HMC and HBM Market Size, Share & Trends Analysis Report By Product (Hybrid Memory Cube (Hmc), High-Bandwidth Memory (Hbm), Other), By Applications (Graphics, High-performance Computing, Networking and Data Centres), By End Use, By Region, and Segment Forecasts - 2035

HMC and HBM Market Overview

hmc and hbm market size was valued at USD 3725.3 million in 2025 and is poised to grow from USD 4626.82 million in 2026 to USD 8864.36 million by 2035, growing at a CAGR of 24.2% during the forecast period (2026-2035).

The HMC and HBM Market is expanding rapidly as artificial intelligence, accelerator-based computing, graphics processing, high-performance computing, networking, and hyperscale data centres require significantly greater memory bandwidth than conventional memory architectures can provide. High-Bandwidth Memory is estimated to represent approximately 82% of market demand in 2026, while Hybrid Memory Cube accounts for around 11% and Other approximately 7%. HBM now dominates new high-performance deployments because vertically stacked DRAM, through-silicon vias, wide interfaces, and advanced packaging enable several terabytes per second of memory bandwidth within compact footprints. Current HBM4 products exceed 2.8 TB/s per stack and use 2,048-bit interfaces, representing more than twice the bandwidth available from previous-generation HBM3E configurations. High-performance Computing is estimated to account for approximately 47% of application demand in 2026, followed by Networking and Data Centres at around 31% and Graphics at approximately 22%. Artificial intelligence training and inference have become the principal demand accelerators, pushing memory suppliers toward 12-layer, 16-layer, higher-capacity, and increasingly customized HBM architectures.

The United States represents a critical HMC and HBM Market because the country hosts leading AI accelerator designers, hyperscale cloud providers, semiconductor companies, data-centre operators, high-performance computing institutions, and advanced packaging initiatives. Micron, Advanced Micro Devices, and Intel provide U.S. representation within the supplied competitive landscape, while Korean HBM suppliers play a central role in supplying U.S.-designed AI platforms. Micron entered high-volume production of 36GB 12-layer HBM4 during the first quarter of 2026, delivering more than 2.8 TB/s bandwidth and over 20% better power efficiency than its HBM3E generation. The company has also sampled 48GB 16-layer HBM4, increasing capacity per HBM placement by approximately 33% compared with 36GB 12-layer products. AI accelerators increasingly combine multiple HBM stacks within a single package, making memory bandwidth one of the principal system-performance constraints. U.S. demand is consequently moving beyond raw capacity toward bandwidth per watt, thermal efficiency, advanced packaging yield, stack height, signal integrity, and tighter co-design between processors and memory.

• 

  Download Free sample to learn more about this report.

Key Findings

• Leading Product Type: High-Bandwidth Memory (Hbm) is expected to lead with approximately 82% market share in 2026 as AI accelerators increasingly require multi-terabyte-per-second bandwidth and vertically stacked high-capacity memory.
• Leading Application: High-performance Computing is projected to account for approximately 47% of demand in 2026, driven by AI training, inference, scientific computing, accelerator systems, and memory-intensive data-processing workloads.
• Leading Region: Asia-Pacific is estimated to hold approximately 58% market share in 2026, supported by dominant HBM manufacturing capabilities in South Korea and extensive semiconductor fabrication and packaging infrastructure across the region.
• Fastest Growing Region: North America is projected to expand at approximately 27.6% annually as hyperscale AI infrastructure, accelerator deployment, advanced data centres, cloud computing, and domestic semiconductor investment accelerate memory consumption.
• Technology Trend: HBM4 is redefining bandwidth performance with a 2,048-bit interface and commercial products delivering more than 2.8 TB/s per stack, substantially increasing data movement for next-generation AI processors.
• Market Driver: Artificial intelligence remains the strongest demand catalyst, with Samsung expecting its HBM sales to increase by more than 3 times in 2026 compared with 2025 as AI infrastructure expands.
• Competitive Landscape: The supplied landscape contains 5 companies across 2 countries, while competition is intensifying around 12-layer and 16-layer stacks, custom base dies, packaging yield, thermals, capacity, and bandwidth.
• Future Outlook: HBM4E will push memory performance further, with 12-layer samples reaching up to 16 Gbps per pin and approximately 4.0 TB/s bandwidth for next-generation AI computing platforms.

