AI reshapes the DRAM cycle: HBM erupts, Samsung ramps up capacity, and the era of severe shortages in storage chips in 2028 fully begins

On July 15, 2026, Samsung Electronics was reported to be planning to build a DRAM factory with a monthly capacity of 100k wafers in the Giheung-gi province Uijengwon area, with total investment reaching tens of trillion Korean won. Artificial intelligence is fundamentally reshaping the global supply-demand structure of storage chips and the logic behind pricing.

As a core component of AI infrastructure, DRAM (dynamic random-access memory) is undergoing a cycle reshaping driven by a sudden shift in demand structure. This reshaping is entirely different from traditional storage cycles: it is not driven by consumer electronics upgrade waves, but by the explosive demand for high-bandwidth memory caused by AI training and inference. Understanding the logic behind the evolution of this DRAM cycle is of important reference value for grasping the direction of the semiconductor industry and changes to the underlying infrastructure of the digital asset market.

Supply-demand gaps expand to a peak in nearly 15 years

As of July 2026, the global storage chip price increases have entered a continuous third consecutive quarter. In a report released in early July, UBS significantly raised its storage price forecasts, expecting DDR contract prices to rise 32% quarter-over-quarter in Q3 and 18% quarter-over-quarter in Q4. The firm believes that DRAM supply-demand tightness will last at least through the first half of 2028.

This assessment is cross-validated by multiple sets of data. TrendForce data from TrendForce Consulting shows that in Q3 2026, the overall DRAM landscape remains extremely tight, with contract prices expected to rise 13% to 18% quarter-over-quarter. In a report dated July 7, Morgan Stanley raised its forecast for the quarter-over-quarter increase in PC DRAM blended average selling prices from the previous 3%-8% to 15%-20%, and raised server DRAM to 13%-18%.

The depth of the supply-demand imbalance is approaching historical extremes. UBS analysis indicates that, excluding the buffering effect of downstream customers rebuilding inventories, the actual supply-demand gap in 2027 will further widen from -8.1% in 2026 to -13.6%, reaching a level rarely seen in the past 30 years. A Goldman Sachs research report judges that the current global storage supply-demand imbalance has reached a peak in nearly 15 years. Jin Dongwon, an analyst at KB Securities, also predicts that 2027 will be the most supply-constrained period in the 70-year history of the memory semiconductor industry.

Upward price pressure has moved from upstream original manufacturers to the entire industrial chain. Samsung Electronics plans to raise the quarter-over-quarter average selling price of general-purpose DRAM in Q3 2026 by 20%, and the increase in LPDDR mobile memory is also above 20%. According to Yicai Global, some consumer electronics terminal makers have already received verbal notices from Samsung about DRAM price hikes.

HBM becomes a structural variable of the DRAM cycle

The key to understanding this DRAM cycle lies in the rise of HBM (high-bandwidth memory). HBM stacks multiple DRAM chips vertically to achieve ultra-high bandwidth, making it an indispensable memory solution for AI GPUs.

The demand-side momentum is astonishing. Feng Li, president of SEMI China, clearly stated that the HBM market size in 2026 is expected to grow 58% to $100k, accounting for nearly four-tenths of the DRAM market. Although the three major original manufacturers—Samsung, SK hynix, and Micron—have directed 70% of incremental capacity toward HBM, the overall capacity shortfall still reaches 50% to 60%. Omdia data shows that global HBM output in 2026 is expected to grow 103% year over year, achieving more than a doubling.

HBM’s “squeezing-out effect” on DRAM capacity is the core mechanism behind supply-demand imbalances. Industry data shows that HBM’s share in DRAM wafer capacity has risen from 2% in 2020 to an estimated 25% in 2026. The capacity consumed by one HBM wafer is roughly equivalent to three DDR5 wafers. A single HBM chip is more than twice the size of a standard DRAM chip; under the same storage capacity, the wafer area and cleanroom space consumed by HBM are at least three times those of ordinary DRAM.

