Deep Dive: Switches
Unveiling the core of AI networking: switch technology explained | Key players in the swtich market | Future trends in CPO and three beneficiary sectors| 16-minute read
In my previous post on data center networking, I briefly highlighted switches as a core component of AI data center networks. However, due to the complexity and importance of the topic, I decided it warranted a dedicated discussion. As promised, this follow-up post delves into the key aspects of switches and examines the investment implications of the evolution of this industry.
Part 1. What are switches?
Switches
In essence, switches are the central hubs used for managing data exchange among devices within the network. Their primary role is to receive data from one device and route it to the appropriate destinations. Switches’ tasks have become increasingly complex in modern AI-driven data centers, where hundreds or even thousands of computing and storage units need to be seamlessly interconnected. Beyond basic data routing, cutting-edge switches from companies like Broadcom and Nvidia have also incorporated processing capabilities into their products - by pre-processing data, they can offload certain computational workloads, and improve overall transmission efficiencies.
Traditionally, switches serve four primary end markets: enterprises, data centers, telecom, and industrial. For this discussion, we will focus on the data center segment, and more specifically switches deployed in modern AI data center networks.
Bandwidth
One of the most key measurements for switches is bandwidth/ throughput, which indicates the total data transmission capacity of a switch. The current state-of-art single-switch chips, such as AVGO’s Tomahawk 5, offer a bandwidth of 51.2 Tbps, which means they are capable of transmitting a maximum of 51.2 terabits of data per second for combined input and output.
Producing advanced switch chips is highly complex. The 51.2T switch chips require 5nm or smaller production nodes, whereas earlier generations (25.6T/12.8T) can be produced using 7nm. This means that China is currently unable to produce 51.2T switch chips, due to the U.S. semiconductor restrictions that have blocked China’s access to nodes beyond 7nm.
SerDes
The main bottleneck in increasing switch bandwidth lies in Serializer/Deserializer (SerDes), which performs the task of converting parallel signals into serial ones (and vice versa) - a key function for enabling high-speed data transmission. Two factors contribute to SerDes’ role as the bottleneck: 1) as an analog chip, SerDes is inherently more challenging to scale in speed compared to digital chips; currently the top in-production SerDes speed is 112 Gbps. 2) SerDes must be placed around the core switch chip (ASIC) so that it is not severely impacted by signal interference from digital signals; for a 5nm switch ASIC, a maximum of 512 units of SerDes can be arranged on the top and bottom sides of the chip (the shorter left and right sides are reserved for HBM). Therefore, the total maximum throughput of a single-switch chip is currently at 51.2 Tbps (= 112 Gbps/ SerDes × 512 SerDes channels).
Companies capable of producing 112G SerDes include Broadcom ($AVGO), Nvidia ($NVDA, via Mellanox), Marvell ($MRVL), Cisco ($CSCO), and Huawei (private). However, even though having the necessary SerDes technology, Huawei remains unable to produce 51.2T switches due to its inaccessibility to 5nm manufacturing. The next milestone for SerDes technology is reaching 224G speed. AVGO and NVDA are at the forefront of this advancement, with 224G solutions already prepared, while MRVL and others are trailing in development.
Part 2. Competitive landscape of switches (in AI data centers)
Switch chips
For switch chips, NVDA and AVGO are the leading providers, with MRVL being a small third. In the IB environment, NVDA undoubtedly dominates, thanks to its ecosystem advantage. In the Ethernet environment, however, AVGO has traditionally been the market leader, holding c.70% share, while MRVL accounts for c.20% and other firms (incl. INTC) represents the remaining 10%.
Recognizing the growing interests of Ethernet for AI reasoning and even training workloads in the future, however, NVDA has been actively expanding its presence in Ethernet to ensure it remains relevant. Through its Spectrum series, NVDA has been closing the gap, and pushing for a duopoly alongside AVGO in the high-end Ethernet switch chip segment.
According to industry experts, NVDA's and AVGO's current Ethernet solutions have quite comparable performance metrics. NVDA’s advantage lies in its deeper understanding of the entire network ecosystem, thanks to its dominance in other critical components within the network, like GPUs and NICs. AVGO, on the other hand, benefits from a longer track record in switch chip production and strong, established partnerships with key manufacturers. The expectation is that NVDA and AVGO combined can capture a majority share (75%+) of the high-end Ethernet switch chip market going forward.
Switch boxes
Switches can be classified as either black-box (branded) or white-label (ODM/JDM and rebranded). In today’s AI data center, less than 50% of the switches are white-labels, already a higher percentage than what’s typically seen in traditional enterprise networks, as most major customers for AI network switches are hyperscalers that have strong in-house technical expertise and substantial purchase volume/ power.
In the black box segment, the main players are Arista ($ANET), Cisco ($CSCO), and Juniper ($JNPR); however, JNPR’s presence in AI data centers remains limited, and its primary market is still the traditional enterprise networks. Among the vendors, ANET stands out as the leading player. Its solutions rely heavily on AVGO’s chips. ANET is known for its strong software and system integration capabilities, which complement well with AVGO’s strong hardware (chip) expertise.
White label solutions, on the other hand, have advantageous costs, typically around 20% cheaper than comparable black-box offerings. The major players in this segment are Celestica ($CLS) and Accton (2345.TP), both of which also have primarily used AVGO’s chips to date.
Among major customers in this space, Microsfot ($MSFT) is viewed as relatively "conservative," as it favors more mature black-box solutions or ODM offerings that require minimal customization on its part. As a result, Arista serves as its primary supplier. In contrast, Google ($GOOG), Meta ($META), and Amazon ($AMZN) have taken a more balanced approach, showing greater openness to white-label solutions and deeper involvement in the customization process.
