With the rapid development of artificial intelligence, large model training, and cloud computing, the construction of global data centers has entered a new round of explosion.The requirements of AI servers for bandwidth, power density, and power consumption far exceed those of traditional servers, which directly promotes the rapid iteration of high-speed connector technology and the rapid growth of market demand.This paper systematically analyzes the changes in demand for high-speed connectors in AI servers and data centers from the four dimensions of technological evolution, market size, product type, challenges and trends.
1. AI power burst drives the surge in demand for high-speed connectors
1.1 Rapid expansion of the AI server market
In 2026, the global AI server market has exceeded the $100 billion mark, with a compound annual growth rate of more than 40%.The three major scenarios of large model training, inference deployment, and edge AI computing jointly promote the construction of computing power infrastructure into the acceleration channel.The connector usage and value of a single AI server are significantly higher than those of general-purpose servers, which brings structural growth opportunities to the high-speed connector market.
The core differences in AI servers compared to traditional x86 servers are:
- High number of GPU/AI accelerator chips: 8 cards and 16 cards on a single machine have become mainstream configurations
- High interconnect bandwidth requirements: Super high-speed interconnection between GPUs and between GPUs and CPUs is required
- High power density: Stand-alone power consumption up to 10kW or more, higher requirements for power connectors
- Large storage capacity: More DDR5, HBM, NVMe SSDs available
- Differential Pair Structure: Separate shielding for each pair of signal lines to reduce crosstalk
- Orthogonal midplane: Front and back connectors are arranged orthogonally to increase wiring density
- Cable backplane: Reduce signal loss by replacing traditional PCB backplanes with high-speed cables
- Coplanar structure: Alternate signal and ground pins to optimize impedance
- Operating voltage reduced from 1.2V to 1.1V, reducing power consumption
- Data prefetching increased from 8n to 16n
- Improve reliability with the addition of on-chip ECC
- Using a dual channel architecture, the same memory bar is divided into two independent channels
- Insertion Loss: The attenuation of high-speed signals in conductors and media increases with frequency
- Return Loss: Signal reflection due to impedance discontinuity
- Crosstalk: Electromagnetic interference between adjacent signal channels
- Jiggle: The time offset of the signal transition edge, including deterministic jitter and random jitter
- Contact resistance rise: High temperature accelerates the oxidation of the contact surface, increasing the contact resistance
- Material softening: Plastic insulators may soften and deform at high temperatures
- Thermal expansion mismatch: Differences in thermal expansion coefficients between materials can lead to stress accumulation
- Accelerated aging: Accelerates the chemical aging process of materials at high temperatures
- Mating cycles
- Retention force
- Vibration and impact resistance
- Contact resistance stability
- Environmental adaptability (temperature, humidity, salt spray, etc.)
- PCIe 6.0 will be commercially available on a large scale around 2027
- 112g SerDes will become standard for high-speed interconnection
- Memory rate will move above 10000MT/s
- Optical interconnection rate evolution to 1.6T, 3.2T
- In-rack interconnect from DAC to AOC
- Onboard light engine inside the server
- CPO technology matures and becomes commercially available
- The cost of silicon-optical technology continues to decline
- CXL Interconnect: Compute Express Link enables consistent interconnection of CPU, GPU, and memory
- UCIe: Universal Chiplet Interconnect Express enables high-speed interconnection between cores
- Wireless connectivity: 60GHz mmWave and other short-range wireless technologies for inter-board communication
- Liquid-cooled connector: The popularity of liquid cooling heat dissipation drives the demand for special liquid cooling connectors
1.2 Increased value of high-speed connectors
In AI servers, the stand-alone value of high-speed connectors is 2-3 times higher than that of traditional servers.The high-speed connectors of traditional servers are mainly concentrated in network cards and storage interfaces, while AI servers have added GPU interconnects, NVLink, CXL, OCP accelerators and other high-speed interconnect scenarios.
II. Main product types of data center high-speed connectors
2.1 PCIe connector
PCI Express (PCIe) is the most popular high-speed interconnect inside servers.With the popularity of PCIe 5.0 and the advancement of PCIe 6.0, the single-channel data rate has increased from 16GT/s in PCIe 4.0 to 32GT/s (PCIe 5.0), and PCIe 6.0 has reached 64GT/s.The improvement of the rate has put forward extremely high requirements for the signal integrity design of the connector.
