Analysis · Optics

Is Marvell (MRVL) stock a buy right now? The case, and the one thing that breaks it

7 August 2026 · 11 min read

The short version: in the six months to May 2026 Marvell raised its fiscal 2028 revenue target twice - from around $15 billion in March to around $16.5 billion in May - closed a $5.5 billion acquisition that fills the one hole in its product line, and took a $2 billion investment from Nvidia. Over roughly the same window the stock fell to about a third below its 52-week high, on a sector-wide selloff that had nothing to do with any of it.

That gap between what changed in the business and what happened to the price is the entire argument. It is not an argument that the stock is cheap. It very clearly is not, and this piece says so plainly further down.

What actually changed in the business

Start with the reported numbers, because the rest of this rests on them.

In the first quarter of fiscal 2027, reported 27 May 2026, Marvell posted:

Net revenue of $2.418 billion, a record first quarter, up 28% year on year
Data centre revenue of $1.833 billion, up 27% year on year - roughly three-quarters of the whole company
Non-GAAP earnings of $0.80 per diluted share, on 58.9% non-GAAP gross margin
Record operating cash flow of $638.8 million

Guidance for the July quarter was $2.7 billion at the midpoint, which would be 35% year-on-year growth - accelerating, not decelerating. Chief executive Matt Murphy described "exceptional AI-related bookings" across custom XPU, XPU-attach, optical and switching.

The guidance track matters more than any single quarter. In March 2026 Marvell told the market to expect fiscal 2028 revenue approaching $15 billion, against a consensus of $12.92 billion. Eleven weeks later it raised that to roughly $16.5 billion, with fiscal 2027 at about $11.5 billion. Companies that raise a two-year target by $1.5 billion in a single quarter are usually reacting to signed business, not enthusiasm - and Marvell backed it by disclosing around $1 billion of prepayments to secure manufacturing capacity. You do not prepay for wafers you are not confident you will fill.

The July selloff was not about Marvell

Then the sector fell over. Chip stocks shed more than $1 trillion in market value across late July, on three things: SK Hynix delaying HBM4 expansion in favour of higher-margin DDR5, which investors read as softening AI memory demand; renewed worry about the debt financing behind AI datacentre buildouts; and fresh signs of progress in Chinese domestic chipmaking.

Read that list again and notice what is not on it. Nothing there concerns interconnect demand, Marvell's order book, or its guidance. HBM allocation is a memory-pricing story - we wrote about the Chinese domestic chip question separately, and the short answer there was that it is a regional substitution story rather than a global demand one.

Multiple compression on unchanged-or-rising estimates is the most common way a good business becomes an interesting entry point. It is also, sometimes, the market seeing something first. Both are true often enough that the only useful response is to look at what specifically would have to be wrong.

Three structural things worth owning

1. Custom silicon, on someone else's capex budget

Every hyperscaler now wants its own accelerator, and almost none of them want to build a full silicon team to get one. Marvell sells into that gap. At its custom AI investor event it put the 2028 data centre semiconductor market at $94 billion, of which $55.4 billion is custom accelerated compute and $40.8 billion is custom XPU alone, growing at a 47% compound rate. It disclosed engagement on 18 custom projects - 12 devices across four major hyperscalers and six for emerging AI customers.

The important structural feature: this revenue is not a bet on which model wins or which cloud grows fastest. It is a toll on the act of building custom compute at all.

2. Celestial AI closes the one real gap

Marvell announced the acquisition of Celestial AI on 2 December 2025 and completed it on 2 February 2026 - $1 billion cash plus about 27.2 million shares upfront, with up to 27.2 million more shares on revenue milestones, for a headline value of roughly $5.5 billion.

What it buys is Photonic Fabric: optical interconnect at the package, system and rack level, claimed at twice the power efficiency of copper with nanosecond latency and 16 terabits per second of bandwidth. The strategic logic is the same physics that drove an entire ETF launch - copper runs out of road past a couple of metres at modern data rates, and scale-up domains are getting bigger than a rack. Data centre president Sandeep Bharathi framed it as combining Marvell's UALink scale-up switch roadmap with optical scale-up so customers can "build AI systems that scale beyond copper limits."

Marvell expects meaningful revenue from it in the second half of fiscal 2028. That is a long way out, which is precisely why it is available at today's price rather than priced in.

