Marvell Pivots To AI Silicon, Looks Poised To Profit

It is hard to bet against the GenAI boom, and thus far it is also hard for anyone other than Nvidia to profit from it.

No one knows these facts better than Marvell Technology, who along with rival chip maker Broadcom, is seeking to benefit from the bevy of custom silicon that the hyperscalers and cloud builders are etching as they seek to build more efficient compute and networks to support their own AI aspirations to drive revenues and secure profits.

It is all a very hopeful act, really. But it seems like a reasonably safe bet, and like the commercialization of the Internet, which started in the late 1990s and early 2000s but which did not really go commercial until a second wave that began perhaps in 2005 or 2006 and gained steam after the Great Recession, separating those with the economic and technical might from those who were less fortunate, we expect for the second GenAI wave that might begin in perhaps 2028 or 2029, or maybe even earlier in 2027, to separate the pretenders from the contenders.

It is safe to bet that the existing hyperscalers and large cloud builders will be the contenders, and that is why we see the providers of the dominant foundation models (mainly OpenAI and Anthropic) tightly coupled to those who have vast AI fleets already. And therefore it is no surprise at all that Marvell is betting its business on these companies needing some intellectual property, packaging expertise, and manufacturing help to bring their homegrown chips to fruition.

In the third quarter of fiscal 2025 ended on November 2, Marvell posted revenues of $1.52 billion, up 6.8 percent year on year but 19.1 percent sequentially as it has two major hyperscaler and cloud builder customers ramping compute engines and DPU chips as well as adopting Marvell’s own networking ASICs to build out their infrastructure.

The pivot to AI came just in time for Marvell, and this is not an accident for chief executive officer Matt Murphy and his executive team.

The company’s enterprise networking, carrier infrastructure, and consumer chip businesses have all been sliding for the past three years after hitting their peaks in fiscal 2023 (which is almost the sale thing as calendar 2022). That timing is not an accident, we think, because the GenAI boom hit in late calendar 2022 and the hyperscalers and cloud builders all pivoted their own investments away from their vast general purpose infrastructure and towards building a new class of AI supercomputers. Just like we think companies of all sizes and stripes are keeping their general purpose servers around longer so they can spend on new AI infrastructure, we think the same thing happened in networks. And hence the declines that Marvell has seen in these core businesses as it has become one of two key suppliers of chip expertise for at least some of the titans of AI.

In the third quarter, Marvell booked a $358.3 million charge related to the impairment of acquired intangible assets and purchased technology licenses, the latter of which have future contractual obligations. Willem Meintjes, Marvell’s chief financial officer, was not more specific about these writeoffs, but said they related to the enterprise networking and carrier infrastructure business units. Murphy stressed, however, that Marvell had made deep investments and acquisitions in these two areas over the past several years and will be more targeted going forward. More money will be poured into research and development around the AI opportunity.

Even without this charge, Marvell was not going to be profitable in the quarter, just as it has not been for the past six quarters. This is a very tough business to make money in, particularly when Broadcom is your main competitor and the hyperscalers and cloud builders are your main customers. This is just inherently tough, and we have tremendous respect for Marvell’s people staying in the fight and getting chip done day in and day out. Most businesses are like this, and Wall Street sometimes forgets this and puts too much pressure on companies to do better than Marvell and its ilk can.

To put some numbers on it, in our model running from February 2008 through August 2024, Marvell had total revenues of $54.91 billion, but only had $1.11 billion in operating revenues and thanks to some asset sales had net income of $1.8 billion. That is an operating income of 2 percent of revenues and a net income of 3.3 percent of revenues. A big box retailer or an ODM or OEM IT supplier does about this level of profitability. (Small wonder, then, that Broadcom bought the legacy cash cows CA and VMware. Hock Tan’s momma didn’t raise no dummy. . . .)

In Q3 F2025, Marvell had an operating loss of $702.8 million, 4.8X worse than the year ago period and 7X worse than in Q2 F2025. The net loss was $676.3 million, which was 4.1X higher year on year and 3.5X higher sequentially. Even without the writeoffs, the red ink is getting deeper as Marvell makes this pivot to AI at the same time that its enterprise networking and carrier infrastructure businesses shrink from their fiscal 2023 peaks of $1.37 billion and $1.08 billion, respectively.

We do not think that Marvell thought these businesses would shrink like this any more than Intel thought its server CPU business was going to shrink in the face of the GenAI boom. People no doubt thought of this as a chance for incremental – and juicy – sales. But the companies building AI training clusters saw a need to curtail existing spending to compensate for the crazy budgets of this new GenAI area. There has been incremental spending, to be sure, but maybe only half as much net increase as many had been expecting.

The good news for Marvell is that its datacenter business is booming, and we cannot know if it is profitable or not because the operating income is not allocated by group in its financials. For all we know, the enterprise networking and carrier infrastructure businesses are losing money as well as the AI portion of the datacenter business. Marvell was just barely profitable before its AI pivot a few years back, and we think this is the likely scenario. And that means as enterprise networking and carrier infrastructure gradually grow, they should return to profits and as Marvell gains expertise in getting custom chips shepherded through the foundries for hyperscalers and cloud builders, this should also flip to a profit.

Just how much profit remains to be seen. If history is any guide, hope springs eternal but profits are elusive.

In the third quarter, datacenter revenues at Marvell nearly doubled to $1.1 billion and was up 25 percent sequentially. Based on statements made by Murphy and Meintjes, we believe that enterprise networking and carrier infrastructure businesses will grow 15 percent sequentially in Q4 F2025, consumer chips will be down, and auto/industrial will be down while datacenter will grow 25 percent sequentially to $1.38 billion. Add it all up, and Marvell could be looking at $1.82 billion in sales, and there is a chance that it might be at breakeven or even a little better. Or perhaps a little worse. We shall have to see.

What we also know is that Marvell had pegged its AI revenue opportunity at $1.5 billion in fiscal 2025 and $2.5 billion in fiscal 2026. (About $500 million of the F2025 number is for custom AI ASICs, and in F2026 it is about $1 billion). Now, Marvell says it is running at well above that rate for the overall AI revenues it can take down this year and next. The question is by how much, and how fast can it reach its stated goal of $8 billion in AI revenues by 2028, which we discussed at length back in June.

It doesn’t hurt that Marvell just inked a five-year technology supplier agreement with AWS this week, covering AI compute engines and a range of networking products that includes optical DSPs, active electrical cable DSPs, PCI-Express retimers, datacenter interconnect optical modules, and Ethernet switching ASICs. Conversely, Marvell is going to be using EDA simulation software on the AWS cloud, which means some of the money that AWS spends on gear is coming back to it through services. We don’t know the specifics of the revenues back and forth between the two.