Latest Trends

The strongest trend shaping the HMC and HBM Market in 2026 is the transition from HBM3E into HBM4 and early HBM4E as AI accelerators demand wider interfaces, higher stack capacities, and improved power efficiency. Commercial HBM4 uses a 2,048-bit interface, doubling the bus width compared with previous-generation HBM designs. Micron's 36GB 12-layer HBM4 delivers more than 2.8 TB/s per stack at speeds above 11 Gbps per pin and offers over 20% better power efficiency than HBM3E. Samsung's commercially shipped HBM4 operates at a consistent 11.7 Gbps and can scale to approximately 13 Gbps. These performance increases are changing memory from a supporting component into a primary AI-system design constraint. High-performance Computing, estimated at approximately 47% of 2026 demand, benefits directly because model training and inference increasingly require processors to move enormous quantities of parameters, activations, and intermediate data. Suppliers are therefore optimizing DRAM dies, logic base dies, through-silicon vias, bonding, thermal resistance, and package architecture as one integrated system.

Another important trend is the migration toward higher layer counts and customized HBM. Micron has sampled 48GB 16-layer HBM4, providing approximately 33% more capacity per stack than its 36GB 12-layer product, while Samsung is developing HBM4E and custom HBM architectures. Samsung's 12-layer HBM4E samples reach speeds of up to 16 Gbps and approximately 4.0 TB/s bandwidth, while next-generation hybrid copper bonding is being developed to support 16 or more layers and reduce thermal resistance by more than 20% compared with conventional thermal-compression bonding. This shift is particularly important for Networking and Data Centres, which are estimated to represent approximately 31% of application demand. Hyperscale AI systems increasingly require thousands of accelerators connected through high-speed fabrics, making capacity per package and bandwidth per watt important at both processor and data-centre levels. The market is therefore moving toward closer memory-processor co-design, customized logic base dies, higher-density stacks, and improved cooling architectures.

Market Dynamics

Driver

""Rapid AI accelerator deployment is creating unprecedented demand for high-bandwidth stacked memory.""

The principal driver of the HMC and HBM Market is the extraordinary growth of artificial intelligence computing, where memory bandwidth increasingly determines how effectively GPUs and specialized accelerators can utilize their computational resources. High-performance Computing represents approximately 47% of estimated market demand in 2026 because AI training and inference require simultaneous access to billions or trillions of model parameters. Conventional memory channels cannot provide sufficient bandwidth without increasing power consumption and board complexity, making HBM's extremely wide interfaces increasingly important. Current HBM4 products exceed 2.8 TB/s per stack, while HBM4E samples can reach approximately 4.0 TB/s. Multiple stacks integrated around an accelerator can therefore provide aggregate bandwidth measured in tens of terabytes per second. This capability allows processors to spend less time waiting for memory transfers and more time performing calculations.

Supplier capacity expansion reinforces this driver. Samsung expects its HBM sales to increase by more than 3 times during 2026 compared with 2025 and is proactively increasing HBM4 manufacturing capacity. Micron entered high-volume HBM4 production during the first quarter of 2026, while SK Hynix remains a major supplier to leading AI accelerator ecosystems. High-Bandwidth Memory is estimated to capture approximately 82% of total HMC and HBM demand in 2026, demonstrating how strongly current investment has shifted toward HBM architectures. Networking and Data Centres represent another approximately 31% of demand because hyperscalers deploy accelerator clusters containing hundreds or thousands of processors. As cluster sizes expand, memory bandwidth, capacity, thermals, and energy efficiency become increasingly important to total system performance and operating economics.

Restraint

""Complex stacking and advanced packaging requirements constrain manufacturing yield and supply scalability.""

The primary restraint is the manufacturing complexity associated with vertically stacking multiple DRAM dies, integrating logic base dies, forming through-silicon vias, controlling stack height, bonding dies, and achieving acceptable thermal behavior. A 12-layer HBM stack requires reliable interconnection across 12 DRAM dies plus the base structure, while 16-layer configurations increase the number of interfaces and potential defect points further. Micron's 48GB 16-layer HBM4 increases capacity per placement by approximately 33%, but achieving this density demands highly precise thinning, bonding, alignment, and thermal management. If the effective yield of each stacked stage is not tightly controlled, cumulative package yield can decline substantially as layer counts rise. Manufacturers therefore require advanced inspection, known-good-die screening, and packaging technologies before final assembly.