Comparison of PCIe specifications by generation:
In AI servers, PCIe connectors are used not only to connect GPU acceleration cards, but also to connect NVMe SSDs, CXL memory expansion devices, DPU smart NICs, and more.The number of channels, crosstalk control, and impedance matching capabilities of the connector directly determine the overall performance upper limit of the system.
2.2 High-speed backplane connector
The backplane connector is the “skeleton” of large servers and switches and is responsible for signal transmission between the motherboard and the functional boards.Due to the large number of GPUs and complex interconnect topologies, the density and rate requirements of backplane connectors have been greatly improved.
Current mainstream high-speed backplane connector technology routes include:
2.3 Memory Connectors
DDR5 memory has become the standard for AI servers. The data rate of a single DDR5 memory bar can reach 6400MT/s, and will increase to more than 8400MT/s in the future.AI servers are usually equipped with 8, 12, or even 16 memory bars, with a total memory capacity of several terabytes.
The DDR5 connector has the following key changes compared to DDR4:
In addition, the way HBM (High Bandwidth Memory) is interconnected as a core component of GPUs is also evolving.HBM3/HBM3e achieves three-dimensional stacking through TSV (silicon vias) and micro bumps, with a single capacity of more than 24GB and a bandwidth of more than 1TB/s.
2.4 Optical Interconnects and Optical Modules
With the expansion of AI clusters, the interconnection bandwidth between servers has become a new bottleneck.InfiniBand, RoCE and other high-speed network technologies have developed rapidly, 400g and 800g optical modules have been deployed on a large scale, and 1.6T optical modules have begun to be commercialized.
Inside the data center, optical interconnects are penetrating from rack to rack and even inside servers.The demand for optical connectors, active optical cables (AOCs), optical engines and other products on the panel continues to grow.CPO (co-packaged optics) is to package the optical module with the ASIC chip of the switch to further shorten the transmission distance of the electrical signal.
III. Technical Challenges Facing High-Speed Connectors
3.1 Signal Integrity Challenge
As the rate increases above 25Gbps, signal integrity issues become extremely complex.Key challenges include:
In order to meet these challenges, connector manufacturers need to continue to invest in material selection, structural design, simulation and optimization.New technologies such as low-loss LCP materials, air media, and 3D shielding structures continue to emerge.
3.2 Thermal Management Challenges
The power consumption density of AI servers has been greatly improved, and the power consumption of stand-alone servers has jumped from the traditional 1-2kW to 10kW or even higher.Under high temperature environment, the contact resistance, material aging, mechanical strength and other properties of the connector will be affected.
The impact of thermal management on the connector is mainly reflected in:
3.3 Balance of high density and mechanical reliability
In order to achieve higher bandwidth in limited space, the pin spacing of the connector is continuously reduced, and the pin density continues to increase.However, the smaller the spacing, the more difficult it is to process, and the more difficult it is to ensure mechanical reliability.
Key reliability metrics include:
IV. Future Trends
4.1 Rates continue to improve
Server interconnection rates will continue to double over the next 3-5 years:
4.2 Photoregressive copper regression continues to deepen
With the increase in speed, the transmission distance limit of copper cables is becoming more and more obvious.Optical interconnects will continue to penetrate shorter-distance scenarios:
4.3 Exploration of new interconnected technologies
In addition to the traditional electric interconnection and optical interconnection, some new interconnection technologies are also being explored:
4.4 Domestic substitution acceleration
In the context of autonomous and controllable AI computing power, the domestic replacement process of high-speed connectors has been significantly accelerated.Domestic manufacturers continue to make breakthroughs in PCIe connectors, backplane connectors, memory connectors and other fields, and some products have reached the international advanced level.In the future, with the maturity of the domestic AI chip industry, the synergy effect of the industry chain will be further enhanced.
Conclusion
The rapid development of AI servers and data centers is profoundly reshaping the technical route and market landscape of high-speed connectors.Rate increase, density increase, optical copper retreat, and domestic substitution are the four clearest trends at present.For connector companies, this is both a once-in-a-lifetime market opportunity and a comprehensive test of the strength of technology research and development.
Shenzhen Ruixin Shengye Electronic Technology Co., Ltd. closely follows the development trend of the data center and AI computing power industry, and continues to invest in research and development in the fields of high-speed connectors, precision pin hubs, FPC connectors, etc., and is committed to providing customers with high-quality, high-reliability connection solutions.
