BlockBeats News, July 28th, Citrini analyst Jukan quoted Fubon Securities' "2027 Semiconductor Outlook" report, pointing out that although the cost of HBM4 is expected to increase significantly from HBM3e's $17-18/GB to $31-32/GB by 2026, raising the price of NVIDIA's Rubin GPU to around $78,000 to $80,000, NVIDIA is still able to maintain a high gross margin level of 75% to 80%, and its pricing power and cost pass-through ability remain unshaken. The report also mentioned that the cost of custom ASICs with HBM may be even higher, reaching $35-36/GB, indicating that the cost of HBM will double by 2027.

The Fubon Securities report remains optimistic about the AI market demand, believing that despite recent market concerns about AI inflation, Token costs are still a more critical factor driving capital expenditure for cloud service providers. At the architecture level, NVIDIA's new rack still plans to use the same number of compute chips, with Scale-up done within the rack using cables and Scale-out between racks using NPO/CPO. Additionally, Google plans to deploy 12 to 15 million TPUs by 2028, with capacity consumption more than doubling compared to 2027; Intel's EMIB capacity is expected to increase to 24,000 to 25,000 wafers per month by the end of 2027, while TSMC will slow down its SoIC expansion to prioritize expanding CoWoS capacity.

The key conclusion is that, despite changes in the cost structure, NVIDIA's pricing power, gross margin, and demand support in the AI computing ecosystem have still built formidable competitive barriers.