On the eve of Marvell earnings, will custom AI chips break out?

Introduction

With the rapid evolution of artificial intelligence, the chip industry is undergoing an unprecedented structural transformation. While general-purpose GPUs continue to dominate AI training, custom AI chips (ASICs), with higher energy efficiency and targeted optimization, are gradually becoming the strategic choice of large technology companies. On June 15, U.S. Eastern Time, semiconductor giant Marvell Technology will release its latest quarterly earnings, and the market is closely watching whether its custom AI-chip business can become a new growth engine. On the eve of earnings, one question is being asked inside and outside the industry: has the breakout phase for custom AI chips already arrived?

Main text

1. Marvell’s custom AI-chip layout: a strategic leap from connectivity to compute

As a global leading semiconductor solutions provider, Marvell has traditionally been known for networking, storage, and connectivity chips. In recent years, however, through a series of acquisitions and in-house R&D, the company has gradually entered the custom AI-chip arena. Acquisitions such as Innovium and Inphi not only strengthened its data-center networking capabilities, but also accumulated key technologies for custom compute chips.

Marvell’s custom AI-chip business mainly targets cloud giants such as AWS, Microsoft Azure, and Google Cloud. These customers need deeply optimized chips for specific AI workloads such as inference, recommendation systems, and natural language processing, rather than general-purpose GPUs. With deep expertise in SerDes, high-speed interconnect, and SoC integration, Marvell can provide end-to-end custom services from design to mass production. Industry estimates suggest that Marvell’s custom AI-chip business could double in fiscal 2026 and become one of the company’s most important revenue sources.

2. Market backdrop: why are custom AI chips breaking out now?

Custom AI chips are not new, but demand has exploded over the past two years for several reasons.

First, AI inference demand is surging. As large-model applications shift from training to deployment, inference workloads far exceed training. General-purpose GPUs are too power-hungry and expensive in inference scenarios, while custom chips can deliver tens of times better efficiency. For example, the custom chip Marvell designed for AWS uses only about one-tenth the power of a general GPU in video encoding and recommendation use cases.

Second, cloud vendors want to cut costs and improve efficiency. Large cloud service providers spend tens of billions of dollars in capex every year, with chip purchases taking a huge share. Self-developed or custom chips can significantly reduce total cost of ownership (TCO) over the long run while reducing over-reliance on a single supplier such as Nvidia. As a neutral design-services provider, Marvell fits these customers’ needs for both performance optimization and supply-chain security.

Third, manufacturing processes and design tools are maturing. With TSMC’s 3nm/5nm processes becoming more widely used and EDA tools gaining AI-assisted design capabilities, custom-chip development cycles have shortened from the traditional 2–3 years to 12–18 months, and costs have fallen sharply. That allows even mid-sized tech companies to join the custom-chip race, further expanding the addressable market.

3. Competitive landscape: Marvell’s differentiated strengths and challenges

In custom AI chips, Marvell faces competition from Broadcom, MediaTek, and many startups. Compared with Broadcom, Marvell is more focused on customized networking connectivity; compared with MediaTek, Marvell has deeper roots in the data-center market. Marvell’s core strength is its “connectivity + compute” synergy—being able to seamlessly integrate custom compute cores with high-speed network interfaces for end-to-end system-level optimization.

However, the challenges are real. First, custom-chip projects are long and customer-dependent; once the design is frozen, switching costs are extremely high, so acquiring new customers is difficult. Second, Nvidia is also launching customized GPU series for inference scenarios, such as L40S and H200. Although these products are not as energy-efficient as dedicated ASICs, their mature software ecosystem (CUDA) still suppresses custom chips. In addition, geopolitical risk deserves attention, as U.S. export controls to China may affect Marvell’s expansion in some markets.

4. Earnings preview: key metrics and market expectations

For this earnings report, the market will focus on several dimensions: revenue growth in the custom AI-chip business, visibility on customer orders, and guidance for the next 12 months. If Marvell can announce new large custom-chip customers or expansion plans from existing customers, the stock could see another leg higher. Conversely, if guidance comes in below expectations, concerns may rise that AI-chip investment is overheating and due for a pullback.

Historically, Marvell’s custom AI-chip business grew more than 150% year over year in 2025, but from a low base. Analysts generally expect fiscal 2026 revenue from this business to exceed $3 billion, accounting for more than 40% of total company revenue. That would mark Marvell’s successful transformation from a “connectivity-chip company” into a “custom compute company.”

Conclusion

The breakout of custom AI chips is not accidental; it is the inevitable result of technological progress, market demand, and deeper industry specialization. Marvell, as a core player in the space, is at a critical earnings inflection point. Whatever the numbers show this time, the long-term trend for custom AI chips is already set—it is reshaping profit distribution across the semiconductor industry, shifting power from general chip vendors to custom design-service providers. For investors and industry participants alike, while watching Marvell’s earnings, it is even more important to grasp the broader industrial change it represents: AI computing is moving from “standard answers” to “custom solutions,” and that may be the true source of long-term investment value.