A Chip Company Is Quietly Rising

A Chip Company is Quietly Rising

In the midst of an AI supercycle, Marvell has rapidly transformed into an AI-first data center semiconductor company. Marvell is entering a new era of its historic development.

On December 4, Marvell's market capitalization exceeded $100 billion for the first time after Marvell released its financial results for the third quarter of fiscal 2025. Net income for the third quarter of fiscal 2025 was $1,516 million, $66 million higher than the company's median guidance on August 29, 2024 and an increase of 19% sequentially. Eight years ago, when Marvell CEO Matt Murphy took over, Marvell was in the midst of a restructuring and had a market capitalization of just $3 billion, which jumped from $3 billion to $100 billion now. How does Marvell do that?

Marvell stock price movements

0040-01973 REV.004 CHAMBER BOTTOM RADIANCE 200MM RTPw

High-growth ASIC business

Marvell's custom chip (ASIC) business is becoming one of the core drivers of its strong growth. The ASIC business unit specializes in custom chips designed to meet specific customer needs. In the field of custom chip services, Broadcom and Marvell are the two main contenders. Since both companies operate at scale, it is difficult for other manufacturers to enter this track and gain a competitive advantage.

In the 25 years since the launch of the business, Marvell has designed more than 2,000 custom chips. Through its strategic partnership with Arm Holdings, Marvell has been able to incorporate Arm's core blueprint into its ASIC designs, particularly in the Arm server chip space that has emerged in recent years, adding greater flexibility and market adaptability to its custom chips.

In the era of AI, custom chips have ushered in a new climax of development. Marvell believes that the key to the rise of AI applications is the consideration of total cost of ownership (TCO). But adopting a custom chip design doesn't mean replacing commercial solutions entirely, it's a co-existence. In some large-scale workloads, custom chips can provide better optimization, and based on TCO considerations, it makes sense to move to custom chips. TCO includes not only the cost of the product and the cost required to realize the product, but also the performance that can be obtained.

Custom AI chip services are becoming a cash cow for Marvell. It is understood that Marvell currently has three major customers, and it was previously reported that Marvell has been invited by companies such as Amazon, Google and Microsoft to provide them with custom AI chips as an alternative to Nvidia, whose products have been dominating the market.

In early December, Marvell announced the expansion of its strategic partnership with Amazon Web Services (AWS) for a five-year, multi-generation agreement. For Marvell, this revenue is significant, and more importantly, the multi-generational nature of the agreement, which covers Marvell's broad portfolio of data center semiconductor products, including custom AI products, optical DSPs, active cable DSPs, PCIe retimers, data center interconnect optical transceivers and Ethernet switching silicon solutions. In addition, Marvell will partner with AWS on an EDA (Electronic Design Automation) cloud service that leverages AWS's advanced and scalable computing capabilities to accelerate chip design.

The vast majority of Marvell's custom silicon business is driven by AI demand, particularly in the data center space. According to Marvell's forecast at AI Day, the total data center market size (TAM) is expected to be $75 billion, of which about $40 billion will come from custom chips. Marvell has set a target of 20 percent of the $40 billion market, or $8 billion. With the rapid growth of its custom AI chip business, Marvell has already exceeded $500 million in revenue this year and expects to surpass $1 billion next year.

Marvell's competitive advantage is not only reflected in the design of custom chips, but also in its wide range of interconnect, switching, networking, and data infrastructure IP products, which provide strong support for its custom chip services.

The fast-rising connected business

Marvell has a long history of leadership in connectivity. Marvell pioneered PAM4 SerDes technology more than a decade ago and is an industry leader in PAM4 interconnect shipments, powering front-end and back-end networks in data centers. Today, most data centers are based on PAM4 technology for optical interconnects.

3nm Optical DSP

In the age of AI, as the power consumption of AI and machine learning (ML) applications increases, the need for connected technologies with higher data rates and higher efficiency is becoming more urgent. Marvell has also made significant achievements in optical and optoelectronic hybrid interconnect technologies, further strengthening its leadership in this area.

