Marvell (MRVL): what it does, how it makes money, and why it's the #2 in custom AI silicon

The standard $MRVL story is a momentum chart: up about 255% in a year, then down roughly 16% over three months. A fast AI name that ran hard and wobbled. That framing tells you how the stock traded. It tells you almost nothing about what you own.

The more accurate frame: Marvell is a ~$212B chip designer whose thesis now rests on one uncomfortable fact. It is the clear number two in custom AI silicon, directly behind Broadcom, and being the number two in what is effectively a two-name race is a very specific place to stand. This piece walks through what Marvell actually sells, how each leg makes money, where it sits in the AI-compute stack, and the concentration risk underneath the growth. Figures are as of 2026-09-11 unless noted.

Why it matters now

In the three months to 2026-09-11, MRVL gave back about 16% while still sitting up roughly 255% on the year. That combination, a sharp give-back inside a large uptrend, is exactly what makes the evergreen question worth answering carefully: is this an expensive stock that got a little less expensive, or a business that broke? The tape cannot answer it. The answer is in whether the custom-silicon ramp that the multiple is paying for actually converts into shipped chips. Everything below is aimed at that one question.

What does Marvell do?

Marvell hides three different businesses under one ticker. Two of them are the reason the multiple is what it is; the third is ballast.

The first and most important leg is custom AI silicon. When a hyperscaler wants its own accelerator instead of buying an Nvidia GPU, it does not design the whole chip alone. It brings the architecture and hands the hard silicon work to a co-design partner: the IP, the SerDes, the packaging, the manufacturing relationship with $TSM. Marvell is that partner for two named, disclosed programs: Amazon's Trainium line and a Microsoft Maia successor. The revenue on any one program ramps over three to four chip generations, which is what makes these relationships slow and expensive for a customer to unwind. There is talk of a third and even a fourth large customer. Those are not confirmed, and the honest way to hold them is as upside scenarios, not facts.

The second leg is the quiet compounder: coherent-optical DSP. In 2021 Marvell bought Inphi for $10B, and that deal handed it the dominant share of the digital signal processors inside high-speed optical modules, the 400G, 800G, and now 1.6T pluggables that move data between AI servers. The mechanism is simple and durable: as AI clusters scale out, the bandwidth between them explodes, and almost every incremental optical module shipped carries a Marvell DSP. It is a toll booth on AI scale-out rather than a bet on any single accelerator winning.

The third leg is the older DPU and networking SoC business: data-processing units plus legacy switch and connectivity silicon. Real and cash-generative, but not the growth engine. So when you own Marvell, you are really underwriting legs one and two.

How Marvell makes money

Marvell sells silicon, so the model is per-chip pricing across those three segments, weighted by whichever ramps fastest. Trailing revenue is about $9.5B with a net margin near 28% (as of 2026-09-11). The market is not paying near 78x trailing earnings for what Marvell earned last year. It is paying for the custom-silicon ramp it expects over the next several years. That is the entire bet, in one sentence.

The contracts that define the AI leg:

• Amazon Trainium (multi-generation custom accelerator co-design) and a Microsoft Maia successor are the two disclosed large programs.
• A third and possibly fourth hyperscaler program has been discussed but not confirmed. Treat each as upside that widens the base and reduces single-customer risk if and when it is disclosed, not as revenue in hand.
• Inphi (closed 2021, ~$10B) is the optical-DSP franchise: dominant DSP content inside 400G/800G/1.6T pluggables.

The headline risk sits right under the growth: concentration. The top few hyperscalers represent the majority of Marvell's AI-silicon revenue. When a small number of buyers drive the ramp, any one of them pausing or re-architecting a program moves the whole leg. That is the number to hold onto through the rest of this piece.

Where it sits in the AI compute bubble

Marvell is a primary-tier name in QA's AI Compute Accelerators bubble, the cluster of pure-play silicon the AI capex cycle runs through, and it carries a genuine secondary weight in the Networking / Optical bubble. That dual membership is rare and it is the whole point of the company: most names sit in one bucket, Marvell spans the compute and the optical fabric that connects it. It also maps onto the AI Hardware, Photonics, and AI Inference themes, because the custom-XPU case is ultimately an inference-economics bet: at hyperscaler scale, a purpose-built accelerator can deliver materially lower total cost of ownership than a general-purpose GPU for a fixed model.

