ASICs for AI on the rise at GPUs expense

Application-specific AI inference accelerators are gaining market share in the datacenter, a trend that is benefiting chip vendors Broadcom and Marvell, rather than the dominant AI player Nvidia, according to The Information Network. The semiconductor analyst predicts that in 2023, GPUs held 72 percent of the market for AI acceleration, while ASICs accounted for 22 percent. By 2025, these numbers are expected to shift significantly, with GPUs losing 7 percentage points and ASICs gaining 8. This trend is projected to continue through 2027.

While Nvidia remains the dominant vendor in the market for AI acceleration in the data center with its versatile GPUs capable of handling various tasks, ASICs optimized for inference are increasingly making an impact. Robert Castellano, founder and principal analyst with The Information Network, highlights the growing significance of ASICs in this space, emphasizing their role in inference tasks.

Meanwhile, CPUs and FPGAs, the more general-purpose chips, are anticipated to occupy approximately 5 to 7 percent of the market. This indicates a diversification in the types of chips being utilized for AI acceleration, reflecting the evolving needs of data center applications and workloads.

Both Broadcom and Marvell reported strong earnings in December 2024, largely attributed to the sales of ASICs for AI applications. While Nvidia's GPUs are primarily used for model development and training, there is nothing preventing the company from entering the ASICs market to compete with Broadcom and Marvell in this growing segment.

As the landscape of AI acceleration in data centers continues to evolve, the market dynamics are shifting towards application-specific solutions such as ASICs. While Nvidia's GPUs have traditionally dominated this space, the rise of ASICs optimized for inference tasks is challenging this status quo. With Broadcom and Marvell capitalizing on the demand for ASICs, the market share distribution is expected to undergo significant changes in the coming years, presenting new opportunities and challenges for chip vendors.