Overview: The Dawn of the Memory Supercycle

The global semiconductor industry is witnessing a tectonic shift. Between 2025 and 2027, SK Hynix and Samsung Electronics — the world's first and second largest DRAM manufacturers — are executing what analysts are calling the most aggressive, most capital-intensive, and most strategically consequential capacity expansion in the history of the memory industry. The focal point of this expansion is not conventional DRAM or NAND flash, but a specialized, ultra-high-performance product category known as High Bandwidth Memory (HBM) — the indispensable neural substrate of the artificial intelligence revolution.

According to reports from Korean news outlets and corroborated by Data Center Dynamics, Samsung is targeting a production capacity increase of approximately 50% for HBM by end-2026, while SK Hynix has committed to doubling its HBM revenue in 2025 and aggressively ramping HBM4 volume production through 2026. UBS projects that SK Hynix will command approximately 70% market share in the HBM4 segment serving NVIDIA's next-generation Rubin GPU platform. By 2026, HBM is expected to account for roughly 25% of total DRAM wafer production, with year-on-year demand growing around 70%. [[0]](#__0) [[1]](#__1) [[2]](#__2)

This is not a cyclical upturn. This is a structural transformation. The expansion of HBM capacity by these two Korean giants is reshaping global supply chains, redefining national technology strategies, and determining who controls the computational infrastructure of the next decade. To understand why this is happening, we must examine the forces driving demand, the specific plans each company has laid out, and the long-term technological and geopolitical implications of this extraordinary moment.

Part I — Why Expand? The Root Causes Behind the HBM Rush

1.1 The AI Training Imperative: An Insatiable Appetite for Bandwidth

The proximate cause of the HBM expansion race is straightforward: the explosive growth of artificial intelligence workloads has created a demand for memory bandwidth that conventional DRAM architectures simply cannot satisfy. Training a frontier large language model (LLM) such as GPT-4 or its successors requires moving petabytes of data between GPU compute cores and memory every second. Standard DDR5, even at its theoretical peak, provides roughly 100 GB/s of bandwidth per module. HBM3E, by contrast, delivers over 1.2 TB/s — more than twelve times the throughput — by stacking multiple DRAM dies vertically and connecting them via thousands of Through-Silicon Vias (TSVs).

NVIDIA's H100 GPU, the workhorse of the current AI infrastructure buildout, uses 80GB of HBM3 providing 3.35 TB/s of aggregate bandwidth. The successor H200 uses HBM3E with 4.8 TB/s. The forthcoming Blackwell Ultra and Rubin architectures will push these figures even further. Each new generation of AI accelerator demands not just more HBM capacity, but higher-bandwidth, lower-latency, and more energy-efficient HBM. This creates a relentless, generation-over-generation pull on the supply chain that shows no sign of abating. [[1]](#__1)

Beyond training, the explosive growth of AI inference — running trained models at scale to serve billions of user queries — is creating an entirely new demand vector. SK Hynix has explicitly identified the AI inference market as a key growth driver, with reports indicating the company is betting on an 800% increase in DRAM demand from inference workloads over the coming years. Unlike training, which is concentrated in a relatively small number of hyperscale data centers, inference runs everywhere — in cloud servers, at the network edge, and increasingly in end-user devices. This democratization of AI deployment means that HBM demand is not a narrow niche but a broad, pervasive market force. [[3]](#__3)

1.2 The Memory Wall: AI's Most Fundamental Bottleneck

Computer scientists have long warned of the "memory wall" — the growing disparity between the speed at which processors can compute and the speed at which they can access data from memory. For decades, this problem was manageable because most workloads were not memory-bandwidth-bound. AI has changed everything. Modern transformer-based neural networks are almost entirely memory-bandwidth-bound: the GPU's compute cores sit idle for significant fractions of each cycle, waiting for data to arrive from memory. In this context, memory bandwidth is not merely a performance metric — it is the single most important determinant of AI system throughput and efficiency.

