Core Insights - Marvell's recent Custom AI Investor Day indicates a shift from "general GPU assembly" to "highly customized system-level collaboration" in AI chips, positioning the company as a key infrastructure partner in AI computing [1][2] - The projected total addressable market (TAM) for data centers in 2028 has been revised upward from $75 billion to $94 billion, with the custom computing (XPU) and its associated components market reaching $55 billion [1][3] Market Growth - Custom computing (XPU) and XPU Attach (supporting components) are identified as the fastest-growing sectors, with custom computing experiencing nearly 30% growth and interconnects growing approximately 37% over the past year [1][5] - The TAM for custom chips has been increased to $94 billion, with a compound annual growth rate (CAGR) of 35%, including $40 billion for custom XPU (CAGR of 47%) and $15 billion for XPU supporting components (CAGR of 90%) [3] Revenue Transition - Marvell anticipates a complete transition of cloud revenue to AI revenue, reflecting the deep integration of AI across all applications and cloud infrastructure, with cloud becoming the "factory" for AI [1][5] Customer Dynamics - Traditional cloud giants (Amazon AWS, Microsoft Azure, Google, Meta) remain primary customers, but emerging hyperscalers (e.g., xAI, Tesla) are beginning to build their own AI clusters and push for custom chips [6] - "Sovereign AI," driven by national initiatives for local AI infrastructure, is emerging as a new growth direction, prompting increased demand for custom AI chips globally [8] Project Pipeline - Marvell has secured 18 custom socket projects and is actively pursuing over 50 new project opportunities, with potential lifecycle revenue totaling $75 billion [10] - The company aims to achieve a 20% market share in the custom computing market by 2028, up from 13% currently, with previous shares below 5% [10] Technological Advancements - Marvell showcased several key technological breakthroughs, including 2nm custom SRAM with 17 times the bandwidth density of mainstream IP, self-developed HBM solutions, and advanced Die-to-Die interconnects [11][14]