Thermal density presents another restraint because several HBM stacks operate extremely close to processors consuming hundreds or even more than 1,000 watts at package level in advanced AI systems. Wider interfaces and faster signaling increase performance but can also increase thermal and signal-integrity complexity. Samsung's next-generation hybrid copper bonding approach targets more than 20% lower thermal resistance compared with thermal-compression bonding, illustrating the scale of thermal optimization required for future 16-layer and higher configurations. HBM's premium manufacturing requirements also restrict the number of suppliers capable of producing competitive products at scale. The supplied landscape contains only 5 companies, while the actual leading memory manufacturing base is even more concentrated. This concentration can create capacity bottlenecks when AI demand expands faster than cleanroom, wafer, advanced packaging, and substrate supply.

Opportunity

""HBM4E and custom HBM create substantial opportunities for memory-processor co-design.""

Customized HBM represents one of the strongest opportunities through 2035 because accelerator designers increasingly need memory architectures optimized for specific computational workloads rather than standardized interfaces alone. Samsung plans to begin custom HBM sampling during 2027, following commercial HBM4 and HBM4E development. HBM4 already incorporates increasingly sophisticated logic base dies, and future custom products can optimize interface logic, power management, error handling, bandwidth allocation, and accelerator communication. High-performance Computing accounts for approximately 47% of demand, providing a large customer base for co-designed memory. Customization can also help system designers improve energy efficiency because moving data frequently consumes significant power in AI systems. Reducing unnecessary data movement by even 10% can materially affect operating efficiency across thousands of accelerators.

Higher-capacity stacks create another major opportunity. Micron's 48GB 16-layer HBM4 provides approximately 33% more capacity per placement than its 36GB 12-layer product, allowing accelerator packages to support larger AI models without increasing the number of HBM footprints. Samsung's HBM4E reaches approximately 16 Gbps per pin and 4.0 TB/s per stack, providing further bandwidth expansion. Networking and Data Centres, estimated to represent approximately 31% of demand in 2026, can benefit as AI inference increasingly handles long context windows, multimodal workloads, and large retrieval systems. More memory per accelerator can reduce the need to distribute model parameters across additional processors, potentially simplifying communication and lowering latency. Suppliers capable of combining higher capacity, greater bandwidth, improved power efficiency, and advanced bonding should capture increasing demand.

Challenge

""Power density and thermal management are becoming critical constraints as HBM bandwidth accelerates.""

The most important technical challenge is managing heat while increasing memory bandwidth and stack density. HBM4 doubles interface width to 2,048 bits and delivers more than 2.8 TB/s per stack, while HBM4E reaches approximately 4.0 TB/s. These gains place more memory channels and active circuitry within very limited package area. AI accelerators also generate substantial heat, creating a thermal environment where memory must operate close to high-power logic. A package containing 8 HBM stacks, each delivering 2.8 TB/s, could theoretically provide more than 22 TB/s of aggregate memory bandwidth, but such density increases cooling, power-delivery, and mechanical-design challenges. Poor thermal control can reduce operating frequency, increase error rates, and shorten component life.

Industry roadmaps toward 16 or more layers intensify the challenge because thicker stacks increase the distance heat must travel from upper dies toward cooling structures. Samsung's hybrid copper bonding technology aims to reduce thermal resistance by more than 20%, demonstrating the importance of packaging innovation alongside DRAM scaling. Competitive pressure is also accelerating product cycles: HBM4 entered mass production during 2026 while HBM4E samples were already being shipped only months later. Suppliers must therefore qualify new processes, improve yields, and expand capacity within increasingly compressed development windows. High-Bandwidth Memory already represents approximately 82% of estimated market demand, meaning manufacturing delays at one generation can materially affect supplier positioning. Achieving performance, thermals, reliability, and volume yield simultaneously will remain one of the defining challenges of the market through 2035.

• 

  Download Free sample to learn more about this report.

Segmentation Analysis

By Types

Hybrid Memory Cube (Hmc): Hybrid Memory Cube (Hmc) is estimated to account for approximately 11% of HMC and HBM Market demand in 2026, positioning it well below High-Bandwidth Memory (Hbm) but preserving relevance in specialized high-throughput memory architectures. HMC introduced an important architectural shift by stacking multiple DRAM layers above a logic layer and connecting them through through-silicon vias, demonstrating how three-dimensional integration could overcome bandwidth limitations associated with conventional planar memory. Its serial interface approach offered high aggregate throughput and reduced board-level interconnection requirements, but the industry increasingly standardized around HBM for mainstream accelerator and graphics architectures. HMC's estimated 11% share is approximately 71 percentage points below HBM, illustrating the scale of this technology transition. The segment remains technically relevant because several principles associated with HMC, including vertical integration, logic-assisted memory management, short interconnects, and three-dimensional packaging, continue influencing advanced memory development. High-performance Computing accounts for approximately 47% of total application demand, but current AI accelerators overwhelmingly favor HBM ecosystems because standardized interfaces and established packaging supply chains simplify integration. HMC is therefore expected to remain a specialized category through 2035 rather than returning to leadership.