HBM was specifically engineered to attack the memory wall. By placing DRAM dies directly on top of or adjacent to the logic die (GPU or AI accelerator) within the same package, and connecting them with thousands of parallel TSV channels, HBM eliminates the long, slow, energy-hungry traces that connect conventional DRAM modules to processors. The result is not just higher bandwidth, but dramatically lower latency and power consumption per bit transferred — critical attributes for both training efficiency and inference cost. The memory wall problem will only intensify as AI models grow larger and more complex, making HBM's value proposition increasingly compelling and its demand structurally durable. [[2]](#__2)

1.3 Severe Supply Shortage and Historic Price Surges

A critical economic driver of the expansion is the extraordinary supply-demand imbalance that has emerged in the HBM market. Because HBM production is extraordinarily complex — requiring advanced 3D stacking, TSV drilling, precision bonding, and extensive testing — capacity cannot be ramped quickly. The lead time from investment decision to productive output is typically 18–24 months for a new HBM production line. This structural inertia means that the surge in AI-driven demand that began in earnest in 2023 created a supply deficit that has persisted and deepened through 2025 and into 2026.

The consequences for pricing have been dramatic. As noted in broader market analyses, the memory market is experiencing what some analysts describe as the "strongest price increase cycle in history," with contract prices for certain scarce DRAM specifications rising dramatically since early 2025. HBM pricing, which commands a significant premium over standard DRAM even in normal times, has seen even more pronounced appreciation. For SK Hynix and Samsung, this pricing environment transforms capacity investment from a defensive necessity into an extraordinarily attractive financial opportunity. Every additional HBM wafer produced in 2026 and 2027 will be sold into a market where supply remains constrained and buyers — hyperscalers, cloud providers, AI chip companies — are willing to pay premium prices to secure supply. [[0]](#__0) [[2]](#__2)

1.4 Geopolitical Imperatives and the Race for Technological Sovereignty

Beyond pure market economics, the HBM expansion is being shaped by profound geopolitical forces. The United States has implemented increasingly stringent export controls targeting advanced semiconductor manufacturing equipment, including restrictions on sub-18nm DRAM lithography tools and the revocation of Validated End User (VEU) exemptions that previously allowed some technology transfers. These controls are explicitly designed to prevent China from developing competitive advanced memory manufacturing capabilities, thereby preserving the technological advantage of U.S. allies — primarily South Korea and Taiwan.

For SK Hynix and Samsung, this geopolitical landscape creates both opportunity and urgency. The effective exclusion of Chinese competitors from the advanced HBM market (Chinese firms like CXMT are currently limited to older DRAM generations) means that the Korean duopoly faces no meaningful competitive threat from the world's largest manufacturing nation for the foreseeable future. At the same time, the South Korean government has designated semiconductor manufacturing as a strategic national priority, offering tax incentives, infrastructure support, and regulatory streamlining to encourage domestic capacity investment. The expansion plans of SK Hynix and Samsung are thus not purely corporate decisions — they are, in part, acts of national industrial policy. [[1]](#__1) [[3]](#__3)

Part II — The Expansion Plans: Strategic Deployments in Detail

2.1 SK Hynix: Consolidating HBM Dominance

SK Hynix enters this expansion cycle from a position of extraordinary strength. The company was the first to commercialize HBM3E, secured the dominant position as NVIDIA's primary HBM supplier, and currently holds an estimated 60%+ share of the overall HBM market. Its strategic objective is not merely to grow with the market, but to deepen its technological lead and lock in long-term supply relationships with the world's most important AI chip customers.

SK Hynix's expansion strategy is characterized by a deliberate focus on quality over quantity. Rather than simply adding wafer starts, the company is investing heavily in advanced packaging technology, yield engineering, and customer co-development programs. Its close collaboration with NVIDIA — which reportedly involves joint development of HBM specifications tailored to specific GPU architectures — gives it a structural advantage that pure capacity expansion cannot replicate. [[0]](#__0) [[1]](#__1)

2.2 Samsung Electronics: A $73 Billion Bet to Reclaim Leadership

Samsung's position in the HBM expansion race is more complex and, in some ways, more dramatic. As the world's largest semiconductor company by revenue, Samsung has the financial firepower to execute the most ambitious expansion plan in the industry. However, it enters this cycle from a position of relative competitive disadvantage in HBM specifically — having lost ground to SK Hynix due to yield challenges and qualification delays with key customers including NVIDIA.

