闪迪：AI推理改写NAND供需 长期订单延伸至2030年

According to Kuai Technology News on November 1, SanDisk stated in the first fiscal quarter earnings conference call of fiscal year 2027 that artificial intelligence reasoning is changing the supply and demand pattern of the NAND flash memory market. Company management said that customers generally underestimate the NAND capacity required to run artificial intelligence infrastructure and are currently adding long-term orders. Its new business model (NBM) order pipeline has extended to 2030 or even beyond. SanDisk disclosed that approximately 50% of bit shipments in fiscal year 2027 have been locked by the NBM protocol, and this proportion is expected to rise to approximately two-thirds in fiscal year 2028. SanDisk Chairman and CEO David Goeckeler said that the company not only saw additional orders from existing customers, but also welcomed new customers; customer feedback showed that they underestimated the NAND flash memory capacity required to run the infrastructure. Management anticipates that reserve demand from NBM projects may exceed the company's available supply by the end of the decade, meaning demand for flash memory for AI infrastructure is shifting from short-term purchases to longer-term capacity and supply arrangements. Data centers have become the core of NAND demand growth. Goeckeler said that the data center business has grown from about 20% of the total industry ten years ago to nearly half now; the company expects data centers to account for more than 55% of the relevant market in 2027, with demand increasing by about 40% year-on-year. He further predicted that by the end of this decade, key-value cache (KV Cache) may account for about one-third of the AI ​​data center flash memory market, and the AI ​​data center market size will be close to the current entire NAND flash memory market. As the model scale expands and the context length increases, the system needs to store more KV caches, and the requirements for capacity, read bandwidth, and cost increase accordingly. NAND no longer only performs traditional data storage functions, but may also become an important capacity layer for AI inference systems. In terms of technology, SanDisk announced that its first high-bandwidth flash (HBF) storage die has completed tape-out and entered the manufacturing stage, with the first batch of product samples scheduled to be delivered in the next calendar year. HBF is mainly oriented to AI inference scenarios, and its goal is to provide read bandwidth close to high-bandwidth memory (HBM).