The useful way to place Marvell is by what it is not. It is not the GPU ($NVDA) and it is not the foundry (TSM). It is the co-design partner and the optical layer, the piece that turns a hyperscaler's architecture into shipped silicon and then wires thousands of those chips together. Its closest peer on the custom-ASIC side is $AVGO, the number one whose shadow defines Marvell's number-two position; for that framing see the Broadcom (AVGO) explainer and, for the deeper GPU-versus-ASIC picture, Nvidia vs custom ASICs: TPU, Trainium, MTIA. Empirically, among MRVL's tightest correlations in the QA universe (252 trading days, market beta stripped out) are Rambus ($RMBS, 0.63), Applied Materials ($AMAT, 0.59), Coherent ($COHR, 0.58), and Arm ($ARM, 0.58): the supply chain that ships when AI capex ships. The full, daily-refreshed peer table is on /stocks/mrvl.

The numbers

Two of those rows deserve a caveat. The trailing P/E is trailing: on a name growing this fast it overstates what you pay for next year's earnings, which is exactly why bulls quote forward multiples and bears quote this one. And the Street mean target is an estimate, an average of analyst opinions, not a promise or a QA view; targets like that get cut quickly when a growth chip misses a single quarter. Report them as what they are and move on.

The bull case

• Two independent growth legs, not one. Custom silicon (Trainium, a Maia successor) and optical DSP each ride AI capex through a different mechanism, so the story does not rest on a single program.
• The optical leg is closer to a toll booth than a bet: Inphi's DSP content rides inside most incremental 400G/800G/1.6T pluggables, so it scales with cluster bandwidth regardless of which accelerator wins.
• Multi-generation co-design relationships are slow and expensive for a hyperscaler to unwind, which gives the custom-silicon backlog more durability than a single design win would.
• Optionality on top: each additional confirmed hyperscaler program widens the base and reduces the concentration risk that is the core bear point.

The bear case

• Concentration. The top few hyperscalers are the majority of AI-silicon revenue; a single paused or re-architected program hits a whole growth leg.
• Marvell is not alone. $AVGO competes directly on both custom ASICs and optical DSP, and the same hyperscalers that are customers today are building in-house design teams for tomorrow.
• The multiple prices a clean ramp. Near 78x trailing earnings and above 22x sales, a single leg printing below cadence compresses the multiple fast; that is the kind of air pocket that produced the recent give-back.
• Hyperscaler capex is the whole cycle. If AI infrastructure spend slows, a custom-silicon order book is one of the first places it shows up.

How to access

Marvell trades on the Nasdaq as MRVL, a clean US listing, so direct ownership needs nothing exotic. To trade it from a US-retail account alongside the rest of the AI-supply-chain names, see /stack/ibkr.

Many diversified investors already own some Marvell without buying the single stock. It is a meaningful holding in the VanEck Semiconductor ETF (SMH), and as a Nasdaq-100 constituent it sits inside QQQ. If you hold a broad semiconductor or large-cap tech fund, Marvell may already be one of your AI-chip exposures; the live holdings breakdown is on /stocks/mrvl.

Bubble-correlation shifts and rule-based alerts on $MRVL, the kind that fire when a name crosses a curated level or its bubble correlation breaks, are part of /pro.

What to watch

• The next quarterly print, read against the custom-silicon ramp: whether AI-silicon revenue is accelerating on schedule is the single cleanest tell. The live cadence and levels are on /stocks/mrvl.
• Confirmation (or not) of a third and fourth hyperscaler program. Each disclosed customer widens the base and directly attacks the concentration bear case.
• Optical attach: whether the Inphi DSP franchise keeps its content share as 800G moves to 1.6T, the leg that scales with bandwidth rather than with any single accelerator.
• The curated $126 level (marked 2026-04-25, far below the recent ~$236 tape) is best read as the observable thesis-break line, the place where the AI-silicon story would be in question, not a trade level or a target.
• A bubble-level shift: if the AI Compute Accelerators bubble breaks its correlation with NVDA and the broader capex tape, the structural read on $MRVL changes from "rides the cycle" to "stands on its own backlog."

Live data on this ticker: /stocks/mrvl. Price, ETF holdings, bubble correlation, curated levels, bot positions.

Bubble context: /bubbles/semiconductors. The cluster this name belongs to and how it's moving.

QuantAbundance is educational research. Nothing here is investment advice. See /disclosures.