Feature articles |
By Jean-Pierre Joosting
Network switches form the backbone of the communications infrastructure. They are found in a variety of applications, ranging from homes and entertainment centres to offices and enterprises, through to data centres, server rooms, smart buildings, and security systems. At home, simple switches are used to connect smart TVs, games consoles, and streaming devices for high-speed performance. In offices and enterprises, switches are deployed to connect desktop PCs, VoIP phones, and printers to the company’s Intranet and to aggregate wireless traffic.
In data centres and server farms, high-speed core and edge switches connect servers, storage arrays, network security appliances and other components to deliver massive data flows.
However, in data centres and enterprises, network switching has become a strategic bottleneck and an active area of innovation. Key challenges include AI workload growth regulation for sustainability, and a generational speed transition from 100 GbE (100G) to 400 GbE (400G), then to 800 GbE (800G).
According to Evolvance Market Research, the data centre switch market was valued at around 19.4 billion USD in 2026 and is expected to reach 38.1 billion USD by 2035, representing a CAGR of around 7.8%. [R1]
The race from 100G to 800G and beyond is currently the dominant trend in the market. However, even with the recent introduction of 800G, 1.6T (1.6 Tbps) is already being considered as the next step in the near future, which will deliver 200G/lane. Further standards for 3.2T are expected around 2030.
The Evolvance report states that 100G remains the mainstream standard for enterprise data centres and mid-tier cloud deployments, accounting for 41.2% revenue share of port speed in 2026. 400G is growing at 47.3% annually, driven by hyperscaler AI procurement and the migration of HPC workloads from InfiniBand to Ethernet. Just getting off the ground, the 800G tier is experiencing 81.4% annual growth, with volume hyperscaler deployment expected by 2028 and enterprise adoption by 2030–2031. The market is rapidly moving to 400G and 800G while simultaneously undergoing a structural shift to spine-leaf architectures. [R1]
Ethernet is displacing InfiniBand and consolidating its position as the dominant interconnect standard for large-scale deployments. According to Dell’Oro Group, Ethernet is rapidly replacing InfiniBand for AI back-end scale-out networks, driven by the adoption of 800 Gbps and, in the near future, 1.6 Tbps port speeds. The cumulative AI back-end switch market is expected to exceed $100 billion by 2030. Further, scale-up compute fabrics, which have been typically dominated by proprietary fabrics such as NVLink, are now seeing alternative technologies such as UALink and Ethernet gain momentum. Dell’Oro expect Ethernet to prevail here as well. [R2] [R3]
The data centre switch market in 2026 remains highly concentrated. Evolvance notes that Cisco Systems holds the commanding position at approximately 28.4% revenue share. It provides the Nexus 9000 series and Application Centric Infrastructure (ACI) software-defined networking platform for large enterprise and service provider deployments. Cisco also provide its proprietary Silicon One ASIC for high-end spine platforms. [R1]
Arista Networks, with a 22.1% share, has moved from a challenger to a co-equal at the high end of the market in a short time frame. The company enables high-capacity data centres and AI clusters with its R4 series of routers. The latest in the series, the 7800R4 family includes 576 800GbE ports in a single system. [R1][R4]
The HPE and Juniper Networks combination, following HPE’s $14 billion acquisition of Juniper in July 2025, holds approximately 11.6% market share. Huawei Technologies is the leader in the APAC region with a 9.8% global market share. [R1]
At the silicon layer, Broadcom, with its Tomahawk, Trident, and Jericho ASIC families, underpins approximately 86% of merchant switching deployments across the industry. [R1]
However, NVIDIA, with its Spectrum-X and Spectrum-4 platforms, is gaining traction in GPU clusters. Designed for AI workloads, Spectrum-4 Ethernet switches are built on the 51.2 Tbps Spectrum-4 ASIC. They combine a specialised high-performance architecture with standard Ethernet connectivity. The Spectrum-X networking platform combines the Spectrum-4 Ethernet switch with NVIDIA BlueField-3 SuperNIC. [R5]
A primary challenge for data centres and networks is power consumption. As data speeds rise, power budgets escalate significantly as bandwidth approaches 400G and 800G. Some large hyperscale switches can consume between 10 and 20 kW. The electrical-to-optical conversion in high-speed links consumes a significant amount of power and generates substantial waste heat. Power-saving techniques such as sleep states can cause latency and compromise network resilience, and consequently cannot be used. Similarly, the use of buffers can also cause latency. Furthermore, driving massive bandwidth through a small physical space creates extreme thermal bottlenecks, necessitating advanced liquid or rack-based cooling systems.
