光器件仍是 Marvell AI 业务的主要驱动力，超过定制芯片

CONNECT

The datacenter has become the dominate factor in driving the business of chip maker and chip shepherd Marvell, and the growth in this business is almost exclusively coming from components for AI clusters. While it is exciting that Marvell has big deals with Amazon Web Services, Google, and others for custom chips, this business is still relatively small. Optical components for networks as well as switching are still driving the bulk of the sales.

Legacy chips in the aimed at telcos and other carriers, enterprise networking, and consumer and industrial devices – which once were a bigger part of a much smaller Marvell – are still humming along with their ups and downs, but Marvell does not really talk about these much anymore except as they are lumped into a single group. Like this:

Nvidia, for what it is worth, has done the same thing, lumping everything that is not part of its datacenter business into a group called Edge Computing. In fact, this chart looks very much like Nvidia’s chart, although the numbers on the Y axis are a hell of a lot smaller and the exponential curve is not quite so steep for Marvell. At least not yet. But watch as AI networking and custom XPUs and CPUs take off among the hyperscalers, cloud builders, and AI model builders. Marvell is becoming every bit as much of a counter balance to Nvidia’s largesse in GenAI as is Broadcom.

In the second quarter of fiscal 2027, which ended on August 1, Marvell had $2.17 billion in datacenter product sales, up 45.7 percent year on year and up 18.5 percent sequentially. My model says that legacy enterprise networking, which should arguably be added to the datacenter, brought in an additional $266 million in the quarter, and I think it rose 37.4 percent year on year. Admittedly, any breakdown in the old legacy divisions Marvell used to talk about are guesswork, since the company has not talked explicitly about their revenues since Q3 F2026.

Add all of those legacy businesses up, and they rose by 10.1 percent in the quarter to $567.8 million. The datacenter part of the Marvell business is now 79.3 percent of total sales, and if you add in enterprise networking, that drives it up to 89 percent of revenues.

Just for fun, here is how the datacenter business compares to the other former divisions in my model:

Unless Marvell talks about these explicitly at its financial analyst day coming up soon, the error bars on these estimated division sales are getting fuzzier and fuzzier. And this chart also looks very much like Nvidia’s old divisional divide before and after the GenAI boom.

The thing to note is that the datacenter business has grown a factor of 8.1X since Q2 F2021, while those legacy chip businesses have grown by 1.3X over the same time. Imagine if Marvell had not bought IBM’s custom chip business out of GlobalFoundries as well as Inphi and Celestica AI. . . . Selling of the automotive chip business to Infineon was smart, too. It’s all hands on the GenAI deck. And it will not be surprising if Marvell sells off other divisions to get cash for maneuvering room and to have better focus.

What we really need to do is drill down into that datacenter business, and it would of course be helpful if Marvell would break this much larger business down into divisions. We have gotten enough hints and a few data points here and there from the Marvell top brass to build a model of how such divisions might be created and how much money they are driving. The US Securities and Exchange Commission should have rules about when a business that is clearly comprised of divisions needs to be financially reported as such. The loose rule at the SEC is that the way you run the business and talk about it internally is supposed to be the way you talk about it – in a very general sense, of course – to investors.

Here is my attempt to drill down into Marvell’s overall datacenter and AI-related datacenter products businesses and breaking that down into two main groups And here is a chart breaking out AI electro-optics from AI custom chips and then other datacenter products. (This does not include enterprise networking.) Anyway, take a gander:

As the GenAI boom was just getting started back in early 2023 – call it Marvell’s Q1 F2024 – the company had an AI business that was around $77 million, with on the order of $12 million in early custom ASICs through a deal with Amazon Web Services for Graviton chips and maybe $66 million in electro-optics component sales.

Fast forward to Q2 F2027, a little more than three years later, and I estimate that AI revenue for Marvell was $1.82 billion, up 90.9 percent year on year and 23.6X higher than three years ago. AI XPU sales slumped a little sequentially in the current quarter, to around $425 million, but that was up 43.1 percent and faster than the 20 percent or so growth that Marvell told Wall Street to expect a few quarters ago.

Electro-optical components continue to dominate the AI story at Marvell, with revenues of around $1.4 billion, up 97.2 percent, in my model. Other datacenter products sold outside of the AI customer set accounted for around $347 million, therefore, down 28.4 percent.

Add it all up, and Marvell had $2.74 billion in revenues, up 36.5 percent, with an operating income of $460 million, up 58.5 percent, and net income of $308 million, up 58.1 percent. The company has $3.93 billion in the bank, which it will need to fulfill its obligations to chip shepherding customers.

Looking ahead, Marvell is raising its guidance once again, and now says to expect around $3.15 billion in sales in Q3 F2027, which would be an increase of 51.8 percent year on year and up 15 percent sequentially. Matt Murphy, Marvell’s chairman and chief executive officer, said that growth will accelerate in Q4 2027. Notably, the first revenues from Google’s recently announced expanded partnership with Marvell hit in Q3 F2027. The datacenter business is expected to grow by around 20 percent sequentially and by around 75 percent year on year in the third fiscal quarter.

Murphy sweetened the fiscal 2027 year’s revenues to $12 billion, up from $11.5 billion a quarter ago, and that puts Q4 F2027 at around $3.7 billion, up 66.4 percent year on year. The datacenter in general and AI specifically are driving this growth, obviously. If you do some math in pencil on a scrap piece of paper, custom CPUs and XPUs will represent about $2 billion in sales in F2027, rising at around 33 percent and well ahead of the 20 percent growth that Marvell was talking about a few quarters ago. That legacy comms and other chip business is expected to grow by a mere 10 percent or so in fiscal 2027.

As for fiscal 2028, Murphy upped the guidance by $1.5 billion from the prior quarter’s guidance to $18 billion, which is a 50 percent growth rate, and added that the datacenter business would grow by more than 60 percent in fiscal 2028. Scale up network components for UALink, ESUN, and NVSwitch networks as well as for the Celestial AI optical interposer are expected to be on the order of $300 million, with Celestial AI’s hardware representing about $150 million. The custom CPU and XPU accelerator business in F2028 is expected to more than double in fiscal 2028 and grow even faster in 2029, in part due to the Google deal but also because of chip projects that Marvell is shepherding for AWS and others.

SPONSORED FEATURE: HPE and NVIDIA help organizations apply accelerated computing, software, enterprise infrastructure, networking, control plane, and services to the workloads and business outcomes that matter most.

October 1, 2026

September 30, 2026

AMD Buys World Labs For $8.2 Billion To Get AI Model Cred