Sinolink Securities Co., Ltd. has released a research report indicating that as GPU power consumption rises, switching chip counts increase, and supernode architectures advance, rack power supply capacity will become one of the key bottlenecks in AI infrastructure. Liquid-cooled busbars are expected to emerge as a significant product trend for high-power AI racks, with 800V high-voltage power supply further raising performance requirements for busbars. The increasing penetration of liquid cooling in rack busbars, along with the adoption of 800V HVDC high-voltage direct current architectures in new data centers, will create material and structural upgrade opportunities for copper busbars, a core component of power distribution systems.
The report outlines its primary viewpoints, starting with the observation that AI servers are transitioning from single-machine deployment to rack-level and supernode-level systems, entering a phase of rapid rack power consumption growth. In the AI era, GPUs/ASICs, NVSwitch, high-speed interconnects, optical modules, and other components are being integrated at higher densities within the same rack, making the entire rack the smallest unit for collaborative design of computing, power supply, cooling, and interconnection. Taking NVIDIA's platform as an example, the GB200 NVL72 has an IT-side power draw of approximately 136kW, while the Vera Rubin NVL72 is expected to increase to around 209kW, with NVLink switching capabilities and overall rack power density continuing to rise.
As GPU power consumption increases, the number of switching chips grows, and supernode architectures advance, rack power supply capacity will become a critical bottleneck in AI infrastructure. Rack busbars are the core component of the internal power distribution system within AI racks, primarily responsible for high-current transmission, power distribution, and tray connections, serving as the "power backbone" inside the rack. In a centralized power supply architecture, after the power module outputs electrical energy, the busbar must distribute current to load terminals such as compute trays and switch trays. Therefore, the conductivity, temperature rise control, voltage drop, connection reliability, and safety redundancy of the busbar directly impact the overall power supply efficiency and operational stability of the rack.
Using NVIDIA racks as an example, the GB200 NVL72 has a busbar current of approximately 2,900A under 50V centralized power supply, while the Vera Rubin NVL72's busbar current is expected to increase to over 5,000A, significantly exceeding the capability boundaries of traditional rack power distribution components. The challenge posed by increased power is not just "higher current," but also the need to achieve controlled transmission of high current within the limited space of the rack. TE Connectivity's simulation data shows that under a 48V architecture, when load power increases from 200kW to 400kW, the maximum DC current rises from 4,166A to 8,333A, the maximum temperature rise increases from 7.79°C to 33.59°C, and the maximum voltage drop increases from 0.10V to 0.21V, with temperature rise pressure showing significant non-linear amplification.
At the same time, the internal space of AI racks is shared by chip cooling components, cables, and structural parts, making it difficult for busbars to solve high-current problems simply by widening or thickening copper bars. In other words, after AI rack power consumption increases, the busbar is no longer an ordinary power distribution component but a critical foundational component affecting the rack's power density, thermal management efficiency, and system reliability. Liquid-cooled busbars are expected to become an important product trend for high-power AI racks, and 800V high-voltage power supply will further raise busbar performance requirements.
Compared to traditional air-cooled or naturally cooled busbars, liquid-cooled busbars introduce cooling channels into the busbar structure or coordinate with the rack's liquid cooling system, integrating conductivity, heat dissipation, fluid sealing, insulation protection, and reliability testing into a single component. This upgrades the product barrier from traditional copper bar processing to integrated "electrical-thermal-fluid" design and manufacturing capabilities. TE has already launched liquid-cooled vertical busbar solutions for 200kW, 400kW, and 750kW levels, while Molex showcased a multi-channel liquid-cooled busbar solution at Computex 2026, which can improve cooling efficiency by up to 20% and control temperature rise to around 15°C at 15,000A current. Liquid-cooled busbars have entered the mass production application stage.
Furthermore, the introduction of 800V high-voltage DC architectures will further raise performance requirements for busbars in terms of insulation, safety protection, connection reliability, and structural integration, driving rack busbars to continue upgrading from low-voltage, high-current components to high-voltage, high-reliability assemblies. Copper is the core material for rack busbars due to its lower resistivity, higher conductivity, and higher thermal conductivity, which are more conducive to reducing line losses, controlling temperature rise, and compressing structural dimensions in high-current, high-power-density scenarios.
Liquid-cooled busbars and high-voltage busbars also place higher demands on copper materials. The competition focus will extend from basic copper bar supply to high-purity copper/copper alloy materials, complex cross-section forming, cooling channel coordination, welding and sealing, insulation coating, flatness control, and batch consistency management. Therefore, the upgrade of rack busbars essentially brings a comprehensive upgrade of materials, structures, processes, and verification systems. Suppliers with capabilities in high-performance copper materials, precision machining, collaborative customer development, and batch quality control are expected to benefit first during the AI rack power upgrade cycle. The report also notes risks including slower-than-expected industrialization progress and delays in product certification and capacity release.