AI 워크로드용 CXL 메모리 확장 : 시장 점유율 분석, 업계 동향 및 통계, 성장 예측(2026-2031년)

According to Mordor Intelligence, the CXL memory expansion for AI workloads market size is projected to be USD 0.34 billion in 2025, USD 0.74 billion in 2026, and reach USD 2.79 billion by 2031, growing at a CAGR of 30.16% from 2026 to 2031.

This report is Segmented by Component (Direct-Attached CXL Type-3 Memory Expansion Devices, and More), Physical Form Factor (EDSFF / E3. S CXL Memory Modules, and More), Application (AI Training and Model Development, and More), End User (Hyperscalers, and More), CXL Specification (CXL 1. 1 and Earlier, and More), and Geography. The Market Forecasts are Provided in Terms of Value (USD).

Global CXL Memory Expansion For AI Workloads Market Trends and Insights

AI Model Parameter Growth Outpacing HBM and DIMM Capacity

The CXL memory expansion for AI workloads market is being accelerated by a problem that lies above raw accelerator throughput, because many AI workloads run into memory limits before compute limits. A 70-billion-parameter model trained in BF16 precision requires close to 140 GB of memory for weights alone, and long-context inference pushes practical memory needs much higher once KV cache growth is added. The same research path also shows that 1-million-token contexts can drive KV cache needs to nearly 330 GB for a 70-billion-parameter model, which is beyond the HBM capacity available on any current GPU. This keeps the CXL memory expansion for AI workloads market closely tied to inference economics, because adding coherent external memory can extend usable model contexts without redesigning the accelerator package or the host socket. Penguin Solutions stated in 2026 that inference workloads are heavily shaped by memory pressure rather than pure compute pressure, which supports the view that idle GPU time has become a direct cost problem in production clusters. As a result, the CXL memory expansion for the AI workloads market is gaining support not only as a hardware category, but also as a way to recover utilization from expensive deployed accelerators.

Hyperscaler Shift Toward Memory Disaggregation to Improve Utilization

The CXL memory expansion for AI workloads market is also advancing because hyperscalers are no longer treating memory disaggregation as a lab exercise. Microsoft Research published its Octopus architecture to demonstrate that sparse CXL pod topologies can connect large memory domains without requiring an expensive full-mesh switching model, thereby lowering the cost barrier to scaled pooling. That matters because the commercial path for the CXL memory expansion for AI workloads market depends on whether memory pooling can grow without forcing proportional switch cost at every expansion step. Astera Labs announced in November 2025 that its Leo controllers were enabled on Microsoft Azure M-series virtual machines, which marked the first publicly announced commercial cloud deployment of CXL-attached memory. This public deployment created a usable reference point for enterprise buyers and smaller cloud operators that had been waiting for evidence beyond internal hyperscaler projects. The commercial implication is that vendors in the CXL memory expansion for AI workloads market now need to support both merchant deployments and custom cloud-architecture paths simultaneously.

Limited Production-Qualified CXL Type 3 Supply Base

The CXL memory expansion for AI workloads market still faces a supply ceiling because the pool of production-qualified Type-3 memory vendors remains narrow. The qualified supplier base is centered on Samsung Electronics, SK hynix, and Micron Technology, yet the range of tested densities, form factors, and specification combinations is still much smaller than data center buyers want for broad deployment. SK hynix confirmed customer validation of its 96 GB CMM-DDR5 CXL 2.0 module in 2025 and noted a 128 GB variant in parallel validation, which shows progress but not full breadth across deployment options. This matters because the CXL memory expansion for the AI workloads market competes for the same underlying memory manufacturing capacity that also serves standard DDR5 and HBM demand. When HBM volumes rise, internal allocation can shift away from emerging CXL products even if customer interest remains strong. The result is that the CXL memory expansion for AI workloads market can see demand faster than shippable volume, which slows real deployment even when the technology case is already accepted.

Other drivers and restraints analyzed in the detailed report include:

1. CXL 2.0 and CXL 3.x Ecosystem Maturity
2. PCIe Gen 5 and Gen 6 Platform Rollout
3. Platform Interoperability and Validation Complexity

For complete list of drivers and restraints, kindly check the Table Of Contents.

