분산형 GPU 시장 : 점유율 분석, 업계 동향 및 통계, 성장 예측(2026-2031년)

According to Mordor Intelligence, the disaggregated GPU market size is projected to be USD 3.97 billion in 2025, USD 5.34 billion in 2026, and reach USD 22.63 billion by 2031, growing at a CAGR of 33.48% from 2026 to 2031.

This report is Segmented by Component (Hardware, Software, and More), Accelerator Type (PCIe-Based Disaggregation, NVLink/NVSwitch-Based Disaggregation, and More), Deployment Mode (On-Premise, and Cloud), Application (AI and High Performance Computing, and More), End User (Hyperscale Cloud Providers, Cloud Service Providers, and More), and Geography. The Market Forecasts are Provided in Terms of Value (USD).

Global Disaggregated GPU Market Trends and Insights

Rising Demand for GPU Pooling in AI Training Clusters

The disaggregated GPU market is gaining direct support from GPU pooling, as large AI training runs now depend on treating many physical GPUs as a single shared logical resource. NVIDIA's sixth-generation NVLink delivered 1.8 TB/s of bidirectional bandwidth per GPU, and NVLink Switch systems supported 72 GPUs in a single NVL72 rack, delivering 130 TB/s of aggregate in-domain bandwidth, reducing the communication limits that used to separate adjacent racks into isolated compute islands. That shift matters for trillion-parameter models because memory and bandwidth pressure now extend across full training clusters rather than within a single node, and the disaggregated GPU market is responding with architectures that let idle capacity join active workloads much more quickly. Agentic AI is adding another layer of demand, since prefill stages are compute-heavy while decode stages are memory-bandwidth-heavy, and this pattern requires pools to be reassigned in seconds rather than through slower manual scheduling cycles. AWS demonstrated this operating model in production using llm-d and the NVIDIA Inference Xfer Library on Blackwell nodes, where disaggregated inference improved throughput by 70% at high concurrency.

Shift Toward Separated Compute and Memory Architectures

The disaggregated GPU market is also moving forward as operators separate compute from local memory and treat memory capacity as a pooled infrastructure layer. Research published in Tsinghua Science and Technology described CXL-based memory disaggregation as a major architectural step for intelligent computing centers, because shared memory pools can be reached by multiple compute nodes through standard load and store operations. ByteDance researchers extended this direction by demonstrating a 768 GB shared memory pool built from Micron CXL Type-3 cards and a TITAN-II CXL switch, enabling collective GPU communication across 3 nodes with NVIDIA H100 GPUs and no application code changes. This matters because long-context inference exceeds 1 million tokens and KV cache demand exceeds the HBM available on each accelerator, making adding pooled memory more practical than replacing whole GPU fleets. The disaggregated GPU market, therefore, gains from CXL not only as a technical option but also as an economic path for operators that need more memory capacity without forcing a full silicon refresh.

High Interconnect And Fabric Integration Complexity

The disaggregated GPU market still faces a significant integration barrier because many deployments must coordinate NVLink, PCIe, InfiniBand, Ethernet, and CXL within a single cluster. Research presented at HotNets 2025 showed that effective path management across intra-host heterogeneous interconnects requires simultaneous coordination of PCIe switches, GPU fabric controllers, RDMA NICs, and Ethernet switches, and current software stacks do not fully abstract that complexity. NVIDIA's NVLink Fusion reduced part of that problem by letting custom CPUs from MediaTek, Marvell, Qualcomm, and Fujitsu connect natively with NVIDIA GPUs, but that move expanded a proprietary path rather than creating a fully neutral multi-vendor fabric. The Ultra Accelerator Link standard is meant to offer a vendor-neutral alternative, yet the input indicates that production hardware was not expected until late 2026 and broad deployments were likely to stretch into 2027. The disaggregated GPU market also faces a regional interoperability ceiling in China, where domestic accelerator ecosystems use incompatible communication protocols, limiting the practical pooling benefits operators can achieve at scale.

Other drivers and restraints analyzed in the detailed report include:

1. Expansion of Hyperscale and Cloud-Native AI Infrastructure
2. Power And Thermal Efficiency Gains Through Resource Disaggregation
3. Software Orchestration Fragmentation Across GPU Stacks

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

Segment Analysis

Hardware accounted for 81.32% of the disaggregated GPU market in 2025, reflecting the high upfront costs of interconnect fabrics, switch trays, memory modules, and rack-scale compute systems. Each Vera Rubin NVL72 rack combined 72 Blackwell GPUs, 36 Vera CPUs, NVLink switch chips, BlueField-4 DPUs, and Spectrum-X Ethernet networking into a single fabric-based capital unit, keeping early spending in the disaggregated GPU market concentrated in large hyperscale procurement cycles. That procurement model matters because buyers are not purchasing isolated cards or servers; they are committing to tightly integrated platforms that bundle compute, networking, and management capabilities within a single deployment event. The disaggregated GPU market, therefore, shows a hardware-heavy revenue mix in its current phase, especially as hyperscalers build new AI capacity in full racks rather than through gradual node-by-node additions. Services remained the smallest component layer, yet integration, managed operations, and GPU-as-a-service delivery still carried attractive margin potential for system integrators and specialist cloud operators.

