SemiAnalysis founder Dylan Patel predicts that by 2030, the computing demand for AI inference from just OpenAI and Anthropic could exceed 100GW. Adding in major AI companies like Meta and Google, the scale of inference computing will further expand, and by 2040, AI inference infrastructure could reach the terawatt (TW) level. Unlike the training phase, which primarily relies on GPU computing power, the inference phase—especially the Decode process—depends more on memory bandwidth and capacity. GPUs need to constantly read data from memory, making HBM a critical component in unlocking AI computing power. Currently, the HBM industry is rapidly expanding: HBM4 has entered the high-volume shipment phase, with the ramp-up speed of HBM4 12-high being about twice that of HBM3E 12-high. HBM4 revenue has already surpassed $1 billion. At the same time, the expansion of AI servers is driving growth in data center DRAM and NAND demand. By 2026, the bit shipments of data center DRAM/NAND in the industry are expected to more than double compared to two years ago.