HBM Price is coming? Samsung disputes supply NVIDIA GB300
Korean media reported that Samsung Electronics is inquiring with NVIDIA (Nvidia Corp.) about the 12-layer stacking "HBM3E" to supply the GB300 Blackwell Ultra.
BusinessKorea quoted Seoul Economic Daily on the 1st that Jun Young-hyun, the responsible person of Samsung Device Solutions (DS) department, visited NVIDIA's headquarters in Silicon Valley last week to discuss the supply of HBM3E, less than two months from the visiting trip in early May.
Although it is currently impossible to confirm whether NVIDIA executive president Jensen Huang attended the meeting, the two sides discussed the possibility of 12-layer HBM3E quality verification and the possibility of starting supply in 2026.
A well-known insider claims that Samsung strongly states that its 12-layer HBM3E is based on the fourth generation of 10-nanometer DRAM (1a), combined with improved substrate and logical grains, showing that it is completely uncompetitive. Samsung is discussing the possibility of supplying Blackwell Ultra. Blackwell Ultra is scheduled to ship in large quantities next year.
According to reports, Samsung believes that the results of this partnership are not only because the product quality does not increase others, but the company has also obtained customers including Ultramicro (AMD).
Samsung recently announced that AMD's MI350X series AI accelerator will supply 12-layer HBM3E. As the performance performance of the MI350X series is better than expected, the market's expectations for AMD are increasing day by day, and the belt also dispels doubts about Samsung's 12-layer HBM3E. As a result, Samsung is becoming more and more convinced that it is only too early to pass NVIDIA quality verification and will have a large supply next year.
HBM is still in short supply. If Samsung becomes the third 12-layer HBM3E supplier, NVIDIA will definitely obtain price judgment advantages for SK Hynix and Micron. An industry insider said that as we all know, SK Hynix's 12-layer HBM3E supply to NVIDIA is about 60% more expensive than the 8-layer version; from the perspective of NVIDIA, if Samsung's HBM quality is verified, there is no reason not to place orders, because this can encourage suppliers to compete with each other.
According to reports, NVIDIA has repeatedly delayed the use of HBM4 timetable for the same reason. HBM4 is mainly used for the next-generation AI chip "Vera Rubin" scheduled to be shipped by the end of next year. SK Hynix has delivered the HBM4 sample in March this year, and Micron has also been sending the sample in June, while Samsung's delivery time is expected to be July or August. NVIDIA will make a decision after Samsung's sample is delivered.
According to industry evaluation, since Samsung appointed Quan Yonghui as the responsible person of the DS department last year, memory competition is rapidly recovering. Quanyong Ridge said at the Stock Conference in March this year that Samsung is working hard to improve product competition and is expected to gain market leading position in the second quarter of this year as soon as possible and at the second half of this year.
