首頁 > 市場調查報告書 > 通訊

數據中心

市場調查報告書

商品編碼

2148213

AI資料中心DPU和智慧網卡的市場規模、佔有率和成長預測（2026-2032年）

AI Data Center DPU and SmartNIC Market Size, Share & Growth Forecast (2026-2032)

出版日期: 2026年09月24日 | 出版商:

Knowledge Sourcing Intelligence | 英文 148 Pages | 商品交期: 最快1-2個工作天內

價格

簡介目錄

預計到 2026 年，人工智慧資料中心的 DPU 和智慧網卡市場規模將達到 19.5 億美元，到 2032 年將達到 83.6 億美元。這意味著從 2026 年到 2032 年的複合年成長率將達到 27.5%。

DPU 和 SmartNIC 位於主機伺服器與網路、儲存和安全基礎架構之間。它們的作用是執行基礎架構服務，這些服務原本會佔用中央處理器 (CPU) 核心或依賴主機作業系統中的軟體傳輸。常見功能包括虛擬交換、疊加網路、封包轉向、加密、防火牆、遙測、儲存協定處理和遠端直接記憶體存取 (RDMA)。由於這些功能獨立於 CPU 運行，雲端供應商可以為 AI 資料準備、編配和推理服務預留更多主機運算資源，同時增強租用戶應用程式與基礎設施控制之間的隔離。 AI 工作負載進一步提升了這套卸載層的價值。在訓練叢集，即使加速器之間的流量使用不同的橫向擴展或縱向擴展架構，也需要高效的前端網路來進行資料攝取、編配和儲存。基於代理的推理提高了快速檢索持久性上下文、遠端儲存和鍵值快取資料的重要性。市場上既有通用部署方案，也有垂直整合部署方案。 NVIDIA 和 AMD 為伺服器和雲端生態系銷售可程式設計DPU 平台。 Marvell 和 Broadcom 提供通用智慧網路卡 (SmartNIC) 以及用於基礎設施處理的晶片和卡片。微軟 Azure Boost、Google的客製化基礎設施處理單元架構以及亞馬遜雲端服務 (AWS) 的 Nitro 系統展示了一種平行模式，即超大規模超大規模資料中心業者設計自己的卸載晶片以滿足其內部雲端需求。北美地區引領著這一趨勢，這主要得益於該地區集中了大量的超大規模資料中心業者、新型雲端平台和人工智慧平台。

市場促進因素

卸載基礎架構可以釋放 CPU 處理能力，用於 AI 工作負載。雲端網路、安全性和儲存功能在軟體運作時會消耗主機大量的 CPU 資源。 DPU 將這些進程與主機解耦，提供一個專用於封包處理、加密、儲存和遙測的可程式引擎。據 AMD 稱，在某些配置下，Pensando Salina 可以為每台伺服器分配多達 22 個 CPU 核心用於 AI 工作負載，而 Microsoft Azure Boost 則將網路和儲存虛擬化轉移到專用硬體。隨著 AI 伺服器成本的上升，釋放主機 CPU 處理能力和降低抖動所帶來的經濟價值，以及加速器運轉率的提高，都在不斷成長。
AI原生儲存正在打造DPU加速的新層面。長上下文和基於代理的推理工作負載需要比傳統推理大規模的鍵值快取和持久記憶體。 NVIDIA的BlueField-4儲存處理架構利用專用儲存處理在叢集和儲存之間傳輸上下文數據，而Marvell則將OCTEON DPU擴展到用於AI推理的網路儲存架構中。 AMD的Salina也以類似的方式加速儲存存取並管理用於擴展GPU記憶體工作流程的NVMe容量。這使得DPU的價值超越了傳統的網路虛擬化，擴展到與每秒令牌數（TPS）和GPU利用率直接相關的儲存和記憶體基礎設施層。
在多租戶人工智慧基礎設施中，安全性和隔離性要求日益提高。新雲端、公共雲端和基於託管的人工智慧平台擴大在共用基礎架構上託管多個客戶、模型和服務。分散式處理單元 (DPU) 可以在宿主作業系統之外強制執行網路和安全策略，從而在客戶工作負載和提供者基礎設施之間建立更強大的信任邊界。 NVIDIA、AMD 和 Microsoft 都在其目前的卸載平台中強調硬體加速的安全性和隔離性。隨著高價值加速器集群在租戶之間共用，以及基於代理的系統存取海量分散式數據，這項要求變得愈發關鍵。
800G 網路提高了可程式卸載的效能上限。目前的 AI 平台正在經歷網路介面從 200G 和 400G 轉向 800G 的轉變。 BlueField-4 支援 800 Gigabit的基礎設施處理速度，而根據已公佈的效能數據，AMD Salina 提供雙 400 Gigabit乙太網路連接埠和約 800 Gigabit的軟體定義網路 (SDN)頻寬。隨著頻寬的增加，純軟體封包處理在 CPU 負載方面變得越來越昂貴，加密、遙測和儲存卸載的效能要求也隨之提高。這進一步增強了部署專用 DPU 硬體的吸引力，尤其是在 AI 前端和儲存網路不斷擴展的情況下。

