Custom Compute Engine Biz Growing More Than Marvell Ever Hoped

When we look back on it several years from now, three acquisitions will have been key in making Marvell a datacenter player that can compete against rival Broadcom in bringing custom compute engines and networking to AI workloads run by hyperscalers, cloud builders, and model builders.

The first is the acquisition of the Avere custom chip design business for $650 million back in September 2020. This custom chip business is a mashup of the chip design and customization businesses of IBM Microelectronics and GlobalFoundries, both of which used to operate cutting-edge foundries in the United States and Europe.

The second is the approximately $10 billion (all-stock, not cash) deal to acquire Inphi Corp in October 2020. Inphi was a maker of digital signal processors used in optical transceivers that also make amplifiers, laser drivers, memory buffer chips, and a bunch of other things.

The third key deal is the $1.1 billion acquisition of hyperscale Ethernet switch ASIC maker Innovium in August 2021, which has given Marvell a business that Matt Murphy told Wall Street this week on a call going over the numbers for Q4 F2026 that drove over $300 million in business in the year ended in February and that will surpass $600 million in sales in fiscal 2027. This year was all about continuing sales of 12.8 Tb/sec TeraLynx 7 and 25.6 Tb/sec TeraLynx 8 ASICs from a few years back and the ramp of the 25.6 Tb/sec and 51.2 Tb/sec TeraLynx 10 switch chip that was revealed way back in March 2023. A future 102.4 Tb/sec TeraLynx chip will start sampling in the first half of this year, which means any day now, bringing Marvell in line with Nvidia, Cisco Systems, and Broadcom.

Without these, Marvell would have missed the GenAI gravy train. And perhaps the first deal to buy the custom chip business will turn out to be the most significant in the long run for the IT industry, which is always seeking out cheaper alternatives than to pay the Cadillac price for Nvidia compute engines and networking but which has not yet really endorsed an alternative with exuberance.

Murphy was perfectly honest about this on the call with Wall Street:

“On the XPU side, we want to be a big time supplier to our customers. We do get strategic advantage in being in that market when we acquired Avere out of GlobalFoundries. And one of the reasons that I wanted to do that acquisition was get Marvell into custom – I mean, I never envisioned it would be this significant business for us. Let's be clear. Back in 2019, we weren't thinking that we were going to buy an asset for $650 million and it was going to open up a $50 billion TAM. But it did.”

“And one of the strategic rationales that I had for that deal was that it would put Marvell in a product area where we had to be at the bleeding edge. We had to be at the bleeding edge on nodes, packaging, IP development, and it was a tip of the spear type of product line that I felt would be really good for us to have a driving force to keep Marvell on best-in-class technology. Because at that time, we were making the move from fast follower to trying to be a technology leader. So now we are in that business. It has got a lot of noise around it, and it has got a lot of controversies over the last year and all the different things that have gone on, and maybe it's affected a multiple. But the fact of the matter is, we're in that business.”

There has been a lot of chatter about Marvell losing the contract to Amazon Web Services to help shepherd Trainium XPUs through their manufacturing, but that has not happened and, based on the forecast, it doesn’t look like it is going to happen.

Let’s talk about the current quarter numbers for Marvell and then get into the rising forecasts cause by the ever-embiggening GenAI boom.

In the quarter ended in February, Marvell’s revenues rose by 22.1 percent to $2.22 billion, right where we figured they would be in our model. Operating income rose by 71.9 percent to $404 million, and nearly all of that dropped down to $396 million in net income, nearly double that of the year ago period. The revenue is growing from AI XPUs and AI electro-optics, but the profits are growing from the latter and certainly not from the former. The reason why Marvell as gigs with Amazon and Microsoft to help them make custom processors is because they won’t pay a lot for this but they will pay for other stuff wrapped around it and give them Marvell profits that way. (The situation is no different with Broadcom and its big CPU and XPU customers, who buy lots of Broadcom network and storage chippery at much higher margins.)