High-Bandwidth Memory (Hbm): High-Bandwidth Memory (Hbm) is estimated to dominate the HMC and HBM Market with approximately 82% share in 2026. The segment's leadership is driven by AI accelerators, high-performance computing, data-centre processors, graphics architectures, and increasingly bandwidth-intensive networking systems. HBM4 represents a major technical transition because its interface expands to 2,048 bits, twice the bus width of the previous generation. Current 12-layer HBM4 products can operate above 11 Gbps per pin and deliver more than 2.8 TB/s of bandwidth per stack, more than doubling bandwidth compared with HBM3E configurations while improving power efficiency by over 20%. A processor package using 8 such stacks could theoretically access more than 22 TB/s of aggregate memory bandwidth, illustrating why HBM has become fundamental to AI system design. High-performance Computing represents approximately 47% of application demand, while Networking and Data Centres contribute another 31%, meaning approximately 78% of the market is associated with compute-intensive and infrastructure-oriented environments. HBM suppliers are now advancing from 12-layer toward 16-layer products, with 48GB 16-layer HBM4 increasing capacity per stack by approximately 33% compared with 36GB configurations. HBM4 validation and shipment timing is also influencing supplier positioning during 2026 as next-generation AI platforms move toward volume deployment.

Other: Other is estimated to represent approximately 7% of HMC and HBM Market demand in 2026, covering specialized stacked-memory configurations and related architectures outside the principal HMC and HBM categories. Although this segment remains 75 percentage points below HBM, it provides an innovation space for memory systems optimized around specialized processors, research platforms, or application-specific computing. Processing-in-memory is one development direction because moving selected calculations closer to stored data can reduce the energy and latency associated with repeatedly transferring information between processors and memory. Experimental HBM-PIM research published in 2026 demonstrated matrix-related performance reaching approximately 14.9 GFLOP/s on a single HBM pseudo-channel, highlighting the potential for memory to participate more directly in computational workloads. Other architectures can also explore alternative interconnects, logic integration, error management, and packaging techniques. The segment is unlikely to challenge HBM's approximately 82% share in mainstream AI acceleration during the forecast period, but specialized technologies may influence future HBM generations. As model sizes, context lengths, and inference volumes increase, reducing data movement becomes increasingly important because memory bandwidth and energy consumption can limit processor utilization.

By Applications

Graphics: Graphics is estimated to account for approximately 22% of HMC and HBM Market demand in 2026, supported by professional visualization, high-end graphics processors, content creation, rendering, simulation, and graphics-intensive computing. HBM was originally commercialized prominently in graphics products because its wide interface offered substantially greater bandwidth within a smaller package footprint than conventional external memory arrangements. Modern graphics workloads process high-resolution textures, geometry, ray-tracing data, frame buffers, and increasingly AI-assisted rendering operations. A 4K image contains more than 8 million pixels per frame, while 8K exceeds 33 million pixels, placing growing pressure on memory bandwidth when high frame rates and complex effects are combined. Graphics represents a smaller share than High-performance Computing at approximately 47%, but it remains an important technology-development environment because many GPU architectures serve both graphics and AI workloads. HBM4 bandwidth above 2.8 TB/s per stack provides significant headroom for next-generation visualization systems, although cost considerations mean HBM is generally concentrated in premium graphics rather than mainstream consumer configurations. Advanced workstation and professional graphics systems should continue using high-bandwidth memory where performance density justifies the greater packaging complexity.

High-performance Computing: High-performance Computing is estimated to dominate the HMC and HBM Market with approximately 47% share in 2026. Artificial intelligence training and inference are the most important growth engines within this application because accelerator performance increasingly depends on moving model parameters and intermediate data rapidly enough to keep computational units active. HBM4 provides a 2,048-bit interface and more than 2.8 TB/s of bandwidth per stack, allowing multi-stack accelerator packages to reach aggregate memory bandwidth exceeding 20 TB/s. The transition is accelerating as HBM4 enters mass production and HBM4E development advances toward 2027. The global semiconductor market is expected to expand by more than 25% during 2026, with memory semiconductors projected to grow around 30%, reflecting the disproportionate impact of AI infrastructure on memory demand. High-performance Computing applications also include scientific simulations, engineering analysis, computational research, and large-scale numerical workloads where processors repeatedly access extensive datasets. HBM's proximity to the processor reduces interconnect distance and enables wide parallel data transfer. With approximately 47% share, this application is 16 percentage points ahead of Networking and Data Centres and 25 percentage points above Graphics, reinforcing its position as the primary demand center through 2035.

