blog about AI Boom Drives Ethernet Dominance in Data Center Switch Market

The explosive growth of artificial intelligence (AI) is reshaping the global technology landscape with unprecedented force. Data centers, as the core infrastructure for AI computing power, are undergoing profound changes in their network architecture. Once a quiet sector, the data center switch market has become a strategic battleground for tech giants. A recent forecast from Dell'Oro Group paints a striking picture: by 2030, spending on data center switches specifically designed for AI backend networks will surpass $100 billion.

This staggering figure not only reflects the soaring demand for AI computing power but also signals an urgent need for high-performance, high-bandwidth, low-latency network infrastructure. This article examines the market transformation driven by AI, explores the underlying factors and emerging technological trends, and analyzes Ethernet's pivotal role in this technological revolution.

AI's rapid development across applications—from natural language processing to computer vision, autonomous driving, and scientific research—has directly fueled demand for powerful computing capabilities. Training large AI models, particularly deep learning models, requires massive datasets and enormous computational resources. GPUs, as the core hardware for AI computing, present unprecedented challenges to data center networks as their performance improves and their numbers increase. Traditional network architectures struggle to meet AI workloads' stringent requirements for data transfer speed, bandwidth, and latency.

Dell'Oro Group's report clearly identifies this trend, projecting that the market for AI backend network switches will reach the $100 billion threshold by 2030. This forecast is grounded in the sustained growth of AI computing needs. As AI models grow larger—with parameters numbering in the tens or hundreds of billions, even trillions—training processes require parallel processing of vast datasets and frequent synchronization of model parameters. High-speed GPU interconnects and efficient data exchange between GPUs and storage demand exceptional network bandwidth and latency. Without adequate network performance, even the most powerful computing resources remain underutilized.

AI workloads differ significantly from traditional data center applications:

• High-density parallel computing: AI training typically involves large-scale parallel computing, with thousands of GPUs working in concert. This requires extremely high GPU-to-GPU communication bandwidth and minimal latency to ensure computational efficiency.
• Data-intensive: Training AI models involves processing enormous datasets, making data loading, transfer, and storage critical. Networks must rapidly and efficiently deliver data to compute nodes.
• Complex communication patterns: AI workloads involve multicast, broadcast, and other collective communication operations, requiring networks to support efficient group communication.
• Dynamic and elastic: AI training tasks' resource needs may fluctuate, requiring networks to adapt dynamically.

These characteristics present serious challenges for data center networks:

• Bandwidth bottlenecks: Traditional network bandwidth may be insufficient for rapid transfer of massive data between GPUs, leading to idle computing resources.
• Latency issues: High latency can significantly reduce AI model training efficiency and even affect model convergence.
• Scalability limits: As AI models grow, more compute nodes are needed, making network scalability crucial.
• Power consumption: High-performance network equipment often consumes substantial energy, making power efficiency a key consideration alongside raw performance.

Dell'Oro Group's projection is not alone; multiple market research firms share optimism about AI-driven data center network growth. A $100 billion market represents enormous commercial potential, attracting attention from network equipment suppliers, chip manufacturers, and technical service providers. This growth extends beyond hardware sales, driving innovation across the entire ecosystem.

To meet growing AI computing demands, data center network architectures continue to evolve. Dell'Oro Group's report identifies several key deployment models driving backend network spending and enabling new architectures.

Scale-up (vertical scaling): This approach increases resources within existing systems to boost performance. For AI, scale-up typically involves integrating more GPUs and memory into individual servers or compute nodes, using high-speed interconnects like NVLink to minimize GPU communication latency. This enables tighter coupling, improving single-node computing density and performance.

Scale-out (horizontal scaling): This strategy adds more servers and compute nodes, connecting them via networks to form clusters. Scale-out underpins large-scale AI training clusters, enabling parallel processing of complex tasks and massive datasets. However, as clusters grow, inter-node communication increases exponentially, demanding greater network bandwidth and lower latency.

Dell'Oro Group emphasizes that both scale-up and scale-out strategies significantly increase backend network spending. Scale-up requires higher-density, more advanced network interfaces and switches to support intra-node GPU connections. Scale-out demands more powerful network backbones and more efficient switches to handle inter-node communication at scale.

More notably, a new architecture called "scale-across" is gaining traction. This model connects geographically dispersed but logically unified data centers to form a single computing cluster. Unlike traditional centralized hyperscale data centers, scale-across distributes computing resources across regions—closer to data sources or users—while integrating them into an efficient whole through advanced networking.

Scale-across addresses challenges in deployment, energy efficiency, and cost faced by traditional hyperscale data centers. Its core benefits include:

• Optimized energy use: Distributing compute resources near energy supplies or cooling sources improves energy efficiency and lowers power usage effectiveness (PUE).
• Maximized performance per watt: Distributed deployment better matches computing needs with energy availability, improving performance per unit of energy consumed.
• Reduced latency: Locating compute resources near data sources or users minimizes data transfer and processing delays, critical for real-time AI applications.
• Enhanced resilience: Distributed systems naturally improve fault tolerance, as individual node failures don't disrupt overall operations while providing better regional disaster recovery.
• Overcoming siting constraints: Finding land, power, and cooling for large data centers can be difficult; scale-across avoids these limitations.

Projects like AWS's Project Rainer and Microsoft's Fairwater exemplify scale-across principles. These initiatives aim to build flexible, efficient distributed AI computing infrastructure, connecting global data centers via advanced networking to create unified, scalable AI platforms.

Implementing scale-across imposes higher network demands. It requires not only high-bandwidth, low-latency internal connections but also efficient, reliable cross-regional links. This may involve advanced routing technologies, more robust network protocols, and smarter network management systems.

In this AI-driven network transformation, switch vendors stand to benefit substantially. At the technology level, Ethernet—with its robust ecosystem—emerges as the likely long-term winner.