市場限制因素

主要阻礙因素在於大型超大規模資料中心業者之間高度的垂直整合。微軟、Google和亞馬遜可以為其所有伺服器客製化卸載平台，而無需採購現成的分散式處理單元（DPU），這限制了獨立供應商的目標市場。此外，DPU 會增加成本、功耗和軟體複雜性，因此並非所有企業級 AI 部署都必須使用 DPU。卸載服務需要與雲端協作、虛擬網路、儲存和安全軟體整合。缺乏成熟軟體堆疊的硬體部署起來可能很困難。傳統的高效能網路卡擴大整合了固定功能的卸載能力，可以以更低的成本滿足部分需求，而主機 CPU 在封包處理和加密方面的效能也不斷提升。

促進因素
- CPU處理卸載與基礎架構效率提升
- 擴展 AI 原生儲存和 KV 快取
- 多租戶環境下的安全性和隔離性
- 800G 基礎設施網路
抑制因子
- 超大規模製造商與客製化半導體之間的競爭
- 軟體和編配。
- 額外的電力和硬體成本
- 與使用專用網路卡進行卸載的功能重疊。

第11章競爭格局

市場結構與競爭強度
可程式DPU與智慧網卡平台的定位
通用商用半導體與超大規模營運商客製架構的比較
網路、儲存和安全軟體整合
雲端、伺服器 OEM 和 AI 儲存合作夥伴關係
競爭優勢：吞吐量、卸載效率、可程式性和生態系統

第12章：公司簡介

NVIDIA
AMD
Intel
Marvell Technology
Broadcom
Microsoft
Google
Amazon Web Services
Dell Technologies
HPE
Lenovo
Supermicro
Achronix
Napatech
Silicom

第13章近期趨勢

第14章附錄

簡介目錄

The AI Data Center DPU and SmartNIC Market is estimated at USD 1.95 billion in 2026 and is projected to reach USD 8.36 billion by 2032, representing a CAGR of 27.5% during 2026-2032.

DPUs and SmartNICs sit between the host server and the network, storage and security infrastructure. Their role is to execute infrastructure services that would otherwise consume central processing unit cores or depend on software forwarding in the host operating system. Common functions include virtual switching, overlay networking, packet steering, encryption, firewalling, telemetry, storage protocol processing and remote direct memory access. Because these functions operate independently from the workload central processing unit, cloud operators can maintain stronger isolation between tenant applications and infrastructure control while preserving more host compute for AI data preparation, orchestration and inference services. AI workloads expand the value of this offload layer. Training clusters require efficient front-end networking for data ingestion, orchestration and storage even when accelerator-to-accelerator traffic uses a separate scale-out or scale-up fabric. Agentic inference increases the importance of persistent context, remote storage and fast retrieval of key-value cache data. The market contains both merchant and vertically integrated deployments. NVIDIA and AMD sell programmable DPU platforms to server and cloud ecosystems. Marvell and Broadcom provide merchant SmartNIC and infrastructure-processing silicon and cards. Microsoft Azure Boost, Google's custom infrastructure processing unit architecture and Amazon Web Services' Nitro system demonstrate a parallel model in which hyperscalers design their own offload silicon around internal cloud requirements. North America leads adoption through hyperscaler, neocloud and AI-platform concentration.

