The era of AI inference has fully arrived, and high-cost-performance customized AI chips are mounting a powerful challenge to Nvidia’s dominant position. On March 6th after the US stock market closed, Marvell Technology (MRVL) released its financial results and outlook, which exceeded expectations across the board. This, together with Broadcom’s (AVGO) previously reported explosive growth, jointly confirms the emergence of a new track in the AI chip sector.
The earnings report shows that for the fourth quarter of fiscal year 2026 ended January 31st, Marvell Technology achieved record-high revenue of $2.22 billion, a year-over-year increase of over 20%, surpassing market expectations. Adjusted earnings per share were $0.80, also exceeding forecasts. The data center business, closely linked to AI, contributed $1.65 billion, accounting for 74% of total revenue, a year-over-year increase of 21%. The company emphasized that data center orders are growing at a “record pace,” causing its stock price to surge over 15% in after-hours trading.
The performance guidance appeared even stronger. Marvell expects the median revenue for the first quarter of fiscal year 2027 to be approximately $2.4 billion, significantly higher than the average analyst estimate of $2.27 billion, which had already been revised upward. The company’s CEO stated that revenue would see “further acceleration” in the current fiscal year.
As one of the largest partners for Amazon’s AWS Trainium series chips, Marvell is deeply benefiting from the global wave of AI data center construction. Demand has exploded for its customized AI ASICs, high-bandwidth networking chips, and data center SSD controller chips. AWS positions Trainium as a dedicated accelerator for generative AI, offering significant cost-performance advantages. Google has also announced that the training and inference for Gemini 2.0 are now fully running on its self-developed TPUs. These signs indicate that large cloud providers developing and commercializing their own ASICs has become an industry trend.
Beyond the AI ASIC business, Marvell Technology’s SSD storage controller chips also play a crucial role in the field of “hidden computing power.” Efficient storage systems are vital during the training and inference of large models. As AI workloads shift from training to inference, enterprise SSDs (eSSDs) are experiencing unprecedented structural growth. Recently, multiple institutions, including BNP Paribas and TrendForce, have predicted that prices for DRAM and NAND Flash memory chips will continue to rise significantly, driven by robust demand from AI data centers. This will undoubtedly bring continued positive momentum to Marvell’s storage controller business.
The strong earnings reports from Broadcom and Marvell jointly confirm the growth logic of AI ASICs. The future AI data center is entering an era of heterogeneous computing: cutting-edge training will be dominated by GPUs, while high-frequency workloads like inference will accelerate their shift towards ASIC solutions that offer lower costs and higher energy efficiency. The golden age for AI ASICs has already begun.