Samsung's response has been to commit extraordinary capital. The company announced a semiconductor investment plan totaling approximately $73 billion (100 trillion Korean Won) over the period through 2026, with a substantial portion directed toward HBM and advanced DRAM capacity. This represents one of the largest single-company semiconductor investment commitments ever announced. [[0]](#__0)

2.3 Comparative Strategic Analysis

Sources: Data Center Dynamics, SK Hynix Newsroom, CNBC, SemiCone [[0]](#__0) [[1]](#__1) [[2]](#__2) [[3]](#__3)

Part III — Future Demand: The Deep Logic of the Memory Supercycle

3.1 HBM Market Size Projections

The scale of the demand opportunity that SK Hynix and Samsung are racing to capture is extraordinary. The broader memory market — already massive — is undergoing a structural expansion driven by AI that is without historical precedent. The global memory market was valued at approximately $220 billion in 2025, with projections suggesting it could reach $890 billion by 2026 as AI-driven demand accelerates across all segments. Within this broader market, HBM is the fastest-growing and highest-margin segment.

HBM demand is expected to grow approximately 70% year-on-year in 2026, driven by the continued buildout of AI training infrastructure and the rapid expansion of AI inference capacity. By 2026, HBM is projected to account for roughly 25% of total DRAM wafer production — up from a low single-digit percentage just three years earlier. This represents a fundamental reorientation of the entire DRAM industry toward a single, AI-driven application segment. [[2]](#__2)

Looking further ahead, multiple industry analysts project that the HBM market alone could exceed $50–60 billion annually by 2028–2030, representing a compound annual growth rate (CAGR) of 40–50% from current levels. These projections are grounded in concrete demand signals: NVIDIA, AMD, Google, Microsoft, Amazon, and Meta have all publicly committed to multi-year, multi-billion-dollar AI infrastructure investments that will require enormous quantities of HBM. [[1]](#__1)

3.2 Multi-Dimensional Demand Drivers

3.3 Supply-Demand Gap and Price Dynamics

Despite the aggressive expansion plans of SK Hynix and Samsung, the consensus view among industry analysts is that the HBM market will remain supply-constrained through at least 2027. The reasons are structural: HBM production is extraordinarily capital-intensive and technically complex, with long lead times between investment and output. As manufacturers shift an increasing share of DRAM wafer capacity toward HBM — projected to reach 25% of total wafers by 2026 — the supply of conventional DRAM (DDR4, DDR5, LPDDR5) is being squeezed, creating secondary shortages and price pressures across the entire memory market. [[2]](#__2)

The global DRAM supply-demand ratio is expected to reach approximately -12% by 2026, indicating a severe structural shortage. In this environment, HBM pricing will remain elevated, supporting the extraordinary profitability of SK Hynix and Samsung's HBM operations. For context, HBM already commands a price premium of 5–8x over equivalent-capacity standard DRAM, and this premium is expected to remain substantial through the forecast period. [[0]](#__0) [[1]](#__1)

Part IV — Future Development: Technology Roadmap & Industry Evolution

4.1 HBM4 and HBM4E: The Next Frontier

The HBM technology roadmap is advancing at a pace that would have seemed implausible just five years ago. HBM3E, which delivers approximately 1.2 TB/s of bandwidth, represents the current state of the art in volume production. HBM4, the next generation, is expected to deliver approximately 2.0 TB/s — a 67% improvement — through a combination of wider I/O interfaces, higher operating frequencies, and improved die-to-die interconnect technology.

Samsung's shipment of industry-first HBM4E samples in May 2026 — confirmed by CNBC — represents a significant milestone. HBM4E is described as offering higher bandwidth than HBM4, potentially positioning Samsung to leapfrog SK Hynix in the next product cycle if it can achieve competitive yields at volume. [[2]](#__2) SK Hynix, meanwhile, is already engaged in HBM4 volume production and is developing HBM4E in parallel, with a focus on maintaining its bandwidth and yield leadership.

Beyond HBM4E, both companies are investing in research for HBM5 and beyond, which will likely incorporate new die bonding technologies (such as hybrid bonding, which eliminates the need for solder bumps and enables much finer interconnect pitches), new materials, and potentially new memory cell architectures. The bandwidth targets for HBM5 are expected to exceed 4.0 TB/s, with power efficiency improvements that will be critical for the thermal management of increasingly dense AI accelerator packages.

4.2 3D DRAM and the Post-Scaling Era

The conventional DRAM scaling roadmap — reducing the size of memory cells to pack more bits onto each wafer — is approaching fundamental physical limits. The industry is moving toward 3D DRAM architectures, which stack memory cells vertically rather than shrinking them horizontally, as the primary path to continued density improvement. Technologies such as Vertical Channel Transistors (VCT) and Vertical Gate (VG) DRAM are expected to enable sub-10nm equivalent memory densities by 2027–2028, paving the way for the next generation of HBM stacks with dramatically higher per-die capacity.

