CICC: NVIDIA Rubin CPX adopts an innovative decoupled inference architecture, which may drive the iteration and upgrade of the PCB market

According to the Zhitong Finance APP, China International Capital Corporation (CICC) released a research report stating that Rubin CPX is a GPU released by NVIDIA (NVDA.US) specifically designed for handling ultra-long context AI inference tasks, featuring an innovative decoupled inference architecture. Rubin CPX brings significant changes on the hardware side, including a wireless cable architecture for connectors/PCBs, along with the addition of the Paladin B2B connector connected to the mid-plane PCB located in the middle of the chassis. Rubin CPX introduces significant innovations in PCB, connectors, and cooling architectures, which is expected to greatly increase its hardware market size. It is estimated that by 2027, the market size of NVIDIA's AI PCB is expected to reach USD 6.96 billion, a 142% increase from 2026.

CICC's main points are as follows:

Rubin CPX achieves a rebalancing of efficiency and cost through hardware-level optimization

Rubin CPX is a GPU released by NVIDIA specifically designed for handling ultra-long context AI inference tasks, featuring an innovative decoupled inference architecture. In terms of computing performance, it provides 30 Peta FLOPS of computing performance at NV FP4 precision; in terms of storage performance, it is equipped with 128GB GDDR7 memory, with a memory bandwidth of 2TB/s; in terms of process and packaging, the Rubin CPX chip adopts a traditional FC-BGA flip-chip packaging single-chip design.

Rubin CPX brings significant changes on the hardware side

In terms of tray architecture, the front end of the tray adopts a modular design, with four sub-cards distributed on both sides of the tray, consisting of 2 Rubin CPX GPX chips and 2 CX-9 network cards; in terms of cooling, the front-end cooling will be upgraded from air cooling to liquid cooling, with the addition of a liquid cooling plate inside the tray, where heat pipes and heat spreaders will conduct heat from the back of each GDDR7 memory module to the cooling plate; in terms of connectors/PCBs, a wireless cable architecture is adopted, along with the addition of the Paladin B2B connector connected to the mid-plane PCB located in the middle of the chassis.

Rubin CPX drives further enhancement of single PCB value

It is estimated that the value of a single cabinet PCB for VR200NVL144 is approximately RMB 456,000, with the PCB value corresponding to a single GPU being RMB 6,333 (USD 880), an increase of 113% compared to GB300. It is also estimated that by 2027, the shipment volumes for GB300 NVL72/VR200 NVL144/VR300 NVL576 will be 10,000/70,000/20,000 racks respectively, totaling 100,000 racks, corresponding to a PCB market size of USD 6.96 billion, a 142% increase from 2026.

Target aspects

Relevant targets in the industrial chain: Shengyi Technology (600183.SH), Shenzhen South Circuit (002916.SZ), Xingsen Technology (002436.SZ), Pengding Holdings (002938.SZ), Dongshan Precision (002384.SZ), Shenghong Technology (300476.SZ), Huadian Co., Ltd. (002463.SZ), Shengyi Electronics (688183.SH), Jingwang Electronics (603228.SH), Founder Technology (600601.SH), Guanghe Technology (300964.SZ), etc Risk Factors

The progress of Rubin CPX is not as expected; the demand for AI and computing infrastructure is below expectations