What Is 400G Ethernet? Standards, Technologies, and Deployment in Modern AI Data Centers

Introduction

According to IDC data, 200G and 400G switches together accounted for 43.9% of data center switch revenue in 2025, while 800G accounted for 16.4%. By the first quarter of 2026, 800G had rapidly increased its revenue share to 35.8%, slightly surpassing the combined 34.1% share of 200G and 400G switches.

This shift shows that 800G is rapidly gaining adoption in new, high-end AI data center fabrics. However, 200G and 400G Ethernet still have a large installed base and broad deployment footprint. The rapid growth of 800G does not make 400G obsolete. Instead, 400G remains an important upgrade path for many existing data center networks.

Against this market backdrop, Asterfusion has introduced three new switches for the 400G market: CX764QO-N, CX764QD-N, and CX764QH-N. All three models provide 64 × 400G ports and up to 25.6 Tbps of switching capacity in a 2U form factor. With native 400G ports and flexible breakout connectivity, these switches can help customers upgrade their networks from 100G to 400G while preserving their existing connectivity architecture.

So, with 800G developing rapidly, why does 400G Ethernet still deserve attention? Why is Asterfusion introducing new 400G switches at this stage?

This article examines the standards, key technologies, and deployment scenarios behind 400G Ethernet. It also explores the role of 400G in modern AI data centers and its relationship with 800G.

What Is 400G Ethernet

400G Ethernet, also known as 400 Gigabit Ethernet (400GbE), refers to a class of high-speed Ethernet technologies that provide a transmission rate of 400 Gb/s. Compared with 100GbE, 400GbE delivers four times the port bandwidth, significantly increasing network aggregation capacity and switching density with the same number of ports.

It is important to note that 400G describes the logical data rate of an Ethernet interface. It does not refer to a specific optical transceiver, optical wavelength, connector, or transmission medium. A 400GbE network typically includes interfaces operating at different speeds. For example, servers or storage nodes may connect to the Leaf layer at 25G, 50G, or 100G, while native 400G links are used between the Leaf and Spine layers for bandwidth aggregation.

The foundational standards for 400GbE are defined by IEEE 802.3bs. Approved in 2017, the standard defines the MAC parameters, physical layer specifications, and management parameters required for 200GbE and 400GbE. It provides the standards foundation for subsequent 400G implementations across different media types, lane rates, and optical transceiver form factors.

In modern data centers, 400G is commonly deployed where high-bandwidth aggregation and large-scale horizontal scaling are required, including:

• Uplinks in a Leaf–Spine architecture or high-bandwidth interconnects between switching layers
• Super-Spine layers and high-capacity interconnects across Pods
• High-bandwidth uplinks for data center interconnect (DCI) or data center edge networks
• Interconnects between compute nodes, storage nodes, and the switching fabric in HPC clusters
• The underlying fabric for GPU clusters, distributed storage, and high-speed data pipelines in AI data centers

For many data centers evolving from 100G to higher speeds, 400G serves as an important intermediate bandwidth tier. Native 400G links can support fabric aggregation and backbone connectivity. With appropriate switch, transceiver, and cabling support, a 400G interface can also be broken out into multiple lower-speed ports to match the existing connection speeds of servers and GPU NICs.

How Does 400G Ethernet Work?

400G Ethernet is not enabled by a single technology. It is fundamentally a 400 Gbps logical interface rate. Actual transmission relies on a combination of PAM4 modulation, multiple high-speed SerDes lanes, Forward Error Correction (FEC), and different optical or cable interconnect technologies.

PAM4 Enables Higher-Speed Transmission

The first step in understanding 400G is understanding PAM4 (4-level Pulse Amplitude Modulation). Although PAM4 was introduced during the 50G era, the transition to 400G marked its large-scale commercial adoption as a defining modulation technology in data center Ethernet.

Before 400G, traditional Ethernet widely used NRZ (Non-Return-to-Zero) signaling. NRZ uses only two voltage levels, high and low. Each symbol can therefore carry only 1 bit of information, either 0 or 1. Increasing network speeds requires a continuous increase in the baud rate. This creates greater challenges for signal integrity, transmission loss, and high-speed hardware design.

PAM4 uses four distinct amplitude levels (00, 01, 10, and 11), allowing each symbol to carry 2 bits of data. This means that PAM4 can double the data rate without increasing the baud rate.

