Marvell: Interconnect Guidance Raised to 70%, Custom Silicon Waits for 2028

Fibre-optic and network cables with connectors on a white surface, illustrating a report on Marvell's interconnect business. Photo by PawinG on Pixabay

BUSINESS · 18 Aug 2026 —

SANTA CLARA, 18 AUG 2026 — Marvell returned 181 per cent over the past year, outrunning both Broadcom and Nvidia. The part of the company that did it does not compute on data. It moves data between the chips that do.

While Marvell is best known for its custom AI silicon, its revenue growth is coming from interconnects. That gap is worth exploring ahead of the company's 27 August earnings report.

The numbers

181%One-year return, ahead of Broadcom and Nvidia

76%Share of fiscal Q1 2027 revenue from data centre

>70%Expected interconnect growth in fiscal 2027, raised from 50%

>20%Expected custom silicon growth in fiscal 2027, doubling in fiscal 2028

Trailing twelve-month revenue is US$8.7 billion, up 34 per cent. The company is targeting US$16.5 billion for fiscal 2028, a figure it raised by US$1.5 billion just last quarter. Analysts expect roughly US$2.71 billion of revenue and non-GAAP earnings near US$0.93 a share when the second quarter of fiscal 2027 is reported.

Interconnect here means optical digital signal processors, transimpedance amplifiers, drivers and pluggable modules for connecting data centres. Custom silicon means designing accelerators for other companies, where Marvell is estimated to hold 20 to 25 per cent of the co-design market, with work reported on Amazon's Trainium and Microsoft's Maia programmes.

Why the unglamorous half is winning

A custom accelerator is a multi-year engagement with one customer, and its revenue arrives when that customer's chip ships in volume. Interconnect sells into every cluster being built, regardless of whose silicon is inside it.

That difference decides the timing. The AI build-out is currently in its construction phase, and construction consumes the parts that connect things long before it consumes the finished designs commissioned two years ago. Guidance raised from 50 per cent growth to more than 70 in the same year is what that phase looks like from inside a supplier.

It also makes interconnect the more diversified business of the two. A custom programme concentrates a large amount of revenue in one customer's success. Optical components sell to whoever is building, which in this market is everyone.

Everything is a data-movement problem

This lands in the middle of a pattern we have been tracking for a fortnight, and it is the clearest instance yet.

We wrote that Kioxia is taking an optical SSD to customer trials because copper limits how far storage can sit from a processor. We wrote that Cerebras sells inference speed by keeping weights in on-chip memory, removing the bandwidth bottleneck rather than mitigating it. And we wrote that SK hynix is committing US$38 billion to memory fabs that feed accelerators which would otherwise sit idle.

In four different layers of the stack, the constraint is the same: the accelerators are fast enough, but getting data to them is the bottleneck. Marvell is the purest financial expression of that so far, because its market has voted for the part of the company that moves bytes over the part that processes them.

What a supplier sees that a buyer does not

This pattern shows up in a component supplier's numbers first for a reason that generalises to the whole market.

An operator buying capacity sees one line item for compute and treats everything around it as plumbing. A component supplier sees which plumbing is being ordered, in what quantity and how far ahead. When orders shift from the parts that process data to the parts that move it, the supplier finds out first, and the shift appears as a guidance revision months before anyone writes an architecture piece about it.

That makes component guidance an unusually early indicator for anybody planning infrastructure. A revision from 50 to more than 70 per cent on optical interconnect is a statement about how many clusters are being wired this year, and it is more reliable than the announced-capacity figures that circulate, because somebody has placed an order behind it.

What this says about the ASIC narrative

The industry has spent two years arguing that custom silicon erodes the general-purpose accelerator business, and we reported that China's domestic shift runs on a dual track of GPUs and proprietary ASICs.

Marvell's figures do not contradict that thesis; they date it. Custom silicon is expected to grow more than 20 per cent this fiscal year and more than double the next, which is a business arriving rather than a business failing. The correction is to the timing: the ASIC transition is a fiscal-2028 event at one of its main beneficiaries, not a 2026 one.

For anyone making procurement or investment decisions on this timing, the distinction is critical: the design wins exist, but the revenue has not yet arrived. The design wins exist. The revenue has not arrived.

The regional read

Optical interconnect is one of the few parts of the AI supply chain where Southeast Asia already has an industrial position.

Optical module assembly, photonics packaging and test are established activities in Malaysia, Thailand and the Philippines, and they sit in exactly the segment growing at more than 70 per cent. That is a more realistic place for regional manufacturing to capture value than accelerator fabrication, which requires capital and process nodes the region does not have.

The second consequence is for buyers. Regional data centre operators specify interconnect at design time and rarely revisit it, and the projects being planned now — including the ASEAN partnership specifying 800-volt distribution and liquid cooling — are choosing an optical generation they will live with for years. That decision deserves the attention usually reserved for the accelerator choice, and rarely gets it.