3. Nvidia paid $2 billion to keep Marvell inside its ecosystem

On 31 March 2026 Nvidia invested $2 billion in Marvell and brought it into NVLink Fusion, letting Marvell's custom XPUs and scale-up networking sit natively alongside Vera CPUs, ConnectX NICs, BlueField DPUs and Spectrum-X switches, with joint work on silicon photonics.

This is the detail I find most underrated, and it cuts in an unusual direction. Marvell's custom business exists largely to help hyperscalers buy fewer Nvidia GPUs. Nvidia's response was not to fight it but to fund it and wire it into its own rack standard. For a Marvell shareholder the effect is that the company now gets paid on both roads: if the hyperscaler buys GPUs, Marvell sells the optics and the attach silicon; if it builds its own XPU, Marvell may well design it. Fewer forks in the road are fatal.

The business that funds all of it

Underneath the custom story sits an optical DSP franchise that is already at scale and already generating cash. Marvell shipped the first 200G-per-lane 1.6T DSPs in 5nm in 2023, followed with the 3nm Ara platform in 2024, and Ara is now shipping in mass volume. On 12 March 2026 it extended the portfolio with Ara X, Ara T - the first 8x200G transmit-retimed optics DSP - plus the Petra gearbox and the Aquila M coherent-lite DSP, all sampling from the first quarter of 2026.

The reason this matters to the investment case is timing. The custom and photonics stories pay out in fiscal 2028 and beyond. The DSP business pays now, and it is what makes the wait fundable. See the full optical group for how the rest of that supply chain is trading, or the MRVL vs CRDO comparison for the closest listed pure-play.

Now the honest part: this is not a cheap stock

Marvell trades at roughly 47.6x forward earnings and about 74.6x trailing, on trailing twelve-month revenue of $8.72 billion growing 34% (figures as of 7 August 2026). Broadcom, running the same playbook at ten times the size, trades near 26.7x forward.

So the comparison people reach for - "Marvell is the cheap way to own custom AI silicon" - is simply wrong on forward earnings. You are paying about 1.8 times Broadcom's multiple for a company with a fraction of its scale and a much more concentrated customer list. What you get for the premium is a growth rate roughly twice as fast off a base small enough that a single socket win moves the whole company, which is a real thing to want, and a real thing to overpay for. The side-by-side is worth a look before deciding which end of that trade appeals.

Analyst consensus sits at Strong Buy with an average twelve-month target of $256.91 across 43 analysts, with individual targets spanning $126 to $400. A low estimate roughly half the high estimate is the most honest summary of this stock available: nobody actually knows, and the spread reflects a genuinely binary set of outcomes rather than ordinary forecasting noise.

The bear case I take seriously

One risk dominates all the others, and it is not valuation.

Socket concentration. Marvell's custom revenue is anchored on a small number of hyperscaler programmes. There have been persistent reports that next-generation Amazon Trainium work moved to Alchip, and separately that Microsoft has held discussions with Broadcom about custom design work. Neither has been confirmed by Marvell or by the customers, and both should be treated as unverified reporting rather than fact. But the shape of the risk is not in dispute: when five customers can each individually reprice your two-year outlook, one lost re-spin is not a bad quarter, it is a thesis change. That is why the stock carries the volatility it does, and no amount of TAM arithmetic makes it go away.

Three more, in order of how much they should bother you:

Broadcom's incumbency. Broadcom holds the large majority of the custom AI accelerator design business. Marvell's pitch is that it is the more flexible partner, letting customers keep more control of their own design. That is a positioning claim, and positioning claims get tested every design cycle.
Hyperscaler capex is the whole floor. There is no defensive segment left to cushion a downturn - this is now a data-centre company. If capex growth slows, interconnect and custom silicon do not decouple from it.
The Celestial earnout is share-funded. Up to 27.2 million additional shares on top of the 27.2 million already issued. The dilution arrives precisely if the acquisition works, which is a fair structure, but it is real.

What I actually think

Marvell is the most interesting risk-adjusted setup on our board right now, and I would not put it in the "safe" half of a portfolio.

The specific reason to be interested at this price is narrow and I want to state it precisely: the derating happened for reasons external to the company, during the same months in which the company twice told the market its two-year revenue would be materially higher than it had said before, and backed that with a billion dollars of capacity prepayments. That is not a valuation call. It is an observation that price and information moved in opposite directions, which is the only kind of setup worth acting on with any urgency.