Technology	NRZ	PAM4
Signal levels	2	4
Bits per symbol	1	2
Signal complexity	Lower	Higher
Signal margin	Larger	Smaller
Use in high-speed Ethernet	Earlier generations	200G/400G/800G and beyond

However, higher transmission efficiency comes with trade-offs. Because the four PAM4 signal levels are closer together, PAM4 has a lower signal margin than traditional NRZ. As a result, the link is more sensitive to loss, noise, crosstalk, dispersion, and receiver equalization.

This means reliable 400G Ethernet operation depends not only on PAM4 itself, but also on several other key technologies.

Forward Error Correction (FEC)

PAM4 increases the amount of information carried by each symbol, but it also reduces the signal margin between levels. To maintain acceptable bit error rates and link stability at high transmission speeds, FEC (Forward Error Correction) is an important part of 400G Ethernet.

FEC allows the receiver to detect and correct a certain number of transmission errors without retransmitting the data. For high-speed Ethernet links based on PAM4, FEC is generally not an optional feature that can be freely disabled. It is part of the link design and interoperability requirements.

For example, some 400G PAM4 implementations use RS(544,514) FEC. The actual FEC type depends on the specific PHY, transceiver, and device implementation.

Components of 400G Ethernet Networking

Deploying a 400G Ethernet network involves more than the switch itself. It includes high-speed SerDes lanes, port breakout, pluggable transceiver form factors, optical PMDs, and different types of interconnect cables. Together, these components determine how a 400G port can connect, break out, and transmit data.

400G Ethernet Switch

A 400G Ethernet switch is the core device in the network. It provides high-speed 400G ports and the required switching capacity.

Asterfusion offers multiple 400G switches for different data center and AI networking requirements, with switching capacities ranging from 12.8 Tbps to 25.6 Tbps.

Product	Throughput	Chip	Ports	Height
CX764QO-N	25.6 Tbps	Marvell Teralynx 10	64 × 400G OSFP2x 10G SFP+	2U
CX764QD-N	25.6 Tbps	Marvell Teralynx 10	64 × 400G QSFP-DD2x 10G SFP+	2U
CX764QH-N	25.6 Tbps	Clounix	64 × 400G QSFP1121x 25G SFP28	2U
CX732Q-N	12.8 Tbps	Marvell Teralynx 7	32 × 400G QSFP-DD2 x 10G SFP+	1U
CX732Q-N-ORv3	12.8 Tbps	Marvell Teralynx 7	32 × 400G QSFP-DD2 x 10G SFP+	1U ORv3

Different models may vary in switching ASICs, port configurations, network features, and deployment scenarios. When selecting a 400G switch, it is therefore important to consider more than port count and total switching capacity. Module compatibility, breakout capabilities, and future network upgrade requirements should also be taken into account.

For 64-port AI switches comparison, please refer to 64-Port 400G AI Switches Compared: Key Differences and How to Choose

400G Switch Port Form Factors

Currently, the common high-speed port form factors for 400G Ethernet switches include:

• QSFP-DD: Backward compatible with QSFP28, QSFP+, and QSFP112. It is widely used in cloud computing and enterprise data center deployments.
• OSFP: Slightly larger than QSFP-DD, with better thermal performance and electrical characteristics. Its higher power-handling capability makes it a common choice for high-density AI compute fabrics, such as those based on the NVIDIA ecosystem.
• QSFP112: Retains the traditional QSFP form factor while increasing the electrical lane rate to 112G PAM4. It is suitable for high-density deployments that aim to maintain the existing physical layout.

These form factors determine which types of high-speed transceivers can be used on the switch front panel. They must also match the corresponding transceiver form factor and host interface specifications.

Multiple High-Speed Lanes Determine Breakout Options

400G refers to the total logical data rate of a port. It does not mean that every 400G interface uses a single physical signal. A 400GbE interface is typically built from multiple high-speed SerDes lanes. The number of lanes and the per-lane data rate vary across implementations.

Early 400G implementations could use 16 × 25G-class NRZ lanes. Today, common 400G port architectures in data centers include 8 × 50G-class PAM4 electrical lanes and 4 × 100G-class PAM4 electrical lanes. These lanes are aggregated to provide a single 400GbE interface.

The lane architecture also provides the foundation for port breakout. For example, a 400G port based on 8 × 50G PAM4 lanes can operate as 1 × 400G. Depending on the port and transceiver implementation, the lanes can also be grouped into 2 × 200G, 4 × 100G, or 8 × 50G independent interfaces.