What we could not establish

The revenue split between interconnect and custom silicon. Growth rates for each have been disclosed and the absolute contributions have not, so it is not possible to say how much of the data centre business each represents, which is the number the whole argument would rest on.

We also could not establish several key details: which customers the custom programmes serve beyond Trainium and Maia; the margin difference between the two businesses; whether the raised fiscal 2028 target assumes unannounced design wins; and how much of the interconnect growth comes from price increases versus volume in a tight market.

The figures here are also drawn from analyst commentary published before the results, not from the company's filings. The second quarter of fiscal 2027 is reported on 27 August and will settle several of these questions.

What to watch

The key detail to watch for on 27 August is how the segments are reported. If Marvell breaks out interconnect and custom silicon separately, the argument above becomes checkable rather than inferred.

Watch whether interconnect guidance is raised again. A second upgrade in a year would suggest the AI build-out is running longer and hotter than the industry has been modelling.

Finally, watch when custom silicon revenue actually lands. Fiscal 2028 is the stated year, and a slip would be the clearest signal yet that designing an accelerator for somebody else is harder to convert into revenue than the design-win announcements suggest.

Eva Chin’s Verdict Business

A hundred and eighty-one per cent in a year, ahead of Broadcom and Nvidia, and the part of the company that did it does not compute on anything. It moves data between the chips that do. Marvell is usually discussed as a custom-silicon story, and the revenue says interconnect — optical DSPs, amplifiers, drivers, the pluggable modules that connect data centres. Guidance for that business went from fifty per cent growth to more than seventy inside the same year. The reason is timing rather than technology. A custom accelerator is a multi-year engagement whose revenue arrives when one customer ships in volume. Interconnect sells into every cluster being built, whoever made the silicon inside it. The build-out is in its construction phase, and construction buys the parts that connect things long before it buys designs commissioned two years ago. That also makes interconnect the more diversified of the two businesses, which is rarely how it is described. Put it beside what we have written this fortnight — Kioxia taking an optical SSD to trials because copper limits how far storage can sit from a processor, Cerebras selling latency by keeping weights in on-chip memory, SK hynix committing thirty-eight billion dollars to memory that keeps accelerators fed — and the pattern is hard to miss. Four companies, four layers, one constraint. The accelerators are rarely the limit; getting data to them is. Marvell is the purest financial expression of it yet. One correction to the prevailing story. This does not falsify the ASIC thesis, it dates it: custom silicon grows more than twenty per cent this year and more than doubles next. The design wins are real, the revenue is a fiscal-2028 event. And for this region, optical module assembly and photonics packaging in Malaysia, Thailand and the Philippines sit in exactly the segment growing fastest.

Sources & cross-checks

Primary: Trefis, 14 August 2026 — for the 181 per cent one-year return ahead of Broadcom and Nvidia; the data centre business at 76 per cent of fiscal first-quarter 2027 revenue; trailing twelve-month revenue of US$8.7 billion, up 34 per cent; the fiscal 2028 target of US$16.5 billion, about US$1.5 billion above the prior quarter's target; interconnect expected to grow more than 70 per cent in fiscal 2027 against earlier guidance of 50 per cent; custom silicon expected to grow more than 20 per cent in fiscal 2027 and more than double in fiscal 2028; and the composition of the interconnect business as optical DSPs, transimpedance amplifiers, drivers and pluggable data-centre interconnect modules.
Also reported: TechTimes, 17 August 2026 — for the second-quarter fiscal 2027 results due on 27 August 2026, consensus expectations of roughly US$2.71 billion revenue and non-GAAP earnings near US$0.93 a share, the estimated 20 to 25 per cent share of the custom silicon co-design market, and reported work on Amazon's Trainium and Microsoft's Maia programmes.
Our earlier reporting: The data-movement pattern is Kioxia's optical SSD trials, Cerebras selling latency through on-chip weights and SK hynix's US$38 billion memory commitment; the ASIC thesis is China's dual-track shift to domestic GPUs and proprietary ASICs; the regional specification point is the ASEAN data centre partnership.
What could not be established: ⚠ The revenue split between interconnect and custom silicon — growth rates are disclosed and absolute contributions are not, so the relative size of each within the data centre business cannot be stated. Also unestablished: which customers the custom programmes serve beyond reported work on Trainium and Maia and on what terms; the margin difference between the two businesses; whether the raised fiscal 2028 target assumes design wins not yet announced; and how much interconnect growth is price rather than volume in a tight supply market. ⚠ These figures come from analyst commentary published ahead of results rather than from company filings; the second quarter of fiscal 2027 is reported on 27 August.
Method and limits: Compiled 18 August 2026. The explanation of why interconnect leads custom silicon during a construction phase and is the more diversified business, the placement of this alongside our reporting on storage, memory and inference architecture as one data-movement constraint, the argument that Marvell's figures date rather than falsify the ASIC thesis, and the regional points about optical module assembly and about interconnect being a long-lived design-time decision are RECATOOLS analysis rather than claims by Marvell or any cited analyst. This article is not investment advice.