NVIDIA 2025 GPU Unit Forecast Increased by Mizuho

Mizuho’s analysts specializing in Asia’s hardware and semiconductor sectors have upgraded their 2025 forecast for Nvidia’s AI GPU units, predicting stronger-than-anticipated growth due to increased production capacity.

The analysts modified Nvidia’s total AI GPU shipments for 2025 to reflect an increase of 8-10% compared to their July 2024 estimates, attributing this adjustment to significant improvements in the supply chain, particularly regarding CoWoS (Chip on Wafer on Substrate) technology.

The report indicates that Nvidia is expected to ship between 6.5 and 7 million units in 2025, which will include 3 million units of CoWoS-S GPUs. The majority of these shipments will consist of the Hopper and Blackwell models, with CoWoS-S achieving a yield rate exceeding 99%.

Additionally, the CoWoS-L process is projected to produce between 3.8 and 4 million units, primarily targeting the in-demand GB200 server racks, which are expected to account for 80-90% of the output.

This updated forecast aligns with the fact that Taiwan Semiconductor Manufacturing Company, Nvidia’s main production partner, is set to double its annual wafer capacity.

The analysts noted, “We anticipate that the current tightness in CoWoS supply will persist into 2025. However, we expect a gradual improvement, with annual wafer capacity projected to exceed 650,000 at TSMC next year, marking a twofold increase year-over-year.”

They also mentioned that, in addition to ramping up front-end (chip-on-wafer) capacity, certain back-end (on-substrate) processes will likely be outsourced to Advanced Semiconductor Engineering, contributing to a better supply-demand balance.

The analysts expressed that investors’ worries about potential delays in the Blackwell GPU launch have been “exaggerated,” highlighting that both TSMC and SK Hynix Inc. remain on track for sales growth following the anticipated Blackwell GPU release in the fourth quarter of 2024.

Overall, Mizuho’s analysts do not foresee any semiconductor-related delays, including those affecting AI GPUs and ASICs, to impact AI server production in 2025. They also noted that Taiwanese back-end equipment vendors, such as Allring Tech, are expected to gain from TSMC’s strategies to localize its vendor supply chain.