Optoelectronic connectivity is one of Marvell's fast-growing businesses in recent years. In the third quarter of 2023, Marvell's electro-optical product revenue exceeded expectations, growing by double digits sequentially, particularly with strong orders for Marvell's 800G PAM products and the beginning of deliveries of the industry's first 1.6T PAM DSP and 5nm process technology.

Marvell's rapid growth in the optoelectronic interconnect space was driven by the forward-looking acquisition of Inphi, a power-optical interconnect company, in 2020. Inphi's electro-optical products provide critical interconnect infrastructure for cloud data centers and wired and wireless carrier networks. The acquisition not only expands Marvell's market potential, but also strengthens its customer base and accelerates Marvell's leadership in hyperscale cloud data centers and 5G wireless infrastructure.

To meet the huge demand for the highest bandwidth and lowest power consumption in AI, Marvell recently introduced the industry's first 3nm 1.6 Tbps PAM4 DSP, Marvell Ara, the industry's first 3nm 1.6 Tbps PAM4 interconnect platform with 200 Gbps electrical and optical interfaces. Building on the success of the Nova DSP, Ara has reduced the power consumption of optical modules by more than 20% compared to its predecessor. In 2023, Marvell introduced the industry's first 5nm 1.6 Tbps PAM4 DSP with a 200 Gbps electrical and optical interface.

Optimized for next-generation AI and cloud infrastructure, Ara is designed to support high-density 200 Gbps I/O interfaces on switches, network interface cards (NICs), and XPUs, while ensuring backward compatibility with previous generations. With best-in-class energy efficiency and integration, Ara meets the growing demand for hyperscale data centers to deliver high-performance accelerated infrastructure at best-in-class total cost of ownership (TCO).

The 1.6T will be a significant source of growth for Marvell next year, with Marvell Ara sampling available to select customers in the first quarter of 2025. "We expect shipments of PAM4 DSPs to more than triple from 2024 to 2029 to nearly 127 million units per year, and will remain the primary optical technology connecting assets within the data center for the foreseeable future," said Bob Wheeler, contributing analyst at LightCounting. Ara marks another first for Marvell and demonstrates that PAM4 technology continues to evolve to meet the challenges of AI infrastructure. "

PCIe retimer for 5nm

In addition to DSPs, PCIe retimers are also an area of Marvell's interconnection business that has only become popular in recent years. The space is heating up with the massive IPO of startup Astera Labs launching PCIe retimer technology. Since then, companies such as Broadcom and Marvell have jumped on board. Now, Marvell has seen success with its PCIe retimer and other product lines.

In May 2024, Marvell introduced a new line of PCIe retimers, the Alaska P PCIe retimer, built on Marvell's industry-leading 5nm PAM4 technology to extend connectivity between AI accelerators, GPUs, CPUs and other components within the server. With a typical power consumption of 10 watts, the 5nm 16-lane PCIe 6 retimer is currently the industry's lowest power retimer. Retimers can be used for on-board or cable-copper connections, or combined with electro-optical components to produce optical PCIe modules to meet different cloud customer data center architectures. Marvell is working with cable and optical module partners to integrate products into cloud-optimized interconnect solutions for different data center customer applications.

Alan Weckel, co-founder of 650 Group, said: "Signal distance is a real challenge for service providers. We estimate that more than 75% of cloud and AI servers shipped in two years will rely on retimers, which will contain multiple retimers. An 8 GPU server may contain 16 or more of these devices. Retimers can even infiltrate enterprise servers. So that's a significant revenue stream for Marvell.

Marvell Alaska P PCIe Retimer

51.2T Ethernet switch

On July 25, 2024, Marvell introduced the Teralynx 10, a 5nm 51.2T switch for data centers. The Teralynx 10 Ethernet Switch is a low-power, programmable, 51.2 Tbps Ethernet device with the industry's lowest latency that delivers leading performance for training, inference, general-purpose computing, and other workloads to scale the accelerated infrastructure of cloud data centers. Here's how it works: it has a throughput of up to 51.2 Tb/s, latency as low as 500ns, and latency below 600ns for all packet sizes. The switch consumes only 1W per 100 gigabit of bandwidth per second. The 512 switching base enables operators to reduce the number of switch tiers in large clusters, resulting in significant reductions in power and total cost of ownership (TCO).

