AMD is scheduled to hold the AMD Advancing AI 2026 conference in San Francisco on July 22, releasing its next-generation AI accelerator Instinct MI450. Each GPU comes with 432GB of HBM4 memory—about 150GB more than NVIDIA’s next-gen GPU “Rubin”—and it has become a key backdrop for this wave of expanded HBM4 orders from Samsung Electronics and SK hynix.
AMD Advancing AI conference: MI450 432GB HBM4 specs released
Based on disclosures from industry insiders on July 21, 2026, AMD is expected to unveil MI450’s detailed specifications and release schedule at the “AMD Advancing AI 2026” conference, along with its HBM4 procurement plan. The most closely watched part is memory capacity: each MI450 GPU is equipped with 432GB of HBM4, which is about 150GB higher than the 288GB of NVIDIA’s next-gen GPU “Rubin.”
Industry sources say the competitive focus in AI semiconductors has recently shifted from compute speed to memory capacity, and a single GPU’s memory capacity directly determines the size of AI model workloads it can run at the same time. NVIDIA officially announced at Computex 2026 (Taipei International Computer Exhibition) that its Rubin graphics cards have entered full-scale mass production.
Oracle, OpenAI, and Meta: AMD MI450 reported customer roster
According to industry reports, AMD MI450 has established the following customer relationships:
Oracle: Plans to deploy 50,000 MI450 chips on its cloud platform starting in Q3 this year
OpenAI: The contract value between AMD and OpenAI is reportedly about $10 billion
Meta: The contract value between AMD and Meta is reportedly about $10 billion
A semiconductor industry official said: “As the number of customers increases, AMD’s negotiating leverage on pricing obviously improves, and AMD’s rise is clearly providing the memory industry with a new growth axis. The dynamic of these two companies competing while continuously expanding scale is creating a favorable environment for memory suppliers Samsung Electronics and SK hynix.”
Samsung Electronics HBM4 supply: AMD prioritized, yields improve to 60%-70%
According to industry news, Samsung Electronics has secured its position as AMD’s preferred supplier of HBM4. At the same time, its HBM4 supply for NVIDIA’s Rubin has passed final quality verification and has started mass production; yields have improved from about 50% in last year’s Q4 to 60%-70% in the second half of this year.
SK hynix is currently the main supplier for NVIDIA’s Rubin GPUs, and the likelihood of supplying AMD is also beginning to take shape. If AMD also releases the roadmap for its next-generation product “MI500” during the event, market attention will shift to the rollout timeline of HBM4E (7th-generation HBM) and whether the mass-production plans of the two Korean memory makers can align.
FAQ
How much HBM4 memory does AMD MI450 have, and how does it compare with NVIDIA Rubin?
According to disclosures from industry insiders on July 21, 2026, each MI450 GPU is equipped with 432GB of HBM4 memory—about 150GB more than NVIDIA’s next-gen GPU Rubin’s 288GB. The specific specifications will be subject to AMD’s official July 22 conference announcement.
What is Samsung Electronics’ position in AMD’s HBM4 supply chain?
According to industry news, Samsung Electronics has secured its position as AMD’s preferred HBM4 supplier. At the same time, its HBM4 supply for NVIDIA’s Rubin has passed final quality verification and is being mass-produced, with yields improving from about 50% in last year’s Q4 to 60%-70% in the second half of this year.
Who are the main customers of AMD MI450?
According to industry reports, Oracle plans to deploy 50,000 MI450 chips on its cloud platform starting in Q3 this year. The contract values between AMD and OpenAI and between AMD and Meta are reportedly about $10 billion each. Specific contract details are subject to official announcements.
