A quiet revolution is underway in the global AI computing sector, as top AI model companies—OpenAI, Meta, DeepSeek, and Zhipu AI—extend into chip design, mounting a challenge to Nvidia's GPU dominance. On July 9, an internal Meta memo disclosed by Reuters revealed plans to begin mass production of its fourth-generation custom AI chip, codenamed "Iris," as early as September this year, developed with Broadcom and manufactured by TSMC. Meta plans to deploy approximately 7 gigawatts of AI computing infrastructure in 2026, doubling to 14 gigawatts in 2027. Chinese AI firm DeepSeek was revealed to have secretly initiated a project to develop its own AI inference chip about a year ago, engaging with chip design firms, wafer foundries, and memory manufacturers. Zhipu AI is also weighing the design of its own AI chips, making preliminary inquiries to Chinese chip design firms. In June this year, OpenAI launched "Jalapeño," its first custom inference chip co-developed with Broadcom, with performance comparable to Nvidia's Blackwell chip and Google's TPU. OpenAI President Brockman stated: "The world is moving toward a compute-driven economy." Forrester analyst Mike Gualtieri noted: "If you rely on someone else's chips, you can never become a true AI giant."