At 18.2 percent of revenue, this is the highest net income as a percent of revenue from Marvell since fiscal 2011 and 2012, when it was consistently in the low 20 percent for reasons we no longer remember.

Marvell ended the quarter with $2.64 billion in cash and equivalents and a mere $4.47 billion in debt. That is a pretty good balance sheet, and we suspect it will get stronger. Selling off the automotive Ethernet chip business to Infineon for $2.5 billion was the smart thing to do for Marvell. Don’t be surprised when Marvell decides other parts of its business are not core to its strategy going forward.

In recent years, Nvidia has become dominated by its datacenter business, and AMD is following suit to a lesser degree because it still has a pretty big PC chip business. Marvell’s datacenter business is also taking the lead in terms of revenues, and that is why the company is shifting away from detailed divisional reporting of its financials to two major groups: Datacenter is one, and Communications & Other is the other. So take a good look at this divisional breakdown, which we estimated for Q4 F2026 because Marvell did not provide much information with which to get hard figures:

From here on out, this is the granularity we will get formally, but we will get some insight into how different parts of the datacenter business are doing from time to time, such as the more than $300 million in sales for TeraLynx ASICs in F2026 and the forecast for more than $600 million in sales forecast for F2027 cited at the top of this story.

The Enterprise Networking division was mostly edge, industrial, and campus switchery plus PHYs for Ethernet transceivers. Which is why the Datacenter division – now a group – did not suddenly get larger. This tells us that Marvell didn’t really sell a lot of Prestera datacenter switches, and that Innovium ASICs, which are in the datacenter group, really are what it sells these days.

In the quarter, the Datacenter group had $1.65 billion in sales, up 20.9 percent, while the Communications & Other group had $567.4 million in sales, up 25.6 percent. If would be good to get operating profit figures from these two groups, and perhaps the US Securities and Exchange Commission should require this data for all operating groups – and maybe even the divisions expressed in financial reports, too.

On the call, Murphy said that the custom AI XPU business brought Marvell $1.5 billion in fiscal 2026, and said it would grow at least 20 percent in fiscal 2027, and then double from that in fiscal 2028. When we plug that all into our model, here is how the annual AI XPU revenues look:

• F2024: $413 million
• F2025: $780 million
• F2026: $1,471 million
• F2027: $1,765 million
• F2028: $3,351 million

Best we can figure, Marvell had $573 million in overall AI revenues in fiscal 2024, which grew by a factor of 3.2X to $1.85 billion in fiscal 2025 and then grew by a slower 2.3X to $4.23 billion in fiscal 2026. The curves on a quarterly basis for AI XPUs and AI electro-optics (all the other stuff, mostly Inphi chips) look like this:

Given all of the increased spending by Microsoft Azure and Amazon Web Services, and perhaps other customers who want to buy its AI wares, Marvell now says it can bring in around $11 billion in revenues in fiscal 2027, up from a $9.4 billion forecast last October. The datacenter business was projected to grow by more than 25 percent in fiscal 2027, but now Marvell says it will grow by around 40 percent. That’s about $2.3 billion in the datacenter, which leaves about $547 million in Communications & Other. (You see now why the company want to lump this all together, because that is a 3.7 percent decline in this non-datacenter group.)

Looking ahead, Marvell has completed its acquisition of XConn Technologies, which is a maker of PCI-Express and CXL switches, and is preparing an Innovium-branded UALink switch with 115 Tb/sec of aggregate bandwidth (we don’t know how many lanes), which will start sampling in the second half of F2026 and be in volume production in F2028. The XConn PCI-Express 6.0 and CXL 3.1 switch has up to 256 lanes, which is a lot beefier than the PCI-Express 4.0 and PCI-Express 5.0 switches we have seen. Muphy said that the combination of Celestial AI optical interposers and XConn switches would contribute about $250 million in F2028.

Add it all up, and Marvell is forecasting sales of $15 billion in F2028, up nearly 40 percent compared to F2027.