Segment Analysis

Direct-attached CXL Type-3 memory expansion devices held a 47.32% share of the CXL memory expansion for AI workloads market in 2025, indicating that early demand favored simple single-host deployments over more complex pooled designs. This position reflected near-term buying behavior because initial inference deployments required additional memory capacity without forcing a full change in rack architecture. In the CXL memory expansion for AI workloads industry, direct-attached devices were easier to validate on Intel Xeon 6 and AMD EPYC Turin platforms, which reduced deployment friction for first-wave buyers. Mid-layer products, such as memory pooling appliances and enclosures, served customers who needed more flexibility than single-host cards could offer, but were not ready for full fabric implementations. Memory management and orchestration software held a smaller share, yet it remained central to commercial usefulness because pooled memory has limited value unless it can be assigned dynamically across changing workloads. MemVerge positions this layer around transparent memory tiering and GPU cluster efficiency, which helps explain why the software stack is strategically important even when it is not yet the largest revenue contributor.

CXL memory fabric and rack-scale systems are projected to grow at a 30.96% CAGR through 2031, which makes them the fastest-expanding component segment in the CXL memory expansion for AI workloads market. Their momentum comes from the ability to let multiple compute hosts draw from a shared memory pool, which changes how GPU-to-memory ratios can be set at the rack level. XConn Technologies and MemVerge demonstrated a 100 TB commercial CXL pool in October 2025 and reported more than 5x performance improvement over SSD-based KV cache offload for AI inference workloads. Marvell then announced in March 2026 that the Structera S 30260 switch would support up to 48 TB of shared memory across 16 to 32 hosts at 4 TB/s aggregate bandwidth, providing a clearer commercial roadmap for rack-scale pooling. The shift from device-level expansion to shared-fabric architectures matters because it allows capacity to more closely follow workload demand across the cluster. That is why the CXL memory expansion for the AI workloads market is likely to see value move gradually from simple attachment products toward higher-level pooling systems over the forecast period.

EDSFF/E3.S CXL memory modules held 49.84% share in 2025, which made them the leading physical form factor within the CXL memory expansion for AI workloads market. Their lead came from compatibility with standard server NVMe bays, a lower thermal footprint, and closer alignment with mainstream server integration paths. Intel demonstrated the use of Xeon 6 6900P processors with multiple CXL E3.S modules for AI and HPC workloads, which reinforced this form factor as a practical fit for OEM server configurations. In the CXL memory expansion for AI workloads industry, this gave EDSFF/E3.S modules an advantage in enterprise and OEM qualification pipelines where mechanical fit and platform familiarity matter. Proprietary integrated formats still have a role, especially in custom cloud systems where vendors can optimize board design and latency paths more tightly than commodity server formats allow.

PCIe add-in cards are projected to grow at a 30.92% CAGR from 2026 to 2031, which places them just behind the leading growth tiers in the overall CXL memory expansion for AI workloads market. Penguin Solutions launched its MemoryAI KV cache server in March 2026 with up to 8 x 1 TB CXL add-in cards and a total of 11 TB of CXL-based memory in a 4U chassis, which made the AIC role in inference infrastructure very clear. This points to a split in adoption paths, AICs fit purpose-built inference appliances, while EDSFF/E3.S modules fit mainstream server refresh cycles. The CXL memory expansion for AI workloads market, therefore, is not moving toward a single universal physical format. It is moving toward a dual-track model where enterprise servers and custom inference systems favor different packaging choices. That form factor split is likely to persist because thermal, serviceability, and deployment model needs differ sharply between standard data center fleets and AI-specialized racks.

Complete Report Scope:

• By Component
  • Direct-Attached CXL Type-3 Memory Expansion Devices
  • CXL Memory Pooling Appliances
  • CXL Memory Fabric and Rack-Scale Systems
  • CXL Memory Management and Orchestration Software
• By Physical Form Factor
  • EDSFF / E3.S CXL Memory Modules
  • PCIe Add-In Cards
  • Proprietary or Server-Integrated Form Factors
  • Other Form Factors
• By Application
  • AI Training and Model Development
  • AI Inference and Model Serving
  • AI Data Preparation, Vector Databases, and RAG
  • AI-Enabled HPC and Scientific Computing
  • Large-Scale In-Memory Databases and Analytics
• By End User
  • Hyperscalers
  • AI Cloud, GPU Cloud, and Neo-Cloud Providers
  • Tier-2 Cloud and Managed Service Providers
  • Enterprise Data Centers
  • Telecom, Network, and Edge Cloud Operators
  • Research Institutions, National Laboratories, and Academic HPC Centers
• By CXL Specification
  • CXL 1.1 and Earlier
  • CXL 2.0
  • CXL 3.x
  • CXL 4.0
• By Geography
  • North America
    • United States
    • Canada
    • Mexico
  • Europe
    • Germany
    • United Kingdom
    • France
    • Italy
    • Rest of Europe
  • Asia-Pacific
    • China
    • Japan
    • South Korea
    • India
    • Southeast Asia
    • Rest of Asia-Pacific
  • South America
  • Middle East and Africa

