In the AI era, the demand for computing power networking is surging. Ethernet is rapidly overtaking InfiniBand in the AI backend interconnect arena, continuously elevating the industry value of switching chips.
In July 2026, Centec Networks (hereinafter referred to as 'Centec' or 'the Company') successively welcomed two major positive developments in orders and technology: Centec announced a substantial increase in its related-party sales quota to China Electronics Corporation (CEC) from 600 million RMB to 2.2 billion RMB, a surge of 266% (the Company's revenue in 2025 was 1.15 billion RMB, with a year-over-year growth of only 6%); at WAIC 2026, Centec and Lightelligence jointly unveiled China's first 51.2T CPO (Co-Packaged Optics) prototype.
Switching chips have transitioned from a supporting role in network infrastructure to core computing hardware. As a leading domestic commercial high-end Ethernet switching chip provider, Centec will deeply benefit from AI expansion and domestic substitution.
Deepening Focus on Ethernet Switching Chips and Expanding into the AI Computing Power Track
Centec Networks is a 'core supplier of Ethernet switching chips for network infrastructure.' Ethernet switching chips are dedicated chips used for switching and processing massive amounts of data and packet forwarding, serving as the core platform chips for building enterprise networks, carrier networks, data center networks, and industrial networks. Each of the above four key application scenarios adopts an access, aggregation, and core networking architecture, thus requiring a serialized portfolio of Ethernet switching chip products.
Today, there is a strong demand for interconnecting computing power clusters in AI data centers. As the core of networking, Ethernet switching chips require large switching capacities, high-speed ports of 800G and above, and low-latency features. Relying on lossless networks, they must simultaneously accommodate the collective communications for model training and the short-term burst traffic of inference services, stably carrying massive internal traffic and adapting to large-scale cluster expansion. They are indispensable key hardware for unleashing AI computing power.
Centec's current product portfolio covers switching capacities from 100Gbps to 25.6Tbps and port speeds from 100M to 800G:
- In terms of high-end products, the Arctic series chips are the Company's high-end flagship chips targeting large-scale data center and cloud service demands, featuring a maximum port speed of 800G and switching capacities of 8Tbps and 25.6Tbps;
- In terms of mid-range products, the MX products feature a 2.4Tbps forwarding capability, supporting the FlexE slicing network technology and G-SRv6 technology proposed by domestic carriers for next-generation communication technologies;
- In addition, the GoldenGate series chips achieve a switching capacity of 2Tbps, support 100G port speeds, and feature visualization and lossless network capabilities;
- The TsingMa series chips integrate high-performance CPUs to provide enterprises with secure and reliable networks, and offer features such as programmable tunnels and secure interconnections for edge computing.
In 2025, Centec Networks' total revenue was 1.15 billion RMB, of which revenue from Ethernet switching chips was 831 million RMB, revenue from Ethernet switching chip modules was 126 million RMB, and revenue from Ethernet switches was 107 million RMB. The Company provides 12.8T/25.6Tbps high-end chips and modules to New H3C, Ruijie Networks, Maipu, and domestic carriers.
Why Ethernet?
In the complex interconnect ecosystem of AI data centers, multiple communication routes such as InfiniBand, NVLink, and Ethernet are developing in parallel. Why has Ethernet emerged as the winner?
According to IDC data, data center Ethernet switch revenue grew by 53.5% year-over-year to $32.5 billion in 2025, with the incremental volume of Ethernet switching chips coming from AI data centers.
In AI data centers, scale-up involves the vertical integration of resources, allowing GPUs to directly access each other's memory, effectively forming a 'super GPU.' This enhances the performance of tightly coupled tasks but requires ultra-high bandwidth interconnects, such as NVLink, UALink, or emerging Ethernet variants.
Scale-out, on the other hand, horizontally expands computing by adding more servers or racks, connected through low-latency fabrics such as InfiniBand or Ethernet. It is highly suitable for distributed workloads, emphasizing greater efficiency, deployment, management, and scalability.
According to forecasts by 650 Group, the primary incremental growth for Ethernet remains in the Frontend/Scale-out segment, gradually moving towards Scale-up.
Data sources: 650 Group, IDC
According to Dell'Oro data, among Ethernet, InfiniBand, and NVLink, Ethernet's share is steadily increasing.
In the fourth quarter and the full year of 2025, Ethernet switch sales accounted for more than two-thirds of AI cluster data center switch sales, whereas InfiniBand held nearly an 80% share just two years ago. For the full year of 2025, Ethernet achieved a historic overtaking in the AI backend network market share, with related switch sales nearly doubling compared to 2022 levels.
Entering 2026, this trend has further accelerated. In the first quarter of 2026, Ethernet switch sales in AI backend networks doubled once again, with Ethernet accounting for approximately two-thirds of data center AI cluster switch sales.
This shift is driven by the comprehensive transition of hyperscalers such as Amazon, Microsoft, Meta, Oracle, and Elon Musk's xAI. The continuous expansion of AI cluster sizes, coupled with ongoing supply chain constraints, is driving customers' demand for supplier diversification, thereby boosting the demand for Ethernet. Although InfiniBand is still growing, Ethernet is expanding at a much faster pace.