Market Drivers

Infrastructure offload preserves central processing unit capacity for AI workloads. Cloud networking, security and storage functions can consume a material share of host central processing unit resources when executed in software. DPUs separate this work from the host and provide dedicated programmable engines for packet processing, encryption, storage and telemetry. AMD states that Pensando Salina can return up to 22 central processing unit cores per server to AI work in selected configurations, while Microsoft Azure Boost moves network and storage virtualization onto dedicated hardware. As AI servers become more expensive, the economic value of preserving host central processing unit capacity and reducing jitter rises alongside accelerator utilization.
AI-native storage is creating a new DPU acceleration layer. Long-context and agentic inference workloads create much larger key-value cache and persistent-memory requirements than conventional inference. NVIDIA's BlueField-4 storage processing architecture uses dedicated storage processing to move context data between clusters and storage, while Marvell is extending OCTEON DPUs into network-storage architectures for AI inference. AMD's Salina similarly accelerates storage access and can manage NVMe capacity used to extend graphics processing unit memory workflows. This expands DPU value beyond conventional network virtualization into a storage-and-memory infrastructure layer tied directly to tokens per second and graphics processing unit utilization.
Multi-tenant AI infrastructure increases security and isolation requirements. Neocloud, public-cloud and colocation AI platforms increasingly host several customers, models and services on shared infrastructure. DPUs can enforce networking and security policy outside the host operating system, creating a stronger trust boundary between customer workloads and provider infrastructure. NVIDIA, AMD and Microsoft all emphasize hardware-accelerated security and isolation in their current offload platforms. The requirement becomes more important as high-value accelerator clusters are shared across tenants and as agentic systems access large amounts of distributed data.
800G networking raises the performance ceiling for programmable offload. Network interfaces are moving from 200G and 400G toward 800G in current AI platforms. BlueField-4 supports 800 gigabits per second infrastructure processing, while AMD Salina provides dual 400 Gigabit Ethernet ports and nearly 800 gigabits per second software-defined networking bandwidth in published performance data. Higher bandwidth makes software-only packet processing increasingly expensive in central processing unit terms and raises the performance requirement for encryption, telemetry and storage offload. This strengthens the case for purpose-built DPU hardware as AI front-end and storage networks scale.

Market Restraints

The principal restraint is the high degree of vertical integration among the largest hyperscalers. Microsoft, Google and Amazon can design custom offload platforms rather than purchase merchant DPUs in every server, limiting the addressable market for independent vendors. DPUs also add cost, power consumption and software complexity and are not necessary in every enterprise AI deployment. Offloaded services must integrate with cloud orchestration, virtual networking, storage and security software. Hardware without a mature software stack can be difficult to adopt. Conventional high-performance network interface cards increasingly incorporate fixed-function offloads that can address some requirements at lower cost, while host central processing units continue to improve packet-processing and encryption performance.

Technology and Segment Insights

Programmable DPUs with onboard Arm-class compute cores and hardware accelerators represent the largest 2026 revenue pool. NVIDIA BlueField, AMD Pensando and Marvell OCTEON combine programmable processing with high-speed Ethernet, cryptography and storage functions, allowing infrastructure services to execute independently from the host central processing unit. These platforms carry higher average selling prices than conventional network interface cards because they integrate processors, memory, accelerators and complex software stacks. NVIDIA BlueField-4 raises DPU throughput to 800 gigabits per second for next-generation AI infrastructure. AMD Salina targets front-end AI networking, security, storage and observability offload.
SmartNIC and infrastructure processing unit architectures optimized around programmable packet processing and cloud-specific offload are expected to expand rapidly through 2032. Microsoft Azure Boost and Google and Intel infrastructure processing units illustrate how large cloud providers can customize the balance between application-specific integrated circuit acceleration, embedded compute and software programmability. Storage-focused DPU configurations are expected to record the fastest use-case growth as agentic AI and distributed inference create new requirements around key-value cache, NVMe-over-Fabrics and secure data movement.
By use case, the market spans software-defined networking and virtual switching, storage and NVMe-over-Fabrics acceleration, AI storage and key-value cache data movement, security with encryption and tenant isolation, telemetry and observability, and front-end AI networking. By network interface, the market covers up to 200G, 400G, 800G and an above-800G roadmap. By deployment layer, the market covers AI compute servers, front-end network servers, AI storage nodes, cloud infrastructure services and multi-tenant graphics processing unit platforms. By customer type, demand comes from hyperscale cloud providers, neocloud and graphics-processor-cloud operators, AI storage providers, colocation and managed infrastructure providers, and enterprise and high-performance computing customers.

