NVIDIA GB200: Millions Shipped by 2025? Projections Soar

In March of this year, at the GTC 2024 conference held at the San Jose Convention Center in California, NVIDIA’s founder and CEO, Jensen Huang, introduced the Blackwell architecture GPU. This new data center product has once again cemented NVIDIA’s dominant position in the artificial intelligence market, heralding a new era in AI computing.

The Blackwell platform includes the B200 GPU, designed to replace the H100/H200, and the GB200, which is paired with the Grace CPU. Additionally, at the end of last year, NVIDIA launched the H200 and GH200 product lines based on the existing Hopper architecture as an interim solution before the introduction of Blackwell products, with the GH200 accounting for about 5% of NVIDIA’s high-end GPU shipments. Although there is still some time before these products are shipped, the supply chain has high hopes for the GB200, anticipating shipment volumes in the millions by 2025, which would represent 40% to 50% of NVIDIA’s high-end GPU shipments.

According to TrendForce, the Blackwell platform’s products will employ more complex and precise CoWoS-L packaging technology, requiring more extensive validation and consequently delaying initial shipments of the B100, B200, and GB200 to the fourth quarter of this year, with volume shipments not expected until at least the first quarter of 2025.

The B100, B200, and GB200 products will also consume more CoWoS capacity, compelling TSMC (Taiwan Semiconductor Manufacturing Company) to significantly boost its packaging capacity in 2024. By the end of the year, monthly capacity is projected to reach 40,000 wafers, an increase of at least 150% from 2023. Moreover, TSMC is already planning to potentially double its CoWoS capacity in 2025, with NVIDIA’s requirements constituting more than half of this demand.

Earlier this year, reports indicated that NVIDIA had turned to Intel for packaging services. However, Intel’s technology primarily revolves around CoWoS-S packaging, which can only meet the requirements of NVIDIA’s H series products. Given the limited technological advancements in the short term, Intel’s capacity expansion plans remain conservative, unless they secure additional orders in the future.