Software is projected to record the fastest 34.08% CAGR from 2026 to 2031 in the disaggregated GPU market, as orchestration layers, isolation tools, and inference frameworks increasingly determine the value operators extract from each GPU. NVIDIA Dynamo separated prefill and decode assignments to lift factory-scale utilization, and that design set a reference point that open-source alternatives such as llm-d now need to match for enterprise deployments. This changes the revenue pattern of the disaggregated GPU industry, because software upgrades can continue even when hardware refresh cycles slow or when installed fabric remains in place. Operators that installed disaggregated hardware in 2024 and 2025 are likely to add new orchestration layers before replacing physical interconnect assets, creating a path for recurring software revenue that is less tied to rack replacement timing. Over time, that dynamic should give software a larger role in the disaggregated GPU market even if hardware continues to anchor absolute spending.

NVLink/NVSwitch-based disaggregation held 44.21% of the disaggregated GPU market share in 2025, reflecting NVIDIA's strong position in AI training fabrics and the installed preference for tightly coupled GPU communication. The sixth-generation NVLink delivered 1.8 TB/s per GPU in 2024, and newer NVLink Switch systems enabled all-to-all communication across 72 GPUs with 130 TB/s of aggregate bandwidth, giving the disaggregated GPU market a high-performance option for scale-up clusters that need dense local communication. PCIe-based disaggregation maintained a durable role as the baseline fabric in many multi-GPU systems, offering broad compatibility without the same level of proprietary dependence. InfiniBand and Ethernet-based approaches continued to matter in scale-out settings, and Ethernet in particular gained relevance where operators wanted to extend large AI clusters by building on existing networking investments. That split means the disaggregated GPU market is not moving toward one universal fabric, but toward a layered model in which performance, openness, and installed infrastructure each influence the final architecture.

CXL-based disaggregation is projected to expand at a 34.46% CAGR through 2031, supported by CXL 3.0 memory pooling and the CXL 4.0 specification, which doubled bandwidth to 128 GT/s via PCIe 7.0 physical layers and introduced bundled ports for much higher total throughput. Research in Tsinghua Science and Technology showed that CXL memory disaggregation and GPU compute disaggregation serve different layers of the stack, which suggests that future deployments in the disaggregated GPU market will combine them rather than force a choice between them. Samsung Electronics is developing its Pangea CXL memory platform with Marvell and Liquid AI to expand GPU memory, where HBM limits can restrict inference batch sizes and working context depth. In practice, NVLink is well-suited to tightly coupled GPU-to-GPU communication inside the rack, while CXL extends memory capacity and memory sharing across nodes and broader system domains. That complementary relationship should help the disaggregated GPU market mature into more composable architectures without reducing the role of high-bandwidth GPU fabrics already established in production clusters.

Complete Report Scope:

• By Component
  • Hardware
  • Software
  • Services
• By Accelerator Type
  • PCIe-Based Disaggregation
  • NVLink/NVSwitch-Based Disaggregation
  • Ethernet Fabric-Based Disaggregation
  • InfiniBand Fabric-Based Disaggregation
  • CXL-Based Disaggregation
• By Deployment Model
  • On-Premise
  • Cloud-Based
• By Application
  • AI and High Performance Computing
  • Data Analytics
  • Digital Twin and Simulation
  • Rendering and Visualization
  • Scientific Research
• By End-User
  • Hyperscale Cloud Providers
  • Cloud Service Providers
  • Enterprises
  • Government and Defense Organizations
  • Research and Academic Institutions
  • Telecommunications Providers
• 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 52.71% share in 2025, giving it the largest regional share in the disaggregated GPU market and reflecting the depth of hyperscale buildouts across the US. AWS, Google Cloud, Microsoft Azure, and Oracle Cloud Infrastructure were all named as second-half 2026 deployers of NVIDIA Vera Rubin NVL72 systems, and that concentration of leading buyers continues to anchor the disaggregated GPU market in the region. North America also benefits from a dense hardware ecosystem, major AI research activity, and a mature GPU cloud layer, including providers such as CoreWeave, Lambda, and Nscale. NVIDIA's September 2025 collaboration with Intel to develop custom AI data center CPUs based on NVLink and x86 deepened regional supply chains and enabled stronger integration between orchestration CPUs and GPU fabrics. Canada adds support through proximity to US hyperscale demand and favorable power economics, while South America remains at an earlier stage and is tied more to hyperscale availability zones in Brazil and Colombia than to broad on-premise deployment.

Europe held a meaningful but smaller share of the disaggregated GPU market, with Germany, the United Kingdom, and France as the main deployment centers in user input. The European Union Energy Efficiency Directive requires power usage effectiveness reporting for data centers with an IT load above 500 kW, which supports more efficient, disaggregated designs and requires operators to demonstrate measurable facility performance. Germany's automotive and precision manufacturing base supports demand for digital twin and physics simulation workloads, where burst compute is easier to justify through pooled infrastructure than through fixed server allocations. The United Kingdom contributes through an active GPU cloud segment, while France and Italy are extending sovereign AI compute programs that incorporate disaggregated GPU capacity.