Competitive and Strategic Outlook

The competitive landscape combines merchant semiconductor vendors, vertically integrated hyperscalers and server-platform companies. NVIDIA has the broadest commercially integrated DPU position through BlueField hardware, DOCA software and tight linkage with Spectrum-X, storage and Vera Rubin systems. AMD competes through the Pensando portfolio and is integrating Salina into Helios and Microsoft Azure. Marvell offers OCTEON DPUs across cloud, networking and AI-storage applications, while Broadcom maintains a SmartNIC and high-performance Ethernet offload portfolio.
Hyperscalers remain strategically important competitors and customers. Microsoft uses both its own Azure Boost DPU architecture and external AMD Pensando technology. Google works with Intel on custom infrastructure processing units, and Amazon Web Services has long used its internally developed Nitro architecture to offload networking, storage and management functions. These companies influence merchant roadmap requirements even where they do not purchase standard cards in every deployment.
Server and system vendors determine how quickly DPUs become part of standardized AI infrastructure. Dell Technologies, HPE, Lenovo and Supermicro integrate or qualify BlueField, Pensando, Intel and other offload technologies across data-center platforms. Achronix, Napatech and Silicom participate in programmable acceleration and SmartNIC implementations for specialized networking and cloud environments. Competitive differentiation centers on throughput, central processing unit offload efficiency, programmable software, encryption and security, storage acceleration, power consumption, orchestration integration and ecosystem support. Programmability differentiates DPUs from fixed-function high-speed Ethernet network interface cards, and open software stacks and orchestration integration increasingly determine commercial platform adoption.
Recent developments confirm accelerating deployment. In September 2026, Marvell detailed DPU-powered network-storage architectures for AI inference, using OCTEON DPUs to offload networking, security and storage processing from host central processing units. In August 2026, Marvell showcased OCTEON DPU technology for AI-native network storage and advanced inference infrastructure. In July 2026, Microsoft and AMD announced that Azure is deploying AMD Pensando DPUs in AI backend networking infrastructure and selected Azure services. In May 2026, Microsoft made the next generation of Azure Boost generally available with custom application-specific integrated circuit hardened networking and storage offload infrastructure. In April 2026, Intel and Google announced expanded multiyear co-development of custom application-specific integrated circuit based infrastructure processing units for cloud and AI infrastructure. In March 2026, NVIDIA announced Vera Rubin production systems including BlueField-4 DPU and BlueField-4 storage infrastructure. In January 2026, NVIDIA introduced BlueField-4-powered AI-native context-memory storage infrastructure for long-context and agentic inference.
North America is the largest market for AI data-center DPUs and SmartNICs because the region contains the leading hyperscale cloud providers, AI-platform companies and merchant DPU suppliers. NVIDIA, AMD, Intel, Marvell and Broadcom are headquartered in the United States, while Microsoft, Google and Amazon Web Services operate some of the world's largest internal SmartNIC and infrastructure-processing fleets. This concentration creates both merchant demand and substantial custom-silicon activity. The 2026 product cycle strengthens that position. NVIDIA's Vera Rubin platform integrates BlueField-4 as an AI-factory infrastructure processor, while BlueField-4 storage extends DPU functionality into context-memory storage. AMD's Pensando Salina is being deployed with Microsoft Azure and positioned for AI front-end networking and storage acceleration. Intel and Google are co-developing another generation of custom infrastructure processing units. Microsoft has moved the next generation of Azure Boost into general availability with custom offload hardware for networking, storage and security.
Neocloud operators provide an additional merchant opportunity because they generally lack the scale to design custom infrastructure processors but still require cloud-grade isolation, storage performance and observability. AI storage suppliers are also adopting DPU architectures to reduce central processing unit overhead and move data efficiently between graphics processing units and flash. Through 2032, North American growth is expected to be led by BlueField and Pensando deployments, AI-native storage offload, custom hyperscale infrastructure processing units and higher attach rates in shared graphics processing unit infrastructure. Asia Pacific plays a major manufacturing role and is also expanding AI-cloud deployment across China, Japan, South Korea, Taiwan, India and Southeast Asia. Europe contributes through sovereign AI, cloud infrastructure and SmartNIC software ecosystems. The Middle East is emerging as a large AI-capacity market where merchant DPU platforms can be integrated into imported rack-scale systems and neocloud infrastructure without requiring locally designed offload silicon.

Conclusion

The AI data center DPU and SmartNIC market is positioned for rapid expansion, growing from USD 1.95 billion in 2026 to USD 8.36 billion by 2032 at a compound annual growth rate of 27.5 percent. Programmable DPUs are becoming a dedicated infrastructure-compute layer inside large AI factories, with storage-centric deployments representing the fastest-growing use case through 2032. Merchant vendors compete alongside vertically integrated hyperscaler platforms, and competitive differentiation increasingly depends on throughput, offload efficiency, programmability, orchestration integration and open software ecosystems. Security isolation becomes more valuable as AI clusters support multiple tenants and services. North America leads adoption, with Asia Pacific expanding AI-cloud deployment and Europe contributing through sovereign AI and cloud infrastructure. Vendors that combine high-bandwidth hardware with mature software stacks and broad ecosystem support are best positioned to capture value as infrastructure offload becomes a standard element of AI data-center architecture.

