Industry sources indicate that thermal issues associated with GB200 have prompted major clients including Microsoft, Amazon, Google, and Meta to reduce their orders, impacting not only Nvidia's product sales and revenue but also creating significant implications for its primary HBM supplier, SK Hynix.
The GB200, which combines one Grace CPU and two Blackwell GPUs, delivers a remarkable 30-fold improvement in inference model performance compared to the H100 while cutting costs and energy consumption by 1/25th.
Several companies had already placed orders exceeding US$10 billion for GB200. HBM is a critical component in AI chips, and for the GB200 NVL72 platform, which integrates 576 HBM3E products, SK Hynix serves as the main supplier. With Micron maintaining limited supply capacity and Samsung Electronics still pending Nvidia's certification, SK Hynix's role remains pivotal.
In its third quarter of 2024 report, SK Hynix estimated that its sales to a presumed single client believed to be Nvidia, would surpass KRW6 trillion (US$4.1 billion), accounting for more than 34% of its total revenue of KRW17.5 trillion.
Initially, analysts had predicted that with the ramp-up of Nvidia's Blackwell production, HBM3E would make up about 55% of SK Hynix's total HBM shipments by 2025, with this share rising to 84% by 2026. However, the thermal challenges with the GB200 identified earlier in 2024 have led to reduced order volumes, potentially delaying the expansion of production timelines.
To accommodate adjustments for the GB200, SK Hynix has accelerated its mass production schedule for its 12-layer HBM3E to the fourth quarter of 2024. However, starting production without sufficient demand could force adjustments to SK Hynix's production schedules and inventory, potentially affecting its results for the first quarter of 2025.
Industry analysts suggest that the delay in the large-scale production of Blackwell chips could benefit Samsung, which has yet to pass Nvidia's certification. This delay provides Samsung with additional time to refine its product performance. Meanwhile, Nvidia may not feel the urgency to expedite Samsung's certification process due to the reduced demand from major tech companies. This scenario could lead Nvidia to impose stricter quality standards on Samsung than it applied to SK Hynix.
Nvidia CEO Jensen Huang had previously stated that Samsung's HBM3E required "redesign," a remark that some analysts interpreted as an attempt to prompt Samsung to propose solutions for the thermal issues. Regardless of the reasoning behind this, the delayed certification process means that Samsung will have limited opportunity to capture the AI memory market soon.
Concerns are mounting as reports suggest that the next-generation AI semiconductor, "Rubin," which was initially slated for release in the second half of 2025, could face similar thermal challenges. Such delays would further extend the timeline for mass production. Rubin is expected to rely on HBM4, the next battlefield in the memory industry.
SK Hynix plans to begin mass production of HBM4 by the end of 2025, viewing it as a critical opportunity for growth. For Samsung, which has faced setbacks with both HBM3 and HBM3E, HBM4 represents a potential recovery. However, if Nvidia fails to timely address the thermal issues with its next-generation chips, coupled with the ongoing downturn in the general memory market, the memory industry may face significant uncertainty heading into 2025 and beyond.
Article translated by Jingyue Hsiao