Networking and Data Centres: Networking and Data Centres are estimated to represent approximately 31% of HMC and HBM Market demand in 2026, making the segment the second-largest application. Hyperscale infrastructure increasingly combines thousands of AI accelerators, high-speed network interfaces, switches, storage systems, and specialized processors, creating demand for memory architectures that support rapid data movement with controlled energy consumption. HBM is especially important in AI servers because accelerator packages can integrate several stacks directly beside compute dies. Current HBM4 provides more than 2.8 TB/s per stack and over 20% better power efficiency than comparable previous-generation HBM3E implementations, helping operators improve bandwidth without proportional increases in memory energy. North American cloud service providers accelerated AI infrastructure investment from late 2025 as inference and AI-agent workloads expanded, strengthening demand for HBM4 qualification and supply during 2026. Networking workloads can also benefit from high-bandwidth memory where packet processing, routing tables, analytics, or accelerated data handling require rapid access to large datasets. The approximately 31% application share is 9 percentage points above Graphics but 16 percentage points below High-performance Computing. Continued expansion of AI clusters should increasingly blur the distinction between HPC and data-centre demand because large cloud facilities now operate some of the world's most powerful computational systems.

• 

  Download Free sampleto learn more about this report.

Regional Outlook

North America

North America is estimated to account for approximately 28% of HMC and HBM Market demand in 2026 and is projected to expand at around 27.6% annually, making it one of the fastest-growing regional markets. The United States contains Advanced Micro Devices, Intel, and Micron among the 5 supplied companies, meaning 3 participants, or approximately 60% of the supplied competitive group, are U.S.-based. Regional demand is heavily influenced by hyperscale cloud infrastructure, AI accelerator design, high-performance computing, data-centre expansion, scientific computing, and advanced semiconductor development. High-performance Computing represents approximately 47% of global application demand and aligns closely with North America's rapidly expanding AI infrastructure.

Memory requirements are intensifying as next-generation accelerator platforms transition toward HBM4. Current HBM4 products deliver more than 2.8 TB/s per stack through 2,048-bit interfaces, while 16-layer configurations can provide approximately 48GB of capacity per stack. North American cloud providers have also become major drivers of AI inference demand, particularly as AI-agent applications increase computational intensity. The region remains dependent on globally distributed manufacturing and packaging ecosystems, but U.S. semiconductor investment is encouraging additional domestic capacity. North America's estimated 28% share places it behind Asia-Pacific but substantially ahead of most other regions in high-performance memory consumption.

Europe

Europe is estimated to represent approximately 10% of HMC and HBM Market demand in 2026, supported by high-performance computing centers, scientific research, automotive computing, industrial AI, advanced engineering, telecommunications, and data-centre infrastructure. The region has comparatively limited HBM manufacturing capacity but substantial demand for accelerators and memory-intensive computing platforms. High-performance Computing represents approximately 47% of global market demand, providing Europe with an important consumption channel through supercomputing centers and research infrastructure. Large computational systems increasingly integrate thousands of processor and accelerator nodes, multiplying the amount of high-bandwidth memory required per installation.

European demand is also supported by data-centre development and sovereign computing initiatives. Networking and Data Centres account for approximately 31% of global application demand, while Graphics represents around 22%. HBM4's bandwidth above 2.8 TB/s per stack is particularly valuable for scientific simulations and AI workloads where processor utilization depends on rapid memory access. Europe's estimated 10% market share remains substantially below Asia-Pacific's approximately 58%, reflecting the geographical concentration of memory production in South Korea and broader Asian semiconductor supply chains. However, investment in regional AI infrastructure should sustain growth through 2035.

