Amid persistently strong demand for AI chips, NVIDIA's (NVDA.US) key production capacity remains secure, though the rollout schedule for its next-generation GPU products may be affected by memory technology constraints. Following a review of the Asian supply chain, KeyBanc Capital Markets issued a report stating that NVIDIA has largely secured its supply of advanced packaging technology, CoWoS. CoWoS capacity is projected to reach 650,000 units in 2026 (a 76% year-over-year increase) and further climb to 840,000 units in 2027 (a 29% increase), indicating the company's continued long-term advantage in AI chip production capacity. From a product structure perspective, the firm estimates that the 2026 CoWoS capacity can support approximately 5.5 to 6 million Blackwell GPUs, 1.5 million Rubin GPUs, and 1 million Hopper GPUs, suggesting overall supply can still meet the market's high demand for AI computing power. However, analyst John Vinh noted that the mass production timeline for NVIDIA's next-generation Rubin GPU might experience a slight delay, primarily due to the verification progress of next-generation high-bandwidth memory (HBM4). As HBM4 is still in a critical certification phase, the production target for the Rubin GPU has been revised down from a previous expectation of 2 million units to 1.5 million units. Consequently, KeyBanc has also lowered its shipment forecast for NVIDIA's AI server systems (Vera Rubin racks) from a prior range of 12,000 to 14,000 units to approximately 6,000 units, reflecting a potential short-term slowdown in the deployment pace of high-end AI hardware. Despite this, Vinh characterized the impact as "relatively limited" and affirmed that it does not alter the firm's overall positive view on NVIDIA. KeyBanc maintains an "Overweight" rating and a $275 price target for the stock.
Comments