The influence of Nvidia GPUs in the entire cycle of artificial intelligence (AI) learning is absolute, accounting for 86% of the data center AI semiconductor market as of the end of last year. However, as AI's center of gravity shifts from learning to reasoning starting this year, companies are seeking specialized semiconductors that infer faster and cheaper. Custom AI chip shipments, called ASIC, are expected to increase to 44.6% this year, three times higher than the growth rate of commercial GPUs. OpenAI joined with Broadcom to introduce its own AI semiconductor "Hallapeño" specialized in AI reasoning last month, designed to run ChatGPT and coding agent Codex. Google uses its own AI semiconductor 'Ironwood', Amazon Web Service developed 'Trainium', and Microsoft developed 'Mya 200'. Deloitte predicted that reasoning will account for two-thirds of AI computing resources worldwide this year. Cerebras, which uses an entire round semiconductor wafer as an AI accelerator, is supplying semiconductors to OpenAI and has joined hands with Korea's AI unicorn Upstage. Korean AI semiconductor startups such as Revelion and Puriosa AI are differentiating as general-purpose high-efficiency neural network processing units for data centers.