With AMD expected to hold "AMD Advancing AI 2026" in San Francisco on the 22nd (local time) to announce the official launch of its next-generation artificial intelligence (AI) accelerator "Instinct MI450" and rack-scale system "Helios," attention in the semiconductor industry is focusing on the matching memory specifications from Samsung Electronics and SK hynix.
As the sixth-generation high bandwidth memory (HBM4) capacity required by AMD's MI450 is expected to surpass Nvidia's next-generation product, some expect AMD's share in the supply chains of Samsung Electronics and SK hynix to grow. It is also possible the HBM customer base, which had been concentrated on Nvidia, could be reorganized around two pillars.
◇ AMD's next-generation AI accelerator puts Nvidia on edge
According to industry sources on the 21st, AMD is expected to announce detailed specifications and a release schedule for the MI450, along with an HBM4 sourcing plan, at the event. The most eye-catching part is the memory capacity. The MI450 is said to equip 432 gigabytes (GB) of HBM4 per graphics processing unit (GPU). That is about 150 GB more than Nvidia's next-generation GPU "Rubin (288 GB)." The biggest variable in how many AI models a single GPU can run is memory capacity. That is why the center of gravity in recent AI Semiconductor competition is shifting from compute speed to memory capacity.
As AMD sharply increased HBM4 capacity on its next-generation AI accelerator to boost performance, Nvidia is also said to have moved to respond. Taiwan market research firm Fubon Research said last year that Nvidia, mindful of AMD's MI450, was reworking the Rubin design, which could delay its release schedule. Nvidia rebutted that by officially announcing at Computex 2026 last month that Rubin had entered full-scale mass production, but there is also speculation that Rubin output this year will fall short of initial expectations due to HBM4 supply bottlenecks.
AMD's confidence is also evident in customers it has already secured. It is believed to have lined up large customers including OpenAI, Meta, and Oracle. Oracle is reportedly planning to deploy 50,000 MI450 units to its cloud starting in the third quarter of this year. The contract sizes between AMD and OpenAI and between AMD and Meta are each said to be in the $100 billion range. It means a practical alternative is emerging in the AI chip procurement structure of hyperscalers (large-scale data center operators), which had been dominated by Nvidia.
◇ Samsung Electronics and SK hynix, dual supply chains that secure both sides
The demand expansion is expected to benefit Korea's memory sector. Samsung Electronics has secured the position as the priority HBM4 supplier for AMD's next-generation accelerator. At the same time, it passed the final quality verification for HBM4 for Nvidia Rubin and is already mass-producing it. As for production yield, it is said to have improved from around 50% in the fourth quarter of last year to the 60%–70% range in the second half of this year.
SK hynix is known to have a high supply share for Nvidia Rubin, while the possibility of expanding supply to AMD is also said to be open. Following Nvidia, AMD's rise could, in effect, allow both companies to reduce their dependence on Nvidia. If AMD also unveils the roadmap for its next product, "MI500," at this event, attention will turn to how the adoption timing of HBM4E (seventh-generation HBM) aligns with the mass production schedules of the two Korean memory companies.
A semiconductor industry official said, "Because the structure improves price negotiating power as the number of customers increases, AMD's rise clearly provides a new growth axis for the memory industry," and added, "In the end, the very dynamic of the two companies competing and increasing volumes is a favorable environment for suppliers Samsung Electronics and SK hynix."