Geography Analysis

North America held a 61.44% share in 2025 and remained the largest regional contributor to the CXL memory expansion for AI workloads market, as the region hosts hyperscale cloud operators, advanced validation capacity, and early production deployments. Microsoft Azure's November 2025 enablement of Astera Labs' Leo controllers on M-series virtual machines marked the region's first publicly announced commercial cloud deployment of CXL-attached memory. The United States leads regional demand because major cloud operators and system partners can justify both bespoke designs and long validation cycles. Intel's 2026 Xeon 6+ platform rollout also reinforced North America's position by increasing live host support for CXL-based deployments across AI and scale-out infrastructure.

Asia-Pacific is projected to grow at a 31.08% CAGR through 2031, which makes it the fastest-growing regional block in the CXL memory expansion for AI workloads market. South Korea sits at the center of regional supply because SK hynix and Samsung Electronics are central to the availability of qualified memory, and that supply role increasingly overlaps with local demand growth. SK hynix completed customer validation of its 96 GB CMM-DDR5 CXL 2.0 module in 2025, which supports the region's efforts to move qualified memory into deployable configurations. Japan and India support regional momentum through sovereign AI infrastructure programs and broader data center expansion priorities described in the input. China adds a different layer to the regional picture because domestic controller activity is forming alongside demand for local AI infrastructure. Montage Technology's CXL 3.1 controller entered customer sampling in 2025, which shows that Asia-Pacific is not only a memory manufacturing base but also a growing source of controller innovation.

Europe remains an established but more cautious part of the CXL memory expansion for AI workloads market, with early use cases centered on enterprise data centers running large in-memory database environments before wider AI inference adoption. This creates a practical entry point for CXL memory expansion, as buyers can first justify adoption by addressing database and analytics workloads that already strain standard DRAM limits. The region also has a structural interest in improving memory utilization because energy-efficiency rules and sustainability reporting favor architectures that reduce overprovisioned dedicated memory. The United Kingdom, Germany, and France remain the most important country markets in Europe, while South America, the Middle East, and Africa are still earlier in adoption and are likely to follow the pace of broader AI data center build-out. The overall geographic picture shows that CXL memory expansion for AI workloads market is scaling first, where compute density, validation capacity, and memory supply access are already in place.

1. Samsung Electronics Co., Ltd.
2. SK hynix Inc.
3. Micron Technology, Inc.
4. Intel Corporation
5. Advanced Micro Devices, Inc.
6. NVIDIA Corporation
7. Marvell Technology, Inc.
8. Astera Labs, Inc.
9. Rambus Inc.
10. Broadcom Inc.
11. Microchip Technology Incorporated
12. Renesas Electronics Corporation
13. SMART Modular Technologies, Inc.
14. Montage Technology Co., Ltd.
15. MemVerge, Inc.
16. GigaIO, Inc.
17. Liqid, Inc.
18. Dell Technologies Inc.
19. Hewlett Packard Enterprise Company
20. Super Micro Computer, Inc.

Additional Benefits:

• The market estimate (ME) sheet in Excel format
• 3 months of analyst support

TABLE OF CONTENTS

1 INTRODUCTION

• 1.1 Study Assumptions and Market Definition
• 1.2 Scope of the Study

2 RESEARCH METHODOLOGY

3 EXECUTIVE SUMMARY

4 MARKET LANDSCAPE

• 4.1 Market Overview
• 4.2 Market Drivers
  • 4.2.1 AI Model Parameter Growth Outpacing HBM and DIMM Capacity
  • 4.2.2 Hyperscaler Shift Toward Memory Disaggregation to Improve Utilization
  • 4.2.3 CXL 2.0 and CXL 3.x Ecosystem Maturity
  • 4.2.4 PCIe Gen 5 and Gen 6 Platform Rollout
  • 4.2.5 Multi-Tenant AI Infrastructure Needs for Dynamic Memory Allocation
  • 4.2.6 AI Inference Cost Pressure Driving Lower TCO Memory Pooling
• 4.3 Market Restraints
  • 4.3.1 Limited Production-Qualified CXL Type 3 Supply Base
  • 4.3.2 Platform Interoperability and Validation Complexity
  • 4.3.3 Software Stack Immaturity for Memory Orchestration
  • 4.3.4 High Initial Platform Re-Qualification Cost for AI Data Centers
• 4.4 Supply Chain Analysis
• 4.5 Technological Outlook
• 4.6 Regulatory Landscape
• 4.7 Porter's Five Forces Analysis
  • 4.7.1 Bargaining Power of Suppliers
  • 4.7.2 Bargaining Power of Buyers
  • 4.7.3 Threat of New Entrants
  • 4.7.4 Threat of Substitutes
  • 4.7.5 Intensity of Competitive Rivalry

