What’s Going on with Marvell (MRVL US)?

Recently, several readers have inquired me about the latest developments with Marvell and Alchip, so today I will provide a brief update on this. This article on Marvell is divided into two sections, mainly explaining about the company’s AI-related business: the first half covers Marvell’s data center optical module DSP business for 800G and above, while the second half focuses on Marvell’s ASIC chip business. As opinions among investors on Marvell are highly polarized, I hope to offer an as objective perspective as possible — please refrain from harsh criticism if you disagree.

First of all, we know that Marvell classifies its sales from 800G and 1.6T optical module chips used in data centers as AI-related revenue. Currently, for 800G optical modules, Marvell typically sells its 800G PAM4 DSP chip bundled with TIA + Driver chips together to customers (readers unfamiliar with the structure of optical modules can refer to the diagram below). However, starting last year, this business segment has faced many new competitors: Broadcom and MaxLinear have begun qualifying their 800G DSP chip with optical module customers, while Macom and Semtech are qualifying their TIA + Driver chips with customers as well.

According to my supply chain research, if a customer purchases Marvell’s chip solution, the quoted price for the DSP of an 800G optical module is $90–100, and the TIA + Driver is $40–50, compared to Broadcom and MaxLinear’s 800G DSP price of $50–60 and Macom and Semtech’s TIA + Driver price of $25–30. From the perspective of optical module manufacturers, adopting Marvell’s solution is nearly twice as expensive as using the Broadcom/MaxLinear + Macom/Semtech combination, while the selling price of 800G optical modules is almost the same. Therefore, optical module makers are motivated to push the alternative suppliers’ solution (see table below):

Right now Marvell still holds more than 80% market share in the 800G optical module market, since CSP customers still have concerns about the reliability of optical module solutions using a mix of Broadcom/MaxLinear + Macom/Semtech chips. This is also why Marvell dared to raise prices to optical module customers at the beginning of this year. However, I believe that over time, the penetration rate of alternative solutions will inevitably increase.

In the 1.6T optical module DSP space, in addition to Broadcom launching its own product to compete with Marvell, a more serious issue is that NVIDIA is heavily promoting its own Mellanox DSP chip for 1.6T optical modules. My supply chain check indicates that NVIDIA launched its self-developed 1.6T DSP chip as early as the beginning of last year. However, because the chip was developed by the Mellanox team — which is known for making switch chips — the 1.6T DSP chip initially had a significant issue with high power consumption. But after more than a year of improvements, NVIDIA’s 1.6T DSP chip has now reached a level of power consumption that is generally acceptable to customers. It is currently undergoing final adaptation and validation process with a few optical module manufacturers and is expected to start mass production in the second half of this year.

Objectively speaking, although Marvell faces strong competition in both the 800G and 1.6T optical module chip businesses, this does not necessarily mean that its optical module chip revenue will begin to decline. Recently, Zhongji Innolight optimistically stated that the 800G optical module market could nearly double to 30 million units in 2026. If that turns out to be true, it could drive continued growth in Marvell’s data center optical module chip revenue next year (simply because the company’s share in the 800G market is unlikely to drop by half while the total market size doubles). For detailed calculations, please refer to the table below:

From the table above, we can see that the overall 800G optical module market size is estimated to be 16 million units in 2025, with Marvell still holding 80% market share. In 2026, according to Innolight’s forecast, the 800G optical module market could reach 30 million units. Assuming Marvell’s market share drops from 80% to 60%, its 800G optical module chip revenue could still grow from $1.8 billion to $2.3 billion. On top of this, I estimate that the total 1.6T optical module market in 2025 will reach 3 million units, with one-third of DSPs supplied by NVIDIA in-house solution (assuming mass production begins smoothly in 3Q25), and the remaining 2 million units using Marvell’s DSP. In 2026, assuming the total 1.6T optical module market double to 6 million units, with NVIDIA’s in-house DSPs accounting for half, and Marvell holding 80% market share of the remaining 3 million units, the company’s 1.6T DSP revenue could grow from $320 million this year to $360 million next year. Combining 800G and 1.6T together, Marvell’s data center optical module chip business revenue could still enjoy a 28% YoY growth in 2026.

Nevertheless, although Marvell’s optical module chip business next year may be saved by strong 800G demand growth, the recently announced retirement of Dr. Loi Nguyen (see screenshot below) may signal some long-term issues of this business segment in the future. For those readers who are not familiar with Dr. Loi Nguyen, he was head of Marvell’s optics business, as well as the Co-Founder of Inphi (later acquired by Marvell) and the inventor of the PAM4 DSP.

Having covered Marvell’s optical module chip business, I will now continue to explain the latest development regarding with Marvell’s ASIC chip business, including updates on its AWS Tranium and Microsoft Maia chips.