Part 3. The shift toward co-packaged optics (CPO)
CPO penetration
One notable trend in the switch industry is the growing interest toward adopting the CPO architecture. In October of this year, NVDA introduced the CPO version of its flagship Q3400 IB switch series, aiming for production in H2 2025. As I noted in my previous report, CPO is a technology that integrates optical engines directly with the switch chips, achieving a higher level of integration that significantly reduces power consumption and latency:
CPO is an emerging technology that takes even a step further from SiPh – aimed at integrating optical transceiver (main functional part of optical module) directly with switch chip within the same physical package, often using co-packaging techniques, and effectively eliminating the need for separate optical modules.
Through removing the need for external optical modules, CPO shortens the connection paths and enhances integration, leading to 30-50% reduction in power consumption and lower latency… making it a high-performance solution ideal for large-scale, demanding AI networks. – Deep Dive: Optical Module Market
Currently, the CPO roadmap still faces several challenges:
Compatibility: The CW laser used to drive optical signals in SiPh modules (employed in CPO switches) can have different wavelengths by vendors, leading to signal incompatibility between CPO switches from different manufacturers
Troubleshooting: Integrating optical engines with switch chips increases the complexity and cost of repairs - unlike traditional setups where optical modules can be easily inspected and replaced independently
Price: The supply chain needs time to adapt and scale to this entirely new architecture, resulting in much higher costs for CPO switches compared to traditional models at this stage. For instance, the standard version of NVDA's Q3400 switches is priced at $30-50K, while the CPO version is expected to launch at over $100K
Therefore, the adoption of CPO architecture will take time. Currently, the optical module industry is at the 800G (transitioning to 1.6T) bandwidth nodes, where non-CPO architectures are still sufficient to handle transmission workloads effectively. Industry experts predict that at the next generation of 3.2T bandwidth (2026-2027 timeframe), the standalone optical modules will remain mainstream, with CPO potentially capturing c.15% of the market. However, as the industry advances to the 6.4T node by 2029-2030, the increasing challenges of power consumption and signal loss of traditional methods as bandwidth increases may render them impractical. At that point, CPO may become a necessary advancement, potentially positioning itself as the main approach in the market.
Key beneficiaries of the CPO trend
Optical engines
CPO differs from the traditional setup in that it integrates the optical engine directly with the switch chip, effectively eliminating the standalone optical modules that are previously plugged into the switch. This shift opens the door for new vendors to emerge as optical engine suppliers. In NVDA's case, the key beneficiary is TFC Communication (300394.SZ), which supplies the optical engines for NVDA's upcoming CPO switches.
NVDA began developing CPO technology in 2020. It initially sought partnerships with companies like Zhongji Innolight (300308.SZ), Coherent ($COHR), and Eoptolink (300502.SH). However, these firms were not particularly proactive, as they viewed CPO as still far from mainstream adoption and were wary of its potential replacement of standalone optical modules, which offer higher profit margins for these vendors. This hesitation created an opportunity for TFC, a smaller company at the time with limited exposure to high-margin optical components, to step in and collaborate with NVDA on CPO development. Now, as CPO technology nears production readiness, TFC is positioned to emerge as a key supplier for this technological transition.
Optical/ fiber splitter
Optical/ fiber splitter is a key passive component used in the CPO switches. For example, In the CPO version of NVDA's Q3400 switch, there are 144 ports operating at 800G, each with 16 channels (8 input + 8 output). This amounts to a total of 2,304 channels (= 144 ports * 16 channels per port). To manage these channels, a fiber splitter is used to distribute the CW laser light source to the 36 optical engines in the switch and then organize the processed optical signals into 2,300+ channels that are connected to MPO ports for subsequent transmission through the fiber cables. The key components inside a fiber splitter are MT ferrules and fibers.
Previously, fiber splitter boxes could be manually arranged by human operators, as the number of optical lanes the system needed to manage was relatively limited. However, with the latest CPO solutions, the number of optical lanes has increased to over 1,000, making manual labor impractical and prone to errors. To address this, the latest generation of fiber splitter boxes are equipped with automated mechanisms to arrange fiber channels.
The price of a splitter box largely depends on the number of channels it handles. T&S Communications, a major player in the field, offers its 300-channel splitters at c.$1,000, 500-channel splitters at $1,600, and its 1,000+ channel splitter, used in NVDA's Q3400 (CPO version), at over $3,000.
Key producers of fiber splitters include Molex (private) and T&S Communications (300570.SZ). T&S holds a price advantage of c.20% compared to Molex. Industry experts also highly commend T&S for its R&D capabilities and product quality. For NVDA’s Q3400 switch, T&S has been actively involved in its fiber splitter design for the past two years and is likely to become the main supplier when the product is launched. T&S is well-positioned to benefit from NVDA’s push to CPO.
CW lasers
Another key beneficiary of the CPO trend is the CW laser manufacturers. CPO switches rely on SiPh modules for optical processing. In these SiPh modules, CW lasers are utilized on the transmitter side. Leading producers of CW lasers include Yuanjie (688498.SH), Shijia (688313.SH), and Lumentum ($LITE). For more detailed information, please refer to my previous post on the optical module market.
Part 4: Key producers of switch chips and swiches
In the paid section below, I will profile the four key companies in the switch chip and switch markets. I will discuss their projected revenues, strategic approaches, and main customers within the industry. The four companies are:
Broadcom ($AVGO)
Nvidia ($NVDA)
Arista ($ANET)
Celestica ($CLS)