400G Ethernet Optical Implementations and Physical Medium Dependent (PMD) Specifications

In 400G Ethernet networks, optical PMD (Physical Medium Dependent) specifications define how transceivers transmit signals over fiber, including the optical lane architecture, fiber type, wavelength scheme, connector type, and transmission distance.

The same 400G switch port can support different PMD specifications depending on the application, transmission distance, and fiber type:

• DR4 (400GBASE-DR4): Uses four parallel single-mode fiber lanes with an MPO-12 connector. Each lane carries a 100G PAM4 signal, supporting distances of up to 500 m. It is commonly used for Leaf–Spine interconnects within a data center.
• FR4 (400GBASE-FR4): Uses CWDM to multiplex four different wavelengths over a duplex pair of single-mode fibers. Each wavelength carries a 100G PAM4 signal. It typically uses duplex LC connectors and supports distances of up to 2 km. Compared with DR4, FR4 can transmit 400G over just two single-mode fiber strands, making it more suitable for deployments where fiber count is a constraint.
• 400GBASE-LR4: Uses a multi-wavelength single-mode optical transmission scheme for 400G Ethernet connections over longer distances than FR4. It is suitable for data center interconnects and other applications that require longer fiber links.
• SR8 / SR4.2: Designed for short-reach deployments over multimode fiber (MMF), primarily for connections within the same rack or between adjacent racks. SR8 uses eight parallel optical lanes, while SR4.2 uses four bidirectional fiber pairs with two wavelengths per pair.
• 400G VR4: A very-short-reach multimode optic that uses four 100G PAM4 optical lanes. It is designed primarily for intra-rack and short adjacent-rack interconnects, supporting distances of up to 50 m over OM4 fiber.

400ZR for DCI

In 400G networks, another specification deserves particular attention: 400ZR.

400ZR was defined by the OIF (Optical Internetworking Forum) with contributions from the industry. It is designed to provide a standardized, cost-effective, high-density pluggable coherent optical solution for metro data center interconnect (DCI). It uses single-wavelength coherent transmission to carry 400GbE over DWDM networks, targeting single-span data center interconnects of approximately 80–120 km.

400ZR is used to connect geographically distributed data centers, such as Edge and regional data centers, providing high-bandwidth interconnects across data centers within the same metropolitan area.

Cable Types

The cable types used in 400G networks are broadly similar to those used in 100G and 200G networks. Common options include:

• DAC (Direct Attach Cable): Uses copper as the transmission medium and has no active electronic components inside the cable. It offers very low cost and near-zero power consumption.
• ACC (Active Copper Cable): Integrates redriver chips at the cable ends to compensate for high-frequency attenuation in copper cables.
• AEC (Active Electrical Cable): Integrates advanced DSP/retimer chips at the cable ends. In addition to boosting the signal, it can resample and reduce noise in PAM4 signals, making AEC an important copper interconnect technology in 400G networks.
• AOC (Active Optical Cable): Integrates optical transceiver components directly into both ends of the cable, with optical fiber used for transmission. AOCs offer lower weight, immunity to electromagnetic interference, and longer transmission distances.

400G Ethernet continues to use the DAC, ACC/AEC, AOC, and pluggable optical transceiver options commonly found in high-speed data center interconnects. Compared with 100G, the main changes in 400G are higher per-lane data rates and stricter signal integrity requirements. As a result, the usable distance of passive copper cables is generally shorter, while AECs, AOCs, and optical transceivers become more important for inter-rack and longer-distance connections.

Why 400G Ethernet Matters for Modern Data Centers

400G remains one of the most widely deployed, mature, and practical high-bandwidth Ethernet speeds available today. One reason is that it provides a balanced aggregation layer between 100G/200G access nodes and higher-bandwidth Spine, Super-Spine, or DCI layers.

For many cloud data centers, enterprise private clouds, IDCs, HPC environments, and AI storage networks, an architecture with 100G or 200G access nodes and 400G Leaf–Spine uplinks can provide a reasonable balance between bandwidth, port density, optical costs, power consumption, and reuse of existing infrastructure. The ecosystem around 400G switching ASICs, QSFP-DD/OSFP transceivers, DAC/AEC/AOC cables, fiber cabling, and NOS support has also become mature.

400G also provides a gradual upgrade path. Servers, GPU nodes, and storage nodes can continue to use 25G, 50G, 100G, or 200G NICs. With the appropriate port modes and cabling, 400G switch ports can use breakout to provide multiple lower-speed interfaces or operate as native 400G uplinks between Leaf and Spine switches. As a result, when expanding an existing 100G/200G network, upgrading to 400G generally requires less architectural change than moving directly to an 800G-based design.

