AI Boom Shifts From Model Training to Inference

The artificial intelligence boom is entering a new phase as the industry shifts focus from large-scale model training to AI inference [1, 2, 3].

This transition matters because it changes the technical requirements for hardware. While general-purpose GPUs were essential for building AI models, the deployment and execution of those models—known as inference—may favor different architectures, potentially altering the competitive landscape for chipmakers and investors [1, 3].

Wall Street investors and AI hardware makers are currently reassessing which companies will dominate this next era [1, 2]. For years, Nvidia has been a primary beneficiary of the boom, providing the GPUs used by the world's top AI companies ^[5]. However, some analysts said that Nvidia may not be the biggest winner as the market matures ^[1].

Custom Application-Specific Integrated Circuits, or ASICs, are emerging as a primary challenger to the GPU-centric model ^[1]. Broadcom is positioned as a significant player in this shift, with its custom ASICs potentially conquering the AI inference market ^[1].

This shift is driven by the need for more rational returns on massive AI spending [1, 3]. As the market matures, companies are seeking hardware that offers better efficiency, and lower costs for specific inference workloads rather than the raw power required for initial training [1, 3].

While some firms continue to view Nvidia as the central pillar of AI growth, the rise of specialized hardware creates uncertainty regarding long-term market leadership [1, 5]. The competition now centers on whether general-purpose hardware can maintain its lead or if the era of specialized, custom silicon will define the next phase of AI deployment [1, 2].