First, in terms of performance, Ethernet is continuously closing the gap with InfiniBand. WWT (World Wide Technology) conducted independent testing in its AI validation lab. By eliminating all variables except network transmission, they concluded that Ethernet/RoCE can rival InfiniBand in many rigorously configured environments. The Ultra Ethernet Consortium's Tomahawk Ultra has further reduced Ethernet switch latency to 250 nanoseconds, narrowing the gap with InfiniBand even more.
Although InfiniBand still holds irreplaceable advantages in extreme scenarios that require training trillion-parameter models with sub-microsecond latency, this scenario may only account for 3% to 5% of the overall market.
In terms of bandwidth specifications, both routes are evolving rapidly:
In the InfiniBand camp, NVIDIA's 2025 Quantum-X800 switch features a single-port speed of 800 Gbps and a total bandwidth of 115.2 Tbps across its 144 ports, targeting the deployment of the highest-performance AI clusters;
In the Ethernet camp, Broadcom's Tomahawk 6, launched in 2025, is the world's first Ethernet switching chip with a total bandwidth of 102.4 Tbps, supporting 1.6 Tbps port speeds. In March 2026, Broadcom announced that the Tomahawk 6 series has entered mass production and shipment. Early adopters include Arista's 7060XE7 series and Edgecore's AIS1600-64Os.
In February 2026, Cisco also released its Ethernet switching chip, the Cisco Silicon One G300, featuring a total switching capacity of 102.4 Tbps and port speeds of 1.6 Tbps.
Dell'Oro predicts that 3200 Gbps ports will begin to penetrate the market before 2030.
Furthermore, cost should be one of the most critical driving forces behind Ethernet's remarkable rise.
For a medium-sized cluster of 512 GPUs, the per-port cost of InfiniBand is 1.5 to 2.5 times higher than that of Ethernet, encompassing the comprehensive costs of switches, network interface cards (NICs), and specialized O&M personnel. Beyond hardware costs, Ethernet also possesses the hidden advantage of 'engineer familiarity.'
Mass Production of 25.6T: Huge Room for Domestic Substitution
Ethernet switching chip design companies are divided into two categories: in-house manufacturers and commercial vendors.
In-house manufacturers are mainly represented by Cisco and Huawei. Their self-developed chips are primarily used for their own switches rather than being supplied to their competitors.
On the commercial front, as the global Ethernet switching chip market expands, in-house manufacturers can no longer meet the growing downstream demand. Consequently, commercial Ethernet switching chip vendors such as Broadcom, Marvell, Realtek, and Centec Networks have emerged globally. Some in-house manufacturers also enrich their switch product lines by purchasing commercial chips externally.
Ethernet switching chips can be categorized into 100M, Gigabit, 10-Gigabit, 25G, 40G, 100G, 400G, 800G, and above. Given the rapid growth of technology industries such as cloud computing, big data, and the IoT (Internet of Things), the demand for 400G and above Ethernet switching chips is gradually increasing, and 800G ports have become the mainstream port form within next-generation data center networks.
Centec Networks' chip products cover switching capacities from 100Gbps to 25.6Tbps and port speeds from 100M to 800G, with the highest mass-produced specification currently at 25.6Tbps. In contrast, overseas leaders such as Broadcom and Cisco have already launched 102.4Tbps new products, indicating that a significant technological gap still exists between domestic and overseas manufacturers.
From the perspective of the industry landscape, the global Ethernet switching chip market is highly concentrated and has long been monopolized by overseas giants. According to QYResearch statistics, in 2024, Broadcom, Marvell, and Cisco collectively accounted for 77.14% of the global market share, solidifying the industry's oligopoly pattern.
In the domestic commercial market, the monopoly of overseas manufacturers is even more pronounced. CIC Consulting data shows that after excluding in-house chip companies like Huawei and Cisco, in the 2020 domestic commercial switching chip market, Broadcom, Marvell, and Realtek held a combined market share of up to 97.8%; the domestic leader Centec Networks held a market share of only 1.6%, leaving vast room for domestic substitution.
Centec's Arctic 25.6T is currently the only mass-produced data center switching chip of the same tier in China. It adopts a 16nm process, supports 32 800G ports, and is equipped with RoCEv2 lossless networking and enhanced visualization and programmability features.
Compared to Broadcom's latest flagship, Centec indeed has a certain generational gap. However, this is not the primary contradiction at present; 'good enough' is more important than 'the best.' The core demand of domestic AI computing power clusters is not to pursue extreme performance, but to achieve autonomy and controllability under export controls.
The announcement of the 2.2 billion RMB related-party order on July 15 may have verified this logic from a commercial perspective.
China Electronics Corporation (CEC) is Centec's largest customer. From 2023 to 2025, the proportion of Centec's sales to the CEC system to its total revenue was 50.44%, 42.29%, and 43.03%, respectively. The previous annual related-party transaction quota had been maintained at 500 to 600 million RMB, but this time it directly jumped to 2.2 billion RMB, an increase of 266%.
As overseas chip controls continue to tighten and the security risks of the computing power supply chain intensify, domestic government, enterprise, and cloud vendors are accelerating the adoption of domestic chips. With Centec's next-generation 51.2T switching capacity chip about to be launched, its performance can match the current mainstream computing power cluster networking demands, balancing practicality and autonomy/controllability. The pace of subsequent industry substitution is expected to continue accelerating.