Asia-Pacific

Asia-Pacific is estimated to dominate the HMC and HBM Market with approximately 58% share in 2026, supported by the concentration of advanced DRAM production, semiconductor fabrication, packaging, electronics manufacturing, and AI hardware supply chains. Samsung and SK Hynix represent South Korea among the 5 supplied companies and form 2 of the world's principal HBM suppliers. Industry competition intensified during 2026 as Samsung completed HBM4 validation and began shipments during the second quarter, while other suppliers adjusted qualification and production schedules. HBM capacity expansion remains a major strategic priority across the region as manufacturers redirect advanced DRAM resources toward AI-related memory.

South Korea is especially important because Samsung and SK Hynix are expanding memory and advanced manufacturing capacity to address long-term AI demand. SK Hynix announced approximately 54.3 trillion won of additional investment across production facilities, illustrating the capital intensity of the current memory expansion cycle. Industry forecasts indicate DRAM and NAND demand could grow approximately 19% annually through 2030 as AI infrastructure expands. Asia-Pacific's approximately 58% share is 30 percentage points above North America's estimated 28%, reflecting the region's manufacturing concentration. The transition toward HBM4, HBM4E, 16-layer stacking, advanced bonding, and customized memory should help Asia-Pacific retain leadership through 2035.

Latin America

Latin America is estimated to account for approximately 2% of HMC and HBM Market demand in 2026, reflecting limited domestic production of advanced stacked memory and a comparatively smaller installed base of AI accelerator infrastructure. Demand is primarily associated with imported servers, cloud infrastructure, research computing, graphics workstations, telecommunications systems, and regional data centres. Networking and Data Centres represent approximately 31% of global demand, providing the strongest long-term regional opportunity as digital services and cloud workloads expand. High-bandwidth memory is generally embedded within complete accelerator systems rather than purchased independently by regional end users.

Growth should accelerate as AI inference becomes more geographically distributed and enterprises increase adoption of accelerated computing. A server platform containing 8 HBM4 stacks operating above 2.8 TB/s per stack can theoretically provide aggregate bandwidth exceeding 22 TB/s, illustrating the performance available to regional cloud and research facilities adopting next-generation hardware. Latin America's approximately 2% share remains small compared with Asia-Pacific's 58%, but the growing deployment of accelerator-based data-centre systems should gradually increase HBM consumption through 2035.

Middle East and Africa

The Middle East and Africa is estimated to account for approximately 2% of HMC and HBM Market demand in 2026, supported by emerging AI infrastructure, sovereign computing initiatives, hyperscale data-centre development, telecommunications modernization, and scientific computing. Regional governments and technology investors are increasingly interested in large AI clusters, which can require thousands of accelerators and substantial high-bandwidth memory capacity. High-performance Computing and Networking and Data Centres collectively represent approximately 78% of global application demand, closely aligning with the types of infrastructure projects driving regional adoption.

The region remains dependent on imported processors and memory because none of the 5 supplied companies is headquartered in the Middle East and Africa. However, HBM is increasingly consumed as part of integrated AI accelerator platforms, allowing regional adoption to expand without domestic memory fabrication. HBM4 configurations providing more than 2.8 TB/s per stack and future products with higher capacities can support increasingly sophisticated training and inference clusters. The region's estimated 2% share is expected to remain comparatively modest, but accelerated investment in AI computing and data-centre capacity provides a meaningful growth pathway through 2035.

List of Top HMC and HBM Companies

• Samsung (South Korea)
• SK Hynix (South Korea)
• Advanced Micro Devices (U.S.)
• Intel (U.S.)
• Micron (U.S.)

Top Two Companies Market Share

SK Hynix: SK Hynix is estimated to hold approximately 48% of the addressable HBM supplier market in 2026, maintaining a leading position through established qualification relationships, advanced HBM production, high-volume manufacturing, and strong exposure to AI accelerator demand. The company has been central to the transition from HBM3 and HBM3E toward HBM4 as AI systems require substantially higher memory throughput. High-performance Computing represents approximately 47% of total HMC and HBM demand, providing a favorable demand base for suppliers with qualified high-volume products. SK Hynix has also outlined substantial production investment, including approximately 54.3 trillion won associated with expanded manufacturing infrastructure. Such investment is significant because HBM requires both advanced DRAM capacity and sophisticated packaging resources, creating a more complex supply chain than conventional memory. The transition to 12-layer and 16-layer products increases the number of vertically integrated dies and therefore raises requirements for wafer thinning, through-silicon vias, bonding, thermal management, and known-good-die screening. With HBM representing approximately 82% of the broader HMC and HBM Market, maintaining high production yields and timely customer qualification remains central to the company's competitive position.