Ethernet switches are also a significant market, with shipments of 51.2 Tbps replacement units expected to soar from approximately 77,000 units in 2024 to 1.8 million units in 2028, at a CAGR of 120%, according to 650 Group. Marvell believes this area will not only continue to grow over the next year, but will be a very strong growth area for the company in the long term.

In the switch space, Marvell has also made acquisitions. On August 3, 2021, Marvell acquired switch vendor Innovium. Innovium's TERALYNX switching architecture delivers ultra-low latency, optimized power, high performance, and innovative telemetry, which are critical to today's cloud-scale data centers, and at the time of the acquisition, Innovium was selected as a key vendor for Tier 1 cloud customers.

In addition, on September 19, 2019, Marvell completed the acquisition of Aquantia, a Gigabit Ethernet company. The acquisition of Aquantia complements Marvell's copper and optical-physical layer portfolio and expands its position in the Multi-Gig 2.5G/5G/10G Ethernet space. In particular, Aquantia's innovative Multi-gig automotive PHYs combined with Marvell's industry-leading Gigabit PHYs and secure switch products create the world's broadest and most advanced high-speed in-vehicle networking solutions.

Become an AI company and start a new era of growth

After delivering strong third-quarter results, Marvell expects year-over-year revenue growth to accelerate significantly in the fourth quarter, with fourth-quarter revenue expected to be around $1.8 billion, fluctuating in a range of ±5%. Marvell, the largest data center end market, expects Q4 growth to be in the mid-20% range, driven by another significant increase in custom AI revenue, as these projects continue to scale into large-scale production. Growth will also be driven by Ethernet switching products and our interconnect products, including optical DSPs, TIAs, drivers, AEC and DCI products.

For the upcoming fiscal year 2026, Marvell is optimistic about almost all businesses. Specifically, the customization business is expected to usher in overall growth next year, the optical business will continue to maintain a strong momentum, and the switching business is also growing steadily. At the same time, emerging fields such as AECs are beginning to enter the stage of real mass production.

Marvell expects sales of AI networking and custom processor chips to reach $2.5 billion by fiscal 2026. Analysts expect more from it, though, with Benchmark Research analyst Cody Acree saying on Wednesday that the venture capital firm sees Marvell as a "unique non-Nvidia alternative" in the AI and data center markets, while UBS said in a report that Marvell's AI revenue could reach $3.6 billion by 2025, exceeding the company's $2.5 billion target.

Conclusion

Taken together, the precise grasp of AI opportunities and mergers and acquisitions are undoubtedly key drivers for Marvell to achieve a market capitalization of $100 billion.

Marvell has built a powerful "arsenal" around AI, covering custom chips, connected devices, computing, networking and other areas, forming a diversified and deeply integrated technology ecosystem. From the acquisition of automotive Gigabit Ethernet company Aquantia in 2019, to the acquisition of Inphi, an optoelectronic interconnect solutions company, in 2020, to the acquisition of Innovium, a key switch provider in the cloud computing space, Marvell has rapidly strengthened its technology reserves in the areas of data center, cloud computing and optoelectronic interconnection through these key acquisitions. In addition, Marvell has optimized its business structure by divesting non-core businesses, such as the $1.76 billion sale of its Wi-Fi connectivity business to NXP in 2019 and focusing its resources on data center and AI-related technologies.

During the earnings call, Marvell CEO Matt Murphy emphasized that the company has a clear capital allocation strategy that is 100% focused on the opportunities of the AI supercycle.

As for Marvell's future market capitalization potential, $100 billion may just be a starting point. In the fast-growing AI market and data center space, where can Marvell's business support Marvell's market capitalization? Everything is unknown, after all, I believe that many people did not expect Nvidia to reach a market value of 3 trillion.