5 MARKET SIZE AND GROWTH FORECASTS (VALUE)

• 5.1 By Component
  • 5.1.1 Direct-Attached CXL Type-3 Memory Expansion Devices
  • 5.1.2 CXL Memory Pooling Appliances
  • 5.1.3 CXL Memory Fabric and Rack-Scale Systems
  • 5.1.4 CXL Memory Management and Orchestration Software
• 5.2 By Physical Form Factor
  • 5.2.1 EDSFF / E3.S CXL Memory Modules
  • 5.2.2 PCIe Add-In Cards
  • 5.2.3 Proprietary or Server-Integrated Form Factors
  • 5.2.4 Other Form Factors
• 5.3 By Application
  • 5.3.1 AI Training and Model Development
  • 5.3.2 AI Inference and Model Serving
  • 5.3.3 AI Data Preparation, Vector Databases, and RAG
  • 5.3.4 AI-Enabled HPC and Scientific Computing
  • 5.3.5 Large-Scale In-Memory Databases and Analytics
• 5.4 By End User
  • 5.4.1 Hyperscalers
  • 5.4.2 AI Cloud, GPU Cloud, and Neo-Cloud Providers
  • 5.4.3 Tier-2 Cloud and Managed Service Providers
  • 5.4.4 Enterprise Data Centers
  • 5.4.5 Telecom, Network, and Edge Cloud Operators
  • 5.4.6 Research Institutions, National Laboratories, and Academic HPC Centers
• 5.5 By CXL Specification
  • 5.5.1 CXL 1.1 and Earlier
  • 5.5.2 CXL 2.0
  • 5.5.3 CXL 3.x
  • 5.5.4 CXL 4.0
• 5.6 By Geography
  • 5.6.1 North America
    • 5.6.1.1 United States
    • 5.6.1.2 Canada
    • 5.6.1.3 Mexico
  • 5.6.2 Europe
    • 5.6.2.1 Germany
    • 5.6.2.2 United Kingdom
    • 5.6.2.3 France
    • 5.6.2.4 Italy
    • 5.6.2.5 Rest of Europe
  • 5.6.3 Asia-Pacific
    • 5.6.3.1 China
    • 5.6.3.2 Japan
    • 5.6.3.3 South Korea
    • 5.6.3.4 India
    • 5.6.3.5 Southeast Asia
    • 5.6.3.6 Rest of Asia-Pacific
  • 5.6.4 South America
  • 5.6.5 Middle East and Africa

6 COMPETITIVE LANDSCAPE

• 6.1 Market Concentration
• 6.2 Strategic Moves
• 6.3 Vendor Positioning Analysis
• 6.4 Company Profiles (includes Global Level Overview, Market Level Overview, Core Segments, Financials as available, Strategic Information, Market Rank/Share, Products and Services, Recent Developments)
  • 6.4.1 Samsung Electronics Co., Ltd.
  • 6.4.2 SK hynix Inc.
  • 6.4.3 Micron Technology, Inc.
  • 6.4.4 Intel Corporation
  • 6.4.5 Advanced Micro Devices, Inc.
  • 6.4.6 NVIDIA Corporation
  • 6.4.7 Marvell Technology, Inc.
  • 6.4.8 Astera Labs, Inc.
  • 6.4.9 Rambus Inc.
  • 6.4.10 Broadcom Inc.
  • 6.4.11 Microchip Technology Incorporated
  • 6.4.12 Renesas Electronics Corporation
  • 6.4.13 SMART Modular Technologies, Inc.
  • 6.4.14 Montage Technology Co., Ltd.
  • 6.4.15 MemVerge, Inc.
  • 6.4.16 GigaIO, Inc.
  • 6.4.17 Liqid, Inc.
  • 6.4.18 Dell Technologies Inc.
  • 6.4.19 Hewlett Packard Enterprise Company
  • 6.4.20 Super Micro Computer, Inc.

7 MARKET OPPORTUNITIES AND FUTURE OUTLOOK

• 7.1 White-Space and Unmet-Need Assessment