Custom AI Chips Poised for Growth as Market Shifts Toward Inference

The artificial intelligence semiconductor landscape is entering a new phase as the industry pivots from the training of large-scale models to the deployment of real-time inference. While graphics processing units (GPUs) dominated the initial surge of AI hardware demand, the market is increasingly turning toward application-specific integrated circuits (ASICs) for operational efficiency. Broadcom (AVGO) and Marvell Technology (MRVL) are emerging as primary beneficiaries of this transition. Unlike general-purpose GPUs, ASICs are custom-designed for specific tasks, making them more energy-efficient and cost-effective for inference workloads—the process where trained models generate responses for end-users. Data from Goldman Sachs suggests a significant shift in data center architecture. The investment bank estimates that unit demand for ASICs will reach parity with GPUs by 2027, resulting in a 50-50 split in AI server installations. This marks a sharp contrast to the previous year, where GPUs held an estimated 62% market share. This trend is driven by hyperscalers including Alphabet, Microsoft, and Amazon, who are co-designing custom silicon with Broadcom and Marvell to reduce their dependence on Nvidia. By utilizing custom processors, these tech giants aim to lower operational costs and optimize performance for specific AI workloads.