NVIDIA Sets 2025 AI Superchip Goal: 70% Share for GB200

According to recent reports from reputable sources, NVIDIA's latest AI chip, the GB200, has garnered attention despite falling short of its predecessor. A foreign investment bank has forecasted NVIDIA's AI superchip shipment goal for the upcoming year to be between 6.0 to 6.5 million units, with the GB200 expected to claim a 70% share of these shipments, leaving the remaining 30% for the GH200.

The GB200, comprising two B200 GPUs and one Grace CPU chip, introduces water-cooling technology to manage its high power consumption, ranging from 1,000 to 1,200 watts for one B200 chip. This innovation has become a focal point in the market. The latest report from the foreign investment bank also delves into the supply chain analysis of NVIDIA's GH/GB200 superchips.

Analysis suggests that NVIDIA's GB200 servers will primarily consist of two categories: NVL36 and NVL72. NVL36 is anticipated to dominate shipments, with rack prices ranging from $1.5 to $2 million, while NVL72's price is estimated to be 1.8 times higher.

In terms of cooling solutions, each GPU may require approximately $300 for cooling, a significant reduction compared to market estimates of $2,000. However, market rumors suggest that the overall cooling solution for GB200 may cost between $40,000 to $80,000, with at least four suppliers per component.

The report highlights NVIDIA's self-developed Grace CPU, based on the Arm architecture, which has received positive responses from Cloud Service Providers (CSPs) since its introduction at the NVIDIA GTC 2024 conference. From a cost-performance perspective, every dollar spent yields better computing power. Additionally, NVIDIA will offer x86 architecture products.

Regarding packaging, TSMC's CoWoS remains predominant, but NVIDIA will introduce Amkor and Intel as advanced packaging suppliers, each accounting for half of the share alongside TSMC. NVIDIA's shipment target for next year is 6.0 to 6.5 million units, with GB200 accounting for 70%.

In terms of server assembly, the report suggests that Hon Hai may hold the largest share, with Wistron categorized as Tier 1 and Quanta as Tier 2. However, CSPs will not abandon Application-Specific Integrated Circuits (ASICs), and as time passes without significant performance improvements, the quantity of ASICs may decrease.

The report emphasizes that ASICs are currently primarily focused on Amazon AWS and Google, with poor performance in other areas. However, recent indications suggest a gradual recovery in general server orders, with CSP customers potentially reducing purchases of the previous generation of AI servers, awaiting GB200 shipments. Therefore, this year's budget may be allocated to general server purchases, indicating potential downward risks for existing Original Design Manufacturers (ODMs) and ASIC AI supply chains.