As an open, mature, and widely adopted technology, Ethernet offers unparalleled advantages in data center networking:

• Open standards: Ethernet's open standards foster a vast ecosystem with broad vendor support, reducing costs and improving interoperability.
• Cost-effectiveness: Compared to proprietary technologies, Ethernet delivers better economics, especially at scale.
• Maturity and reliability: Decades of development have proven Ethernet's reliability and performance.
• Continuous innovation: Ethernet standards evolve steadily, with speeds advancing from 100GbE to 400GbE, 800GbE, and 1.6TbE to meet rising demands.

Dell'Oro Group notes that while multiple technologies will coexist, Ethernet's rise in high-performance backend networks is unstoppable. By 2025, Ethernet had already become the dominant interconnect standard in supercomputing. Innovations like HPE's Slingshot Ethernet solutions have been widely adopted in top-tier high-performance systems, demonstrating Ethernet's potential in this domain.

Dell'Oro Group's previous forecast paints an even brighter picture: over the next five years, Ethernet will contribute nearly $80 billion to data center switch sales. This suggests Ethernet will capture the majority of the market, particularly in AI-driven high-performance backend networks.

Ethernet faces competition, however. In scale-up architectures, GPU and memory tight coupling often relies on proprietary fabrics like NVLink. Dell'Oro Group VP Sameh Boujelbene explains: "To meet explosive computing demands, we can no longer rely solely on scale-out networks for GPU-to-GPU connections across racks. This shift is driving scale-up architectures that tightly couple GPUs and memory in shared high-bandwidth environments for distributed inference."

In scale-up, NVLink and similar proprietary fabrics have long dominated, offering exceptional bandwidth and minimal latency for intra-node GPU communication. However, Boujelbene observes: "Just as Ethernet surpassed InfiniBand in large scale-out environments, we now see alternatives like UALink and Ethernet gaining momentum in scale-up architectures."

UALink, a new interconnect technology, aims to outperform NVLink with higher speeds and lower latency, potentially gaining strong adoption in scale-up. Yet Boujelbene concludes: "While we predict strong UALink adoption, we expect Ethernet to emerge as the long-term winner in both scale-up and scale-out." This suggests that despite UALink's near-term potential, Ethernet's openness, cost advantages, and continuous innovation will ultimately prevail.

To further enhance network performance for AI workloads, co-packaged optics (CPO) technology is gaining attention. CPO integrates optical modules with switch chips in a single package, shortening optical signal paths to reduce power consumption and signal loss while enabling higher bandwidth, lower latency, and improved energy efficiency.

Even Nvidia, long associated with InfiniBand, is embracing Ethernet and CPO. Its Ethernet-based Spectrum-X line has evolved from initial configurations to become "roughly comparable" to competitors' offerings. Dell'Oro Group highlights Nvidia's leadership in CPO adoption for silicon photonic switches.

Nvidia's CPO versions of Spectrum-X and Quantum-X (InfiniBand-based) are expected this year. Nvidia engineers report these CPO solutions improve power efficiency 3.5-fold, signal integrity 63-fold, and provide 10x greater network resilience at scale. These metrics indicate CPO's transformative potential for AI networks.

Broadcom is also advancing CPO technology. Since 2021, it has developed CPO switches supporting 800Gb/s and 1.6Tb/s speeds. Late last year, Broadcom disclosed Meta's testing of its photonic products, which achieved zero link failures across millions of hours of 400G-equivalent port operation—demonstrating remarkable reliability.

Nevertheless, Broadcom CEO Hock Tan remains cautious, telling investors: "I can foresee a future where silicon photonics becomes the only solution, but we're not quite there yet—though we have the technology and continue developing it." This tempered outlook reflects CPO's remaining challenges in cost, yield, and mass production, despite its widely recognized potential.

The AI-driven transformation of data center networks heralds not only vast market opportunities but also a new chapter in technological evolution. Whether Ethernet can dominate as predicted—leveraging its openness and cost advantages to meet growing complexity—will be a key focus in coming years.

Ethernet's openness and standardization remain its core strengths, enabling broad ecosystem participation that drives innovation and cost reduction. As AI workloads proliferate, demand for high-performance, high-bandwidth networks will grow, with Ethernet's evolving standards and economics positioning it favorably.

AI workloads' complexity presents new challenges, however. For example, in scale-up architectures, matching proprietary fabrics' performance via Ethernet will require focused R&D. CPO integration also demands better optical component compatibility for efficient signal transmission.

Future data center networks will likely feature diverse solutions. Ethernet will dominate, but in specialized areas like high-performance scale-up computing, emerging technologies like UALink may find niches. Meanwhile, InfiniBand will retain relevance in certain high-performance computing scenarios.

• Higher-speed Ethernet: 800GbE, 1.6TbE, and beyond will accelerate to meet AI training and inference needs.
• CPO adoption: As CPO matures and costs decline, it will gradually enter mainstream data center switches, boosting performance and efficiency.
• Network intelligence: AI will enhance network management and optimization, improving operational efficiency and reliability.
• Software-defined networking (SDN): SDN will provide more flexible control and management to accommodate AI workloads' dynamic nature.

AI's explosive growth is profoundly reshaping the data center switch market, with a $100 billion opportunity signaling unprecedented potential. Dell'Oro Group's forecast outlines a clear future: an AI-driven network revolution is underway, with Ethernet—through its openness, cost-effectiveness, and continuous innovation—poised as the long-term winner. Revolutionary technologies like CPO inject new momentum into AI network performance. In this AI-led transformation, embracing change, strategic planning, and sustained innovation will be essential for competitive success. We stand at the threshold of an exciting new era, witnessing and participating in the remarkable evolution of data center networking.