Samsung: Samsung is estimated to account for approximately 35% of the addressable HBM supplier market in 2026 as it increases HBM4 shipments and expands capacity for next-generation AI systems. The company has indicated that HBM sales could increase by more than 3 times during 2026 compared with 2025, demonstrating the scale of demand associated with AI accelerators and data-centre computing. Samsung's HBM4 operates around 11.7 Gbps in commercial configurations and can scale toward approximately 13 Gbps, while its HBM4E development targets speeds of up to 16 Gbps and bandwidth approaching 4.0 TB/s per stack. These performance levels are increasingly important because High-performance Computing and Networking and Data Centres collectively represent approximately 78% of market demand. Samsung is also developing advanced hybrid copper bonding for future high-layer-count HBM, targeting more than 20% lower thermal resistance compared with conventional thermal-compression bonding. Lower thermal resistance becomes increasingly valuable as stack counts rise from 12 layers toward 16 layers and beyond. Samsung's broad semiconductor manufacturing base provides opportunities to coordinate DRAM, logic base dies, packaging, and process technology as custom HBM becomes more important.

Investment Analysis

Investment in the HMC and HBM Market is accelerating around advanced DRAM fabrication, through-silicon-via processing, wafer thinning, bonding, packaging, logic base dies, testing, thermal management, and production-yield improvement. The market's projected 24.2% CAGR during 2026-2035 creates strong incentives to expand capacity, particularly because High-Bandwidth Memory (Hbm) is estimated to represent approximately 82% of 2026 demand. AI infrastructure is the dominant investment catalyst, with High-performance Computing accounting for approximately 47% of applications and Networking and Data Centres contributing around 31%. Together these categories represent approximately 78% of market demand. SK Hynix has announced approximately 54.3 trillion won in additional production-related investment, demonstrating the capital intensity associated with securing long-term memory supply. Capacity expansion requires more than additional DRAM wafers because HBM consumes packaging resources and requires multiple dies for every finished stack. Moving from an 8-layer configuration to a 12-layer configuration increases the number of DRAM dies by 50%, while moving from 12 layers to 16 layers adds another approximately 33%. Suppliers therefore need coordinated investments across fabrication and advanced packaging to prevent bottlenecks from shifting between production stages.

Investment priorities are also moving toward HBM4E and custom HBM as accelerator manufacturers seek differentiated memory configurations. HBM4 doubles interface width from 1,024 bits to 2,048 bits, while commercial products already provide more than 2.8 TB/s per stack. HBM4E development pushes bandwidth toward approximately 4.0 TB/s, increasing requirements for signal integrity, logic integration, cooling, and package design. Hybrid copper bonding is another investment area because next-generation approaches target more than 20% lower thermal resistance and can facilitate stacks containing 16 or more layers. North America is estimated to represent approximately 28% of 2026 demand and expand at around 27.6% annually, creating opportunities for semiconductor manufacturing and packaging investment closer to major AI customers. Asia-Pacific remains the manufacturing center with approximately 58% market share, meaning investment decisions increasingly balance geographic diversification against established supplier ecosystems. Companies that improve yield by only 5 percentage points can materially increase usable output from expensive advanced production lines, making process-control investment nearly as important as nominal wafer capacity.

New Product Development

New product development is centered on HBM4, HBM4E, higher layer counts, increased stack capacity, customized base dies, lower power consumption, and improved thermal performance. HBM4 represents a substantial architectural advancement because the interface width expands to 2,048 bits, approximately 2 times the 1,024-bit interface associated with earlier HBM generations. Current 36GB 12-layer products can exceed 2.8 TB/s per stack, while 48GB 16-layer configurations increase capacity by approximately 33%. This progression is directly aligned with High-performance Computing, which represents approximately 47% of estimated 2026 application demand. AI accelerator designers increasingly require enough local memory to accommodate larger models and context windows while maintaining rapid access to parameters. An accelerator integrating 8 stacks of 48GB memory could theoretically provide 384GB of directly attached HBM capacity, compared with 288GB when using 8 stacks of 36GB. Such capacity gains can reduce the need to partition memory-intensive workloads across additional processors, although practical system configurations depend on processor architecture and package design.