Both SK Hynix and Samsung are investing heavily in 3D DRAM research and development. The transition to 3D DRAM will require new manufacturing equipment, new process chemistries, and new design methodologies — representing both a challenge and an opportunity for the companies that can execute it first. The winner of the 3D DRAM race will likely hold a decisive competitive advantage in HBM for the following decade. [[1]](#__1) [[2]](#__2)

4.3 Industry Ecosystem Restructuring and Vertical Integration

The HBM expansion is not occurring in isolation — it is catalyzing a broader restructuring of the semiconductor ecosystem. Advanced packaging, which encompasses the TSV drilling, die stacking, and substrate integration required for HBM production, is becoming as strategically important as wafer fabrication itself. Both SK Hynix and Samsung are investing in advanced packaging capacity as a core competency, rather than outsourcing it to third-party OSATs (Outsourced Semiconductor Assembly and Test providers).

This vertical integration trend is reshaping the competitive landscape. Companies that control the full stack — from DRAM wafer fabrication through advanced packaging to final HBM module assembly — will have significant advantages in yield, quality, cost, and time-to-market. It is also driving consolidation in the supply chain, as equipment makers, materials suppliers, and EDA tool vendors align their roadmaps with the requirements of HBM production. [[0]](#__0) [[3]](#__3)

Part V — Future Prospects: Opportunities, Risks & Global Impact

5.1 Historic Opportunity Window

For SK Hynix and Samsung, the current moment represents a generational opportunity. The convergence of explosive AI demand, structural supply constraints, favorable pricing, geopolitical tailwinds (exclusion of Chinese competition), and government support creates a profit environment that may not recur for decades. Companies that invest aggressively now — building capacity, developing technology, and deepening customer relationships — will be positioned to capture disproportionate value from the AI infrastructure buildout for years to come.

SK Hynix's financial performance has already reflected this opportunity: the company reported record revenues and operating profits in 2024 and 2025, driven almost entirely by HBM. Samsung, despite its HBM challenges, has the financial resources and manufacturing scale to execute a recovery and potentially recapture market share in the HBM4/HBM4E generation. Both companies are, in effect, making a multi-decade bet on the centrality of memory bandwidth to the future of computing — a bet that the evidence strongly supports. [[1]](#__1) [[2]](#__2)

5.2 Core Challenges and Risks

5.3 Global Impact: Reshaping the AI Infrastructure Landscape

The expansion plans of SK Hynix and Samsung will have profound ripple effects throughout the global technology ecosystem. Most immediately, the increased supply of HBM — even if it remains below demand — will help moderate the extreme pricing pressures that have made AI infrastructure extraordinarily expensive. This has direct implications for the democratization of AI: lower HBM costs translate into lower GPU costs, lower cloud AI service costs, and ultimately more accessible AI capabilities for businesses and consumers worldwide.

The geographic diversification of HBM production — with both companies investing in U.S. facilities — will also reduce the concentration risk of having the world's AI memory supply concentrated in a single country. This is a deliberate policy objective of the U.S. CHIPS Act, and its success will have long-term implications for the resilience of global AI supply chains. [[0]](#__0) [[3]](#__3)

For the broader semiconductor ecosystem, the HBM expansion is driving investment and innovation across the entire supply chain. ASML's EUV lithography systems, Tokyo Electron's deposition equipment, JSR and Shin-Etsu's photoresists, and dozens of other specialized suppliers are all seeing increased demand driven by the HBM buildout. The economic multiplier effects of this investment — in jobs, in technology development, in national industrial capability — are substantial and will be felt for decades.

Finally, the HBM expansion is reshaping the competitive dynamics of the global AI race itself. Nations and companies that can secure reliable, affordable access to HBM will have a structural advantage in developing and deploying AI capabilities. Those that cannot — whether due to export controls, supply constraints, or cost — will face a growing disadvantage. In this sense, the expansion plans of SK Hynix and Samsung are not merely corporate investment decisions; they are acts that will shape the trajectory of artificial intelligence, and with it, the future of human civilization. [[1]](#__1) [[2]](#__2)

Conclusion

[[0]](#__0) [[1]](#__1) [[2]](#__2) [[3]](#__3)

Industry FAQ