When discussing AI computing power, attention invariably centers on GPUs, large language models and optical modules. Few people notice that the core foundation keeping ten-thousand-GPU AI clusters running stably lies in seemingly trivial specialty wires and cables.

Without high-end cables to deliver power and transmit data, even the priciest AI servers become nothing more than non-functional hardware. As demand for AI computing surges, conventional low-end cables have been phased out. Specialty cables designed exclusively for AI servers are emerging as a new industry hotspot.

01 Why Ordinary Cables Cannot Support AI Servers

Traditional IDC racks only operate at 5–8 kW, with simple wiring and light load. New-generation liquid-cooled AI racks, by contrast, ramp up to 60–120 kW, and a single GPU can draw more than one kilowatt of power.

Extreme power consumption creates three major challenges: high temperature, heavy current and high-density cabling. Ordinary cables heat up rapidly, suffering insulation degradation and severe signal crosstalk, which can easily trigger short circuits and server outages. For AI clusters costing tens of millions, computing losses caused by a single cable fault come at an astronomical price.

This separates AI server cables fundamentally from regular data center cables, turning them into high-threshold specialty products.

02 Two Core Cable Categories Underpinning the AI Computing Ecosystem

AI server cables fall into two main types: power supply cables and high-speed interconnect cables. Each fulfills distinct roles, and neither can be dispensed with.

High-temperature Specialty Power Cables

Most intelligent computing centers adopt an 800 V high-voltage DC power supply architecture, equipped with low-smoke zero-halogen flame-retardant cables. These cables feature high temperature resistance, flame retardancy and low-smoke non-toxic properties, suited for high-density, high-heat environments and mitigating fire risks in data centers.

Around GPUs inside racks, high-temperature silicone harnesses rated for 150–200°C are widely deployed. They withstand repeated thermal cycling and bending, resolving wiring and material aging issues in confined spaces to guarantee stable power delivery for AI servers.

High-speed Interconnect Cables

Large model training requires tens of thousands of GPUs to perform parallel computing in real time, imposing stringent requirements on transmission latency and reliability. The industry has formed the standard strategy: copper for short reach, fiber for long reach.

For short-distance interconnections within racks, high-speed DAC/AEC copper cables are the primary solution. Boasting ultra-low latency, low power consumption and easy pluggability, they are compatible with NVLink high-speed interconnection and support ultra-high-speed transmission at 800G and 1.6T, with consumption more than ten times that of traditional servers.

Long-haul communications across racks or campuses rely on OM5 multimode fiber and ultra-low-loss single-mode fiber. Cutting-edge hollow-core fiber can reduce transmission latency by 30%, making it a key solution for future supercomputing and cross-city computing orchestration.