HBM4E and custom HBM represent the next development stage as memory suppliers integrate more logic functionality and collaborate more closely with accelerator designers. HBM4E samples reaching up to 16 Gbps per pin and approximately 4.0 TB/s per stack demonstrate how quickly bandwidth targets are advancing. Compared with a 2.8 TB/s HBM4 stack, 4.0 TB/s represents an increase of approximately 43%. Development is simultaneously addressing thermal resistance because higher bandwidth and additional layers concentrate more activity within limited package area. Advanced hybrid copper bonding targets thermal-resistance reductions exceeding 20%, supporting future 16-layer and higher configurations. Custom HBM also creates opportunities to tailor logic base dies to specific accelerators, potentially optimizing power management, interfaces, reliability features, and data movement. Networking and Data Centres represent approximately 31% of demand, providing a significant market for these developments as hyperscale facilities deploy larger AI clusters. The product roadmap through 2035 is consequently shifting from standardized memory capacity toward integrated optimization of bandwidth, capacity, energy efficiency, thermals, packaging, and processor-specific functionality.

Five Recent Developments

• June 2026: HBM4E development advanced toward approximately 16 Gbps per-pin performance and bandwidth approaching 4.0 TB/s per stack, representing roughly 43% more bandwidth than a 2.8 TB/s HBM4 implementation and targeting next-generation AI accelerators.
• April 2026: Samsung accelerated HBM4 commercialization as the industry moved toward a 2,048-bit memory interface, approximately 2 times the width of the previous 1,024-bit generation, while preparing higher-performance configurations for expanding AI infrastructure demand.
• February 2026: Micron entered high-volume production of 36GB 12-layer HBM4 with more than 2.8 TB/s bandwidth per stack and over 20% improved power efficiency compared with its preceding HBM3E generation.
• November 2025: Development emphasis shifted toward 16-layer HBM architectures capable of providing approximately 48GB per stack, representing around 33% greater capacity than 36GB 12-layer configurations and supporting larger AI models within constrained package footprints.
• September 2024: The HBM3E competitive cycle accelerated around 12-layer stacking, with configurations reaching approximately 36GB per stack and providing the technological foundation for subsequent HBM4 products targeting significantly wider interfaces and higher bandwidth.

Report Coverage

The HMC and HBM Market report covers industry conditions across the 2025-2035 assessment period, with detailed evaluation of technology evolution, product segmentation, application demand, regional positioning, competitive activity, investment priorities, manufacturing constraints, and next-generation memory development. The market is projected to expand at a CAGR of 24.2% during 2026-2035, reflecting strong demand for memory architectures capable of supporting increasingly data-intensive computing. Product coverage is restricted to the 3 supplied categories of Hybrid Memory Cube (Hmc), High-Bandwidth Memory (Hbm), and Other. High-Bandwidth Memory (Hbm) is estimated to represent approximately 82% of 2026 demand, compared with around 11% for Hybrid Memory Cube (Hmc) and approximately 7% for Other. The assessment examines critical technical parameters including memory bandwidth, interface width, layer count, stack capacity, through-silicon vias, logic base dies, advanced bonding, thermal resistance, energy efficiency, package integration, and manufacturing yield. HBM4 represents a major generational transition through its 2,048-bit interface, which is approximately 2 times the width of the 1,024-bit architecture associated with the preceding generation. Current HBM4 configurations can exceed 2.8 TB/s of bandwidth per stack, while HBM4E development is targeting approximately 4.0 TB/s, representing an increase of around 43% and demonstrating the rapid performance progression occurring across advanced stacked memory.

Application coverage is limited to the 3 supplied categories of Graphics, High-performance Computing, and Networking and Data Centres, estimated to represent approximately 22%, 47%, and 31% of 2026 market demand, respectively. High-performance Computing and Networking and Data Centres therefore account for a combined approximately 78%, illustrating the strong concentration of demand around AI accelerators, scientific computing, cloud infrastructure, and data-intensive processing. Regional analysis covers North America, Europe, Asia-Pacific, Latin America, and the Middle East and Africa, with Asia-Pacific estimated to account for approximately 58% of 2026 demand and North America around 28%. Competitive coverage includes all 5 supplied companies: Samsung, SK Hynix, Advanced Micro Devices, Intel, and Micron. South Korea contributes 2 companies and the United States contributes 3, meaning approximately 60% of the supplied competitive group is U.S.-based by company count. The report additionally evaluates the transition from 12-layer toward 16-layer architectures, where increasing the stack from 12 to 16 DRAM layers represents approximately 33% more layers. It also assesses higher-capacity 48GB configurations, custom HBM, hybrid copper bonding, AI-related memory requirements, production capacity expansion, processor-memory co-design, packaging complexity, and thermal optimization as important factors influencing the HMC and HBM Market through 2035.