As competition intensifies in the artificial intelligence (AI) semiconductor market, NAND flash giant Kioxia of Japan has thrown down a new challenge. While Samsung Electronics and SK hynix are leading the AI memory technology competition centered around high-bandwidth memory (HBM), Kioxia, which has lagged behind, has revealed a strategy specialized in "AI inference."
According to industry sources on the 10th, Kioxia has recently declared at a corporate briefing for global analysts that it will focus its efforts on the "inference" market, a phase of AI utilization. The company aims to capture market leadership with a strategy that encompasses hardware technology and software by prioritizing storage for AI inference.
If creating an AI model is like moving all the books in a giant library to the shelves at once, the actual service of "inference" is more akin to thousands of readers simultaneously requesting different pages immediately. For this reason, the role of high-performance storage that is fast and responsive, in addition to the computational support capabilities of HBM, is becoming prominent in the inference market. Kioxia projects that the AI inference storage market will grow at an average annual rate of 69% until 2029. This figure exceeds the overall projected growth rate for the NAND market (20%) by more than three times.
Kioxia, the fourth largest in the global NAND flash market, is concentrating on the inference market to overcome the limitations of the existing NAND market, which has become a red ocean. According to market research firm TrendForce, the revenue of the top five global NAND flash manufacturers in the first quarter this year was $12.02 billion (approximately 16.3 trillion won), a severe drop of 24% compared to the previous quarter. Due to intensified competition and decreasing demand, the average NAND selling price in the industry has fallen by 15% compared to the last quarter, and shipments have also decreased by 7%. In response, Kioxia plans to focus on territorial expansion through technological differentiation instead of a full-scale war.
The core of Kioxia's proposed strategy goes beyond just component supply; it aims to provide solutions that solve specific AI service problems. The software "AiSAQ" targets the expense issues of retrieval-augmented generation (RAG) technology, which improves the accuracy of AI responses. While RAG systems rely on expensive DRAM for speed, AiSAQ enables rapid search capabilities using large-capacity solid-state drives (SSDs).
Kioxia also introduced a card to resolve the significant bottleneck faced by graphics processing units (GPUs) in AI computations: the "super high IOPS SSD." This storage device is designed to instantaneously provide the data most needed by the GPU, which acts as the brain for AI, without delay, aiming for over 10 million IOPS (input/output operations per second). This speed is 3 to 5 times faster than currently the highest-performing SSDs for data centers, with sample releases planned for the second half of next year. Masashi Yokotsuka, Kioxia's vice president, expressed confidence, stating, "We will collaborate with the world's largest GPU manufacturer to achieve better performance in GPU systems."
Kioxia has revealed that it is developing the next-generation DRAM called "OCTRAM," expressing its ambitions to eventually enter the low-power memory market. A semiconductor industry official noted, "Competition in the NAND flash market is transcending simple price, capacity, and general performance in the AI era. Kioxia has put forth a multilayered portfolio suitable for this, but ultimately, the process of bundling all these technologies into a single solution to prove value to the market and persuade customers will be a key variable in determining future success."