Power consumption concerns due to costs and community backlash from rising electricity bills, as well as regulatory pressure are driving investment in energy-proportional switching, dynamic power scaling, and co-packaged optics (CPO) to reduce transceiver energy losses. CPO integrates the optical transceivers directly next to the routing ASIC to reduce energy loss by shortening electrical paths. Another approach, Linear Pluggable Optics (LPO), bypasses power-hungry DSPs in pluggable modules, thereby lowering power consumption.
Adding to the environmental and power consumption challenges facing data centres and high-speed switches is a new European regulation. The European Commission will formally adopt a unified EU-wide Data Centre Energy Efficiency Package in mid-2026. The framework introduces a rating and labelling scheme applicable across all member states.
The regulation targets data centre operators and owners running facilities with an installed IT power demand of 500 kW or above. Covered facilities will be required to submit annual sustainability data to a centralised European Database on Data Centres, reporting across a defined set of key performance indicators: total power consumption, data traffic volumes, water usage, temperature set points, and the share of renewable energy in the facility’s overall consumption.
Beyond reporting, the package includes mandatory provisions for waste heat recovery, requiring operators to capture and redirect surplus heat for use in nearby building heating systems or local municipal heat networks. An automated efficiency labelling scheme will rank and publicly compare facilities with one another based on their infrastructure footprint, introducing a degree of market transparency and competitive pressure on operators to improve performance.
Several distinct segments are currently driving investment in the network switching market.
AI and high-performance computing infrastructure account for the most demand due to massive data centre investment globally. The requirement to interconnect large GPU clusters at scale is pulling hyperscaler procurement toward 400G and 800G switching and pushing custom ASIC development to meet workload-specific performance requirements. [R1]
Colocation and cloud service providers are expanding capacity rapidly, with particularly strong build-out activity in North America, Germany, and across Asia-Pacific. China accounts for the majority of the APAC data centre switch market by value. [R6]
In telecommunications, the rollout of 5G core and edge infrastructure is driving sustained demand for switching that meets the low-latency and high-throughput requirements of disaggregated radio access network architectures.
As edge computing and IoT expand, managed switches with Power over Ethernet support, cloud-based management, and IoT platform integration are the primary product requirements here.
Government and defence demand is also rising as countries spend more to address nation-state threat activity with hardware-level encryption and verified supply chain provenance increasingly treated as baseline requirements rather than optional premiums.
AI-driven telemetry and intent-based networking are transforming how switches are managed. Rather than a reactive configuration, modern platforms continuously analyse traffic patterns, predict congestion, detect anomalies, and apply policy adjustments autonomously. Arista’s EOS (Extensible Operating System) and Cisco’s ACI (Application Centric Infrastructure) SDN platform are representative examples. The broader trend is toward the network as a self-optimising system, in contrast to earlier static configuration models.
Open networking is also gaining traction. The Open Compute Project is working to standardise OCS as an open, SDN-integrated option for next-generation data centre and AI cluster fabrics through its Optical Circuit Switching (OCS) Subproject. Open ASIC ecosystems and disaggregated software/hardware models will remove the lock-in enjoyed by proprietary system vendors.
Co-packaged optics is an active area of development. CPO and in-package optical I/O (OIO) approaches promise lower power per bit and shorter signal paths. However, mass-production barriers, such as constraints on the supply of laser light sources, remain significant near-term challenges.
The Ultra Ethernet Consortium is developing the next open, high-performance Ethernet standard specifically for AI and HPC workloads, addressing the limitations of traditional RDMA (Remote Direct Memory Access). RDMA lacks multipathing, out-of-order packet delivery, and scalable congestion control. IEEE 802.3dj is progressing toward completion in 2026, standardising 200 Gb/s per lane operation. These standards activities will shape the next generation of switch silicon and fabric architectures.
[R1] https://evolvancemarketresearch.com/reports/data-center-switch-market
[R2] https://www.delloro.com/news/ethernet-is-winning-the-war-against-infiniband-in-ai-back-end-networks
[R3] https://www.delloro.com/news/ai-back-end-switch-market-will-push-past-100-billion-by-2030
[R4] https://www.arista.com/en/company/news/press-release/22541-pr-10292025
[R6] https://www.gminsights.com/industry-analysis/data-center-switch-market
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。