The direct consequence of capacity shifting is a structural contraction in the supply of general-purpose DRAM. KB Securities expects that HBM’s share in global DRAM wafer output will increase from 15% in 2026 to 34% in 2027. This means that most incremental capacity will be concentrated in HBM, and the supply expansion of general-purpose DRAM will be effectively constrained.

AI servers have become the largest application market for DRAM, with server demand breaking through 50% of total DRAM demand. DRAM usage per AI server is 8 to 10 times that of a typical server. The average DRAM load per server is expected to rise from 1,032 GB in 2025 to 1,432 GB in 2026, an increase of about 39%. TrendForce estimates that in 2026, about 70% of global high-end memory capacity will flow to AI data centers.

The capacity race and share game among the three giants

In the new cycle driven by HBM, the strategic choices of the three major players—SK hynix, Samsung Electronics, and Micron—are reshaping the industry landscape.

SK hynix is currently the absolute leader in the HBM market. Counterpoint Research data shows that in Q1 2026, by revenue share, SK hynix ranked first with a 58% share in the global HBM market, while Samsung Electronics and Micron each accounted for 21%. Analysts predict that SK hynix’s 2026 HBM revenue could reach $5.95 billion. SK hynix’s full-year 2026 HBM capacity has already been fully locked in under long-term agreements with cloud vendors.

However, its lead faces challenges. A report from Korea Investment & Securities on July 13 projected that SK hynix’s Q2 revenue would be 80.9 trillion won (up 264% year over year), operating profit 60.4 trillion won (up 556% year over year), but about 8% below market consensus of 65 trillion won. The main reason is that SK hynix’s HBM sales mix has a higher share than competitors, leading to an average selling price increase lower than the market average. Analysts expect that starting from mass production ramp of HBM4 in Q3, the average selling price increase will return to the market mean.

Samsung Electronics is playing catch-up with full force. The company plans to invest more than 110 trillion won (about $73.3 billion) in equipment investment and R&D in 2026, up about 22% from the previous year. The latest news on July 15 indicates that Samsung plans to build a DRAM factory in the Uijengwon area with monthly capacity of 100k wafers, with construction potentially starting as early as Q3 2026.

On the HBM4 product, Samsung has already achieved the world’s first mass production shipment. In a June report, TrendForce noted that significant divergence has emerged among the three major HBM4 original manufacturers in their validation progress, with Samsung leading in validation. Samsung expects its 2026 HBM sales to grow more than threefold compared with the previous year.

Micron, although it has the smallest HBM capacity among the three, has the fastest growth momentum. In its fiscal third quarter of 2026 (through May), the company’s revenue reached $41.46 billion, surging 346% year over year. DRAM business contributed $31.3 billion, accounting for 76% of total revenue. Gross margin was 84.9%, surpassing NVIDIA.

Even more notable is the guidance for the fourth fiscal quarter: revenue is expected to be about $50 billion, gross margin about 86%, and earnings per share about $31. Micron’s full-year 2026 HBM capacity is already sold out under fixed-price contracts. The company’s equipment investment plan is set to increase by more than 90% compared with the previous year.

From cycle to structure: DRAM’s long-term logic shift

The most fundamental difference between this DRAM market cycle and all prior cycles is a structural shift in the driving force.

IDC expects global DRAM market revenue to reach $418.6 billion in 2026, up 177% year over year. NAND market revenue is expected to reach $174.1 billion, up 138.5% year over year. Total annual output value of the storage industry will exceed $550 billion, with server storage replacing phones to become the largest demand scenario. UBS believes that rising prices will drive the memory industry to achieve $992 billion in revenue in 2026.

This growth is not a simple cyclical recovery. IDC clearly points out that this is not another upswing within the traditional memory price-rise-and-fall cycle, but a structural transformation in the industry. Ultra-large-scale data center customers are buying a completely different—and more expensive—type of storage chip, and are willing to pay a premium to ensure supply.

The supply-side constraint also has structural characteristics. After the industry’s deep losses in 2025, original manufacturers generally adopted a strategy of controlling production to support prices. Semiconductor advanced process R&D and fab construction have long timelines, and multiple barriers such as talent, energy, and approvals prevent new capacity from being deployed quickly. A Micron executive judged that the tight supply landscape in the industry will last beyond 2027, with supply-side improvement gradually arriving only in 2028.

