in itself is a risk.

Industry Impact Analysis: Who Is Paying the Price of the Super Cycle?

This super cycle has a price, and the price is being borne by weak links across the industry chain.

On the supply side, leading original manufacturers tightly control pricing through stepwise supply tightening strategies: moving from quota-based pickup and NCNR agreements to prepayment deposit-based procurement in April and May 2026, while also locking in long-term contract orders for 2027 to 2028. Small and medium buyers face a dilemma: either accept exorbitant spot prices or simply be unable to buy. SK Hynix has stated even more directly: in 2026, demand from all customers cannot be fully met, and suppliers hold absolute pricing power.

On the downstream consumer electronics side, the procurement cost of DRAM and NAND is expected to account for more than one-third of the total cost of entry-level smartphones by mid-2026. PC manufacturers are forced into panic buying, and some smaller brands are experiencing a “rationing era” in the supply chain—some large orders have lead times exceeding 20 to 40 weeks. Even Samsung’s own mobile division faces the risk of recording its first-ever annual loss in history due to soaring storage chip costs.

On the AI infrastructure side, shortages and high prices of storage chips are becoming one of the physical bottlenecks limiting the speed of AI compute deployment. Although hyperscale cloud providers have locked up most available capacity through long-term agreements, persistent supply constraints mean that not every AI project can secure the required hardware on time—perhaps the most intriguing irony of this cycle.

Conclusion

The AI memory super cycle is a real, profound structural transformation, not short-term sentiment trading—this should be confirmed first. From Samsung to SK Hynix to SanDisk, the financial data of the three core players all point in the same direction: demand for storage chips driven by AI is real, enormous, and still accelerating.

But super cycles never run on a smooth track. Samsung’s internal cracks, the profit ranking between traditional DRAM and HBM, and the time lag between capex and capacity release—these three cracks are not individually fatal, but together they form a potential trigger for a cycle reversal. Especially with uncertainties like China’s capacity variables and AI investment return rates still unresolved, 2027 could become the true watershed for the direction of this super cycle.

What the market truly needs to watch for is never when the super cycle will end, but whether enough people have prepared answers for that endpoint before it arrives.