Key Challenges When Deploying 400G Ethernet

While 400G Ethernet provides significant bandwidth advantages, its deployment also introduces several challenges.

Signal Integrity

As discussed earlier, 400G Ethernet widely uses PAM4, which makes links more sensitive to loss, noise, crosstalk, dispersion, and receiver equalization.

This means signal integrity in a 400G link cannot be evaluated simply by checking whether the link comes up. During deployment and operations, it is also important to monitor metrics such as pre-FEC BER, post-FEC BER, FEC corrected and uncorrectable codewords, module temperature, and received optical power. For short-reach interconnects such as DAC, ACC, AEC, and AOC, the cable type should also be selected based on the actual distance, cable specifications, and port electrical interface. This helps avoid intermittent errors caused by insufficient electrical margin in the copper channel.

Interoperability

IEEE standards define common MAC, PCS, and physical-layer specifications for 400GbE, but they do not guarantee interoperability across all multi-vendor implementations. Actual links also involve details such as module form factors, host electrical interfaces, FEC modes, optical PMDs, DSP implementations, DOM/DDM thresholds, switch NOS support, and breakout configurations.

For example, two devices may both support 400G but use different port form factors, such as OSFP and QSFP-DD. One end may also be configured as 4 × 100G breakout while the other remains in native 400G mode. Even when the form factor and data rate match, incompatible FEC settings, transceiver support, or fiber connections can prevent the link from coming up or cause excessive errors. Before large-scale deployment, interoperability should therefore be validated across the target combination of switches, NICs, transceivers, cables, and network operating systems.

Thermal Management

400G transceivers, AECs, and high-density switch ports consume significantly more power than lower-speed interfaces. Deploying 32 or 64 high-power 400G optical modules in a 1U or 2U switch can create a substantial thermal load at the front panel.

Thermal design affects both module lifetime and link stability. Excessive temperatures can reduce the performance margin of lasers, DSPs, and receiver circuits, increasing the risk of bit errors and link flaps. During deployment, operators should evaluate airflow direction, inlet and outlet temperatures, module power classes, unused-port blanks, rack airflow, and the thermal density around adjacent ports. The selected modules should also comply with the switch vendor’s specified power and operating-temperature limits.

Optics and Cabling

400G transceivers and cables must be selected based on the transmission distance, available fiber infrastructure, port form factor, and network upgrade path. PMDs such as DR4, FR4, LR4, SR8, SR4.2, and VR4 differ in connector type, fiber count, transmission distance, and cabling requirements. Form factors such as OSFP, QSFP-DD, and QSFP112 determine whether a transceiver is physically and electrically compatible with the switch port.

For example, DR4 typically uses an MPO-12 connector and eight active single-mode fiber strands, while FR4 uses two single-mode fiber strands with duplex LC connectors to carry 400G. If a data center already has extensive two-strand single-mode fiber infrastructure, FR4 may provide better reuse of existing cabling. If 400G optical breakout to 4 × 100G is required, the complete compatibility of the transceiver, MPO polarity, breakout cable, and remote ports should be verified in advance.

Network Congestion

400G increases link bandwidth, but it does not eliminate congestion by itself. In AI, HPC, distributed storage, and large-scale microservice environments, traffic patterns such as incast, all-to-all communication, and checkpoint traffic can still create hotspots at the Leaf, Spine, or receiver.

For RoCEv2 networks, congestion control therefore requires additional design and attention. ECN, PFC, DCQCN, ECMP load balancing, queue configuration, and buffer strategies should be designed according to the application traffic patterns. Telemetry should also be used to monitor queue depth, ECN marking, PFC pause frames, packet loss, and tail latency. AI workloads require predictable latency, lossless throughput, and stable iteration times. These requirements cannot be achieved simply by increasing port speeds.

How to Design a 400G Data Center Network and Use Case

A 400G AI network does not necessarily mean that every link operates at 400G. Spine switches, Leaf switches, and GPU server NICs can use different port speeds, resulting in six common deployment combinations. In practice, three deployment scenarios are more commonly used, as shown below:

Option	Spine	Leaf	GPU Server NIC
1	400G	400G	400G
2	400G	200G	200G
3	400G	100G	100G

Option 1: Full 400G Architecture (400G Spine – 400G Leaf – 400G NIC)

This architecture provides an end-to-end 400G ultra-high-bandwidth pipeline tailored for large-scale AI model training and massive data parallel processing.