The thing that would change my mind is equally specific. Marvell reports on 27 August 2026. The number I care about is not the quarter - the quarter is effectively pre-announced by the guide - it is whether the fiscal 2028 target holds at around $16.5 billion or better. A raise means the bookings were real. A reiteration is neutral and the stock probably drifts. A cut, or any softening of language around a major custom programme, retires the thesis on the spot, because it would mean the socket risk above stopped being hypothetical. I would rather be wrong quickly on a stated test than slowly on a vibe.

One practical note: this is a name where position size does the work, not conviction. A 47x forward multiple with binary socket risk deserves a smaller allocation than the write-up sounds like it deserves.

How the board reads it today

Sentiment on ChipSentiment is a behaviour score, not a valuation one - it reflects momentum, relative strength against the group, trend, volume, volatility and drawdown. A stock recovering off a sector selloff will score differently from a stock making new highs, and neither tells you whether the business is any good. The full methodology is here.

Live positions for the names in this piece:

Full MRVL page - live price, sentiment and history →
All AI chip stocks →
Today's board across all 34 names →

Information and analysis only, not investment advice, and it takes no account of your circumstances. Figures are drawn from the company filings, press releases and reporting linked below, and were accurate as at 7 August 2026. Statements about unconfirmed customer programmes are identified as such above and should not be read as fact. The author holds no position in any security mentioned.

Frequently asked questions

Is Marvell (MRVL) stock a buy right now?

That depends on your tolerance for concentration risk. The case for it is that Marvell raised its fiscal 2028 revenue target from around $15 billion in March 2026 to around $16.5 billion in May 2026 and backed it with roughly $1 billion of manufacturing capacity prepayments, while the share price fell with the broader July 2026 chip selloff for reasons unrelated to its own order book. The case against is that it trades near 47.6x forward earnings against Broadcom's 26.7x, and its custom silicon revenue depends on a small number of hyperscaler programmes.

Why did Marvell stock fall in July 2026?

It fell with the sector rather than on company-specific news. Chip stocks lost more than $1 trillion in market value in late July 2026, driven by SK Hynix delaying HBM4 expansion in favour of DDR5, concerns about the debt financing behind AI datacentre buildouts, and signs of progress in Chinese domestic chipmaking. None of those relate directly to interconnect demand or to Marvell's guidance, which had been raised in March and again in May.

What did Marvell buy Celestial AI for?

Marvell announced the acquisition on 2 December 2025 and completed it on 2 February 2026. The structure was $1 billion in cash plus approximately 27.2 million shares upfront, with up to 27.2 million further shares payable on revenue milestones, giving a headline value of about $5.5 billion. It brings Celestial AI's Photonic Fabric optical interconnect technology, which Marvell expects to contribute meaningful revenue from the second half of fiscal 2028.

Why did Nvidia invest $2 billion in Marvell?

On 31 March 2026 Nvidia invested $2 billion in Marvell and brought it into the NVLink Fusion ecosystem, allowing Marvell's custom XPUs and scale-up networking to work natively with Nvidia's Vera CPUs, ConnectX NICs, BlueField DPUs and Spectrum-X switches, alongside joint silicon photonics work. The practical effect is that Marvell participates whether a customer buys Nvidia GPUs or builds its own accelerator.

Is Marvell or Broadcom the better AI stock?

They are different trades. Broadcom is larger, holds the majority of the custom AI accelerator design business, and trades near 26.7x forward earnings. Marvell trades near 47.6x forward but is growing faster off a much smaller base, so individual design wins move it far more. Broadcom is the lower-variance way to own the theme; Marvell is the higher-variance one.

What are the biggest risks to Marvell stock?

Customer concentration is the main one - custom silicon revenue rests on a small number of hyperscaler programmes, and there has been unconfirmed reporting about next-generation Amazon Trainium work moving to Alchip and about Microsoft discussing custom design work with Broadcom. Beyond that: Broadcom's incumbency in custom ASIC design, complete dependence on hyperscaler capital expenditure with no defensive segment left, the valuation, and share dilution from the Celestial AI earnout.

When does Marvell next report earnings?

Marvell is scheduled to report second-quarter fiscal 2027 results on 27 August 2026. The quarter itself is largely set by the $2.7 billion midpoint guide; the number that matters more is whether the roughly $16.5 billion fiscal 2028 revenue target is maintained, raised or cut.