Mizuho Raises Nvidia AI GPU Forecast for 2025

Mizuho’s analysts for Asia hardware and semiconductors have increased their forecast for Nvidia’s AI GPU units in 2025, expecting stronger growth due to expanded production capacity.

The analysts lifted Nvidia’s total AI GPU shipments for 2025 by 8-10% from the July 2024 estimates, highlighting improvements in the supply chain, particularly with CoWoS (Chip on Wafer on Substrate) technology.

Nvidia (NASDAQ:NVDA) is projected to ship between 6.5 to 7 million units in 2025, which will include 3 million units of CoWoS-S GPUs. The shipment will primarily consist of the Hopper and Blackwell models, with CoWoS-S achieving a yield of over 99%.

Additionally, CoWoS-L is expected to produce 3.8 to 4 million units, focusing on high-demand GB200 server racks, allocating 80-90% of that output.

This revised forecast coincides with Taiwan Semiconductor Manufacturing (NYSE:TSM) doubling its annual wafer capacity.

The analysts stated, “We expect the tightness in CoWoS supply to continue into 2025 but project gradual improvement, with an estimated 650,000 annual wafer capacity coming online at TSMC next year (up 2x YoY).”

Moreover, they mention that while some back-end processes may be outsourced to Advanced Semiconductor Engineering (ASE), the supply-demand situation should gradually enhance.

Concerns about delays in the Blackwell GPU rollout have been termed as “exaggerated” by analysts, noting both TSMC and SK Hynix Inc (KS:000660) are positioned for sales growth with the Blackwell GPU launch in Q4 2024.

Mizuho’s analysts do not foresee semiconductor delays affecting AI server production in 2025 and predict Taiwanese back-end equipment vendors like Allring Tech will benefit from TSMC’s localization efforts.