Key Benefits of this Report

Insightful Analysis: Detailed market insights across regions, customer segments, policies, socio-economic factors, consumer preferences, and industry verticals.
Competitive Landscape: Understand strategic moves by key players to identify optimal market entry approaches.
Market Drivers and Future Trends: Assess major growth forces and emerging developments shaping the market.
Actionable Recommendations: Support strategic decisions to unlock new revenue streams.
Caters to a Wide Audience: Suitable for startups, research institutions, consultants, SMEs, and large enterprises.

What Businesses Use Our Reports For

Industry and market insights, opportunity assessment, product demand forecasting, market entry strategy, geographical expansion, capital investment decisions, regulatory analysis, new product development, and competitive intelligence.

Report Coverage

Historical data from 2021 to 2024, Base year 2025, and Forecast years from 2026 to 2032
Growth opportunities, challenges, supply chain outlook, regulatory framework, and trend analysis
Competitive positioning, strategies, and market share evaluation, and trade analysis
Revenue growth and forecast assessment across segments and regions
Company profiling including strategies, products, financials, and key developments

1. Executive Summary

1.1. Market Opportunity and Key Findings
1.2. DPU and SmartNIC Adoption Outlook
1.3. Principal Revenue Pools

2. Market Overview

2.1. Infrastructure Offload Architecture
2.2. DPU, IPU and SmartNIC Functional Differences
2.3. AI Front-End Networking Requirements
2.4. AI Storage and KV-Cache Acceleration
2.5. Security, Isolation and Observability

3. Market Size and Forecast, 2026-2032

3.1. Global Market Revenue
3.2. Annual Growth Analysis
3.3. DPU and SmartNIC Units Shipped
3.4. Revenue per AI Server and Storage Node

4. Market by Product Architecture

4.1. Programmable DPUs
4.2. Infrastructure Processing Units
4.3. Programmable SmartNICs
4.4. Cloud-Specific Custom Offload ASICs
4.5. Software, SDK and Support Layer

5. Market by Use Case

5.1. Software-Defined Networking and Virtual Switching
5.2. Storage and NVMe-over-Fabrics Acceleration
5.3. AI Storage and KV-Cache Data Movement
5.4. Security, Encryption and Tenant Isolation
5.5. Telemetry and Observability
5.6. Front-End AI Networking

6. Market by Network Interface

6.1. Up to 200G
6.2. 400G
6.3. 800G
6.4. Above 800G Roadmap

7. Market by Deployment Layer

7.1. AI Compute Servers
7.2. Front-End Network Servers
7.3. AI Storage Nodes
7.4. Cloud Infrastructure Services
7.5. Multi-Tenant GPU Platforms

8. Market by Customer Type

8.1. Hyperscale Cloud Providers
8.2. Neocloud and GPU-Cloud Operators
8.3. AI Storage Providers
8.4. Colocation and Managed Infrastructure Providers
8.5. Enterprise and High-Performance Computing

9. Regional Market

9.1. North America
- 9.1.1. United States
- 9.1.2. Canada
9.2. Europe
9.3. Asia Pacific
9.4. Middle East and Rest of World

10. Market Dynamics

10.1. Drivers
- 10.1.1. CPU Offload and Infrastructure Efficiency
- 10.1.2. AI-Native Storage and KV-Cache Growth
- 10.1.3. Multi-Tenant Security and Isolation
- 10.1.4. 800G Infrastructure Networking
10.2. Restraints
- 10.2.1. Hyperscale Custom-Silicon Competition
- 10.2.2. Software and Orchestration Complexity
- 10.2.3. Additional Power and Hardware Cost
- 10.2.4. Overlap with Fixed-Function NIC Offloads

11. Competitive Landscape

11.1. Market Structure and Competitive Intensity
11.2. Programmable DPU and SmartNIC Platform Positioning
11.3. Merchant Silicon versus Hyperscale Custom Architecture
11.4. Networking, Storage and Security Software Integration
11.5. Cloud, Server OEM and AI-Storage Partnerships
11.6. Competitive Differentiation: Throughput, Offload Efficiency, Programmability and Ecosystem