Asia-Pacific is projected to expand at a 34.39% CAGR between 2026 and 2031, giving the region the fastest growth rate in the disaggregated GPU market during the forecast period. The region is being driven by China's hyperscale AI spending, South Korea's vertically integrated memory supply chain, Japan's manufacturing and automation needs, and public AI infrastructure programs in India and Singapore. China is building domestic disaggregated architectures around proprietary interconnect approaches, creating a split regional structure in which Chinese stacks differ from those used elsewhere in the disaggregated GPU market. South Korea benefits from SK Hynix's position in HBM3e output, which helps domestic operators secure earlier access to memory subsystems and supports faster data center capital deployment. India is also moving quickly as government-backed AI initiatives and hyperscale cloud zones expand, while the Middle East and Africa remain earlier in development but are gaining support from sovereign AI investment programs in the UAE and Saudi Arabia.

1. NVIDIA Corporation
2. Advanced Micro Devices, Inc.
3. Intel Corporation
4. Qualcomm Incorporated
5. Apple Inc.
6. Samsung Electronics Co., Ltd.
7. MediaTek Inc.
8. Arm Holdings plc
9. Imagination Technologies Limited
10. Hewlett Packard Enterprise Company
11. Dell Technologies Inc.
12. Super Micro Computer, Inc.
13. Lenovo Group Limited
14. ASUSTeK Computer Inc.
15. Gigabyte Technology Co., Ltd.
16. Micro-Star International Co., Ltd.
17. Alibaba Group Holding Limited
18. Tencent Holdings Limited
19. Amazon.com, Inc.
20. Microsoft Corporation

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 Rising Demand for GPU Pooling in AI Training Clusters
  • 4.2.2 Shift Toward Separated Compute and Memory Architectures
  • 4.2.3 Expansion of Hyperscale and Cloud-Native AI Infrastructure
  • 4.2.4 Power and Thermal Efficiency Gains Through Resource Disaggregation
  • 4.2.5 Faster Fleet Utilization Through Multi-Tenant GPU Allocation
  • 4.2.6 Shorter Upgrade Cycles in Modular Data Center Architectures
• 4.3 Market Restraints
  • 4.3.1 High Interconnect and Fabric Integration Complexity
  • 4.3.2 Software Orchestration Fragmentation Across GPU Stacks
  • 4.3.3 Capital Intensity of Retrofitting Legacy Data Centers
  • 4.3.4 Latency Tradeoffs in Remote Memory and Disaggregated Access
• 4.4 Industry Value Chain Analysis
• 4.5 Regulatory Landscape
• 4.6 Technological Outlook
• 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
• 4.8 Impact of Macroeconomic Factors on the Market

5 MARKET SIZE AND GROWTH FORECASTS (VALUE)

• 5.1 By Component
  • 5.1.1 Hardware
  • 5.1.2 Software
  • 5.1.3 Services
• 5.2 By Accelerator Type
  • 5.2.1 PCIe-Based Disaggregation
  • 5.2.2 NVLink/NVSwitch-Based Disaggregation
  • 5.2.3 Ethernet Fabric-Based Disaggregation
  • 5.2.4 InfiniBand Fabric-Based Disaggregation
  • 5.2.5 CXL-Based Disaggregation
• 5.3 By Deployment Model
  • 5.3.1 On-Premise
  • 5.3.2 Cloud-Based
• 5.4 By Application
  • 5.4.1 AI and High Performance Computing
  • 5.4.2 Data Analytics
  • 5.4.3 Digital Twin and Simulation
  • 5.4.4 Rendering and Visualization
  • 5.4.5 Scientific Research
• 5.5 By End-User
  • 5.5.1 Hyperscale Cloud Providers
  • 5.5.2 Cloud Service Providers
  • 5.5.3 Enterprises
  • 5.5.4 Government and Defense Organizations
  • 5.5.5 Research and Academic Institutions
  • 5.5.6 Telecommunications Providers
• 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 Market 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)
• 6.5 NVIDIA Corporation
• 6.6 Advanced Micro Devices, Inc.
• 6.7 Intel Corporation
• 6.8 Qualcomm Incorporated
• 6.9 Apple Inc.
• 6.10 Samsung Electronics Co., Ltd.
• 6.11 MediaTek Inc.
• 6.12 Arm Holdings plc
• 6.13 Imagination Technologies Limited
• 6.14 Hewlett Packard Enterprise Company
• 6.15 Dell Technologies Inc.
• 6.16 Super Micro Computer, Inc.
• 6.17 Lenovo Group Limited
• 6.18 ASUSTeK Computer Inc.
• 6.19 Gigabyte Technology Co., Ltd.
• 6.20 Micro-Star International Co., Ltd.
• 6.21 Alibaba Group Holding Limited
• 6.22 Tencent Holdings Limited
• 6.23 Amazon.com, Inc.
• 6.24 Microsoft Corporation

7 MARKET OPPORTUNITIES AND FUTURE OUTLOOK

• 7.1 White-Space and Unmet-Need Assessment