Hardware Configuration from Asterfusion:

• Spine Switch: Asterfusion CX764QO-N (400G)
• Leaf Switch: Asterfusion CX764QO-N (400G)
• GPU Server NIC: 400G NIC

Network Topology:

Key Deployment Scenario:

Ideal for premier AI data centers and supercomputing clusters requiring zero-bandwidth bottlenecks between compute nodes and spine layers.

Option 2: Balanced 200G Access Architecture (400G Spine – 200G Leaf – 200G NIC)

A balanced cost-to-performance architecture utilizing 200G interfaces across the leaf-to-server access tier.

Hardware Configuration:

• Spine Switch: Asterfusion CX764QO-N (400G)
• Leaf Switch: Asterfusion CX664D-N (200G)
• GPU Server NIC: 200G NIC

Network Topology:

Key Deployment Scenario:

Perfect for medium-to-large scale AI clusters balancing 400G spine scale with optimized 200G leaf switch density.

Option 3: Cost-Optimized 100G Access Tier (400G Spine – 100G Leaf – 100G NIC)

A cost-effective entry option utilizing proven 100G access infrastructure while maintaining a scalable 400G core.

Hardware Configuration:

• Spine Switch: Asterfusion CX764QD-N (400G, breakout to 100G connecting leaf switches)
• Leaf Switch: Asterfusion CX532P-N (100G)
• GPU Server NIC: 100G NIC

Network Topology:

Key Deployment Scenario:

Best suited for edge AI inference workloads, small-scale model training, or enterprise data centers scaling out core bandwidth on a controlled budget.

Which Architecture Should You Choose?

Cluster requirement	Recommended architecture
Maximum GPU training throughput with 400G NICs	Option 1
Balanced 200G AI training pod design	Option 2
Smaller, cost-sensitive 100G AI/HPC cluster	Option 3

Conclusion

400G Ethernet remains an important building block for modern data center and AI networks, even as 800G Ethernet gains momentum. Its value lies not only in higher port bandwidth, but also in its flexible deployment model. A 400G network can combine 400G Spine links with 400G, 200G, or 100G Leaf and GPU server connections, allowing organizations to adapt network capacity to different GPU, workload, and infrastructure requirements.

As this guide shows, there is no single deployment model for a 400G data center network. From 400G Spine–400G Leaf–400G GPU NIC to 400G Spine–100G Leaf–100G GPU NIC, different combinations can be built with appropriate switches, transceivers, breakout connections, and cabling. The choice of OSFP, QSFP-DD, or QSFP112 interfaces, as well as the appropriate optical or copper interconnect, also plays an important role in deployment.

For organizations building or upgrading AI infrastructure, 400G provides a practical path between existing 100G/200G networks and emerging 800G architectures. With flexible port configurations and breakout capabilities, 400G switching can support both high-bandwidth AI fabrics and gradual network upgrades.

With its 400G switch portfolio, flexible port configurations, and support for high-speed Ethernet and RoCE deployments, Asterfusion provides the switching infrastructure needed to build scalable 400G data center and AI networks across different deployment scenarios.

Explore More in Our 400G Technical Series

Building a resilient, high-performance 400G infrastructure requires a unified understanding of standards, interconnects, and hardware choices. Whether you are scaling an AI data center or upgrading campus backbones, explore our comprehensive 400G guide series to master every layer of the architecture:

• 1. Fundamentals: Learn core standards and deployment models in What Is 400G Ethernet? Standards, Technologies, and Deployment in Modern AI Data Centers.
• 2. Transceiver Specifications: Compare reaches, fiber types, and form factors in 400G Transceivers Explained: VR4, SR4, SR8, DR4, FR4, LR4, ER4 and ZR.
• 3. Coherent DCI Solutions: Explore how 400G ZR and coherent optics enable direct switch-to-switch data center interconnects in What Is 400ZR? How 400G Switches Enable Coherent DCI.
• 4. Switch Architecture: Solve bandwidth and latency bottlenecks by choosing the right platform in How to Choose a 64-Port 400G AI Switch: QSFP-DD vs. OSFP vs. QSFP112.
• 5. 400G QSFP-DD Cable Connection for AI Fabric: 400G QSFP-DD Cable Connectivity for AI Fabrics: DAC vs. ACC vs. AEC vs. AOC
• 6. AI Checkpoint Traffic: How 400G Switches Prevent Network Bottlenecks in GPU Training
• 7. 400G AI Switch Real-World Performance Test: Validating Lossless Ethernet with H100 GPUs