The widespread adoption of long-term agreements (LTA) is changing the industry’s pricing and profitability model. SK hynix, Samsung, and Micron all lock in future supply quantities and prices for the next 1 to 2 years with major customers such as NVIDIA through long-term supply agreements. Micron has signed 16 long-term supply agreements covering data centers, consumer electronics, and automotive sectors; the minimum guaranteed revenue from the already signed agreements alone is as high as $100 billion. Korea Investment & Securities noted that as the memory industry shifts to an LTA structure with tenures of 3 to 5 years, corporate value depends on how long high profitability can be sustained, rather than on the quarterly average selling price growth rate.

Conclusion

AI server demand is pushing the DRAM industry into an unprecedented new cycle. The core feature of this cycle is not just price increases, but a systematic reshaping of demand structure, capacity allocation, pricing models, and industry landscape.

HBM’s rise changes the DRAM product matrix; capacity skewing by the three giants creates a structural shortage in general-purpose DRAM; and the spread of long-term agreements is smoothing out the industry’s cyclical volatility. UBS expects DRAM supply-demand tightness to last at least through the first half of 2028, and KB Securities predicts that 2027 will be the most supply-constrained period in the industry’s 70-year history—these judgments point to the same conclusion: the duration and intensity of this DRAM cycle may far exceed the market’s initial expectations.

For investors, understanding DRAM’s transition from a “cyclical product” to a “strategic asset” may be more valuable long term than forecasting the next quarter’s price increase.

When storage chips become core infrastructure of the AI era, the supply-demand logic and distribution of pricing power will have a profound impact on the entire technology industry chain—from semiconductors to digital assets.

FAQ

Q: How much higher is the DRAM demand for AI servers than for ordinary servers?

DRAM usage per AI server is 8 to 10 times that of a typical server. The average DRAM load per server is expected to increase from 1,032 GB in 2025 to 1,432 GB in 2026. Scaling up large-model inference and AI agent deployments will further drive rapid growth in storage demand.

Q: What are the differences in production cost between HBM and traditional DRAM?

HBM is made by vertically stacking multiple DRAM chips. The size of a single chip is more than twice that of a typical DRAM chip; with the same storage capacity, the wafer area and cleanroom space consumed are at least three times those of ordinary DRAM. The production capacity consumed by one HBM wafer is roughly equivalent to three DDR5 wafers. Higher production costs also correspond to higher gross margins—gross margin for a single HBM chip is 60%-70%, which is 4 times that of standard DRAM.

Q: When is the DRAM supply-demand tightness expected to ease?

UBS believes DRAM supply-demand tightness will last at least through the first half of 2028. KB Securities predicts that the supply shortage is likely to continue through at least 2028. A Micron executive judged that the tight supply landscape will last beyond 2027, with supply-side improvement arriving gradually only in 2028. The core disagreement among the three institutions is not “whether it will be tight,” but “how long it will be tight.”

Q: What is the market share distribution among the three HBM players?

In Q1 2026, SK hynix held the top spot with a 58% revenue share, while Samsung Electronics and Micron each had 21%. For full-year shipment share, analysts predict SK hynix at about 52%, Samsung at about 39%, and Micron at about 8%. In the HBM4 segment, Counterpoint Research expects SK hynix’s market share in 2026 to be about 54%, Samsung about 28%, and Micron about 18%. Samsung is leading in HBM4 validation progress.

Q: How does DRAM price increases affect the prices of consumer electronics terminal products?

DRAM price increases have been transmitted from upstream original manufacturers to consumer electronics terminal products. According to industry analysis, PC and smartphone brands are expected to gradually raise product prices by 15% to 20%. A “three prices in a day” phenomenon has appeared in distribution markets such as Huaqiangbei, and dealers recommend postponing purchases for non-essential needs. Automakers are also starting to re-evaluate certain configurations for intelligent driving features to control component costs.