Marvell Technology

IThe business

Marvell designs semiconductors and does not build them. TSMC and other contract factories make the wafers, and assembly and test partners finish them.⁶ The company had 7,480 employees at the end of FY2026, about 49% of them in the Americas and 42% in Asia-Pacific.⁶ It sells chips that move, store and process data, and about three quarters of revenue now comes from data centers.² The AI buildout is why. Revenue was $9.45B in the twelve months to 1 August 2026, up from $5.77B in FY2025 and $5.51B in FY2024 C (calculated).⁶¹⁴²

Fiscal years end on the Saturday nearest 31 January, so FY2027 ends on 30 January 2027. TTM on this page means the four quarters to 1 August 2026. The October quarter closed on 31 October. On 2 November Marvell kept the guided midpoint of $3.15B and narrowed the range to plus or minus 2%.⁷ Full results are expected in early December, so I use the July quarter, and those results will change some numbers. I would read them beside this page and not assume this page has them.

What it sells

Marvell no longer reports revenue by product. It reports two end markets, data center and communications and other, and describes products in the 10-K.⁶ Four groups matter.

• Custom chips. A cloud company wants its own AI processor, called an XPU, or a chip that sits next to one. Marvell builds the design on its own library of high-speed circuits and builds the chip with TSMC. It has done 5 nanometer designs, is on 3 nanometer now, and is developing 2 nanometer.⁶ Management has put custom revenue at about $1.5B in FY2026 and has said it will more than double in FY2028 and reach $10B or more in FY2029.¹⁹⁴ The Google agreement of July 2026 covers inference accelerators, storage controllers, network interface controllers, memory interface controllers and near-memory compute chips.²⁰
• Electro-optics and interconnect. Light carries data between racks. At each end a small chip turns electrical signals into optical ones and back, and it has to correct errors at very high speed. Marvell makes those digital signal processors, plus the drivers and amplifiers beside them, active copper cable chips and PCIe retimers.⁶ Management says optical DSPs, broadband analog chips and scale-out switching are each on or ahead of a path to $1B of annualized revenue.⁴
• Switching. Ethernet switch chips, the Teralynx line for cloud data centers, plus new switches for linking chips inside a rack, called scale-up, from the XConn deal.⁶²
• Storage and the rest. Controllers for hard drives and SSDs, Fibre Channel adapters, security and network processors.⁶ This is the older business. It sits mostly in communications and other, which was $2.28B in the twelve months to August 2026 against $3.51B in FY2023.²¹²

How it got here

In its fiscal 2017 10-K Marvell, then a Bermuda company, described its products as storage, networking and connectivity chips, with hard drive controllers listed first.¹⁵ Mr. Murphy joined as chief executive in July 2016.¹⁰ Since then it has bought its way into data centers: Cavium, Aquantia, Inphi, Innovium, and in 2026 Celestial AI and XConn.⁶¹⁶ Inphi closed on 20 April 2021 and brought optical DSPs; Innovium followed on 5 October 2021 and brought data center switching.¹⁶ The company became a Delaware company in fiscal 2022.¹⁶ It sold the automotive ethernet business to Infineon for $2.5B in cash on 14 August 2025 and booked a $1.8B pre-tax gain.⁶ From FY2019 to the first half of FY2027 it paid about $8.7B in cash for acquisitions, besides the shares it issued C (calculated).⁵¹⁸² Data center was 35% of revenue in FY2021 and is 76% now.²¹²

Celestial cost $3.5B: $1.28B in cash, 24.5M shares valued at $1.93B, and a $316M earnout at the time of closing.² The earnout is payable in cash and shares through FY2029 if revenue milestones are met. By 1 August its fair value had risen to $749.5M, and $433.7M of that rise went through first-half expense as the share price climbed.² A cash-settled forward contract, set up in April, hedges part of the stock price exposure and was an asset of $131.0M.² I count the net $0.6B as debt.

Who pays

Few buyers pay most of the bills. In the first half of FY2027 one distributor was 45% of revenue, up from 35% a year earlier, and one direct customer was 16%.² A distributor resells to many end customers, so that 45% is not one buyer, but the 10-Q does not show who is behind it. The 10-K says the ten largest customers, distributors included, were 82% of FY2026 revenue.⁶ Four customers were 72% of gross receivables on 1 August.² Shipments to China were 42% of Q2 revenue, up from 29% a year before, though the 10-Q says most of that goes to non-Chinese customers with factories there.²

Who the large buyers are comes from other places. In December 2024 Marvell announced a five-year, multi-generational agreement with Amazon Web Services covering custom AI chips, optical and copper cable DSPs, retimers, data center interconnect modules and Ethernet switching chips.²³ On 29 July 2026 it signed the Google agreement and issued Google a warrant, described below.²⁰ The 2026 proxy says the company has “design wins at all four” large US cloud companies.¹⁰ Press coverage of the October investor day names Amazon and Google as cloud companies for which Marvell designs processors.²⁴ I found no filing or release in which Marvell ties Microsoft to a custom chip program. Microsoft appears in Marvell’s October releases on hardware security and in a payments product announced in September, which are different things.²⁵ I treat Amazon and Google as documented and Microsoft as reported by others.

Design wins are not revenue

This is the part that is easy to get wrong. A custom chip program has three stages. The customer picks a design partner, which is a win. Then the partner spends one to three years designing and getting the chip made, and is often paid something toward that cost. Only after that do volume shipments begin, and the customer’s own demand sets how many. The 10-Q says plainly that “the development process for our products is long” and that there is a delay between the cost and the revenue.² The company says it had record design wins in FY2026.¹⁰ Revenue from them is mostly in the future.

The Google deal shows the gap. Murphy said on the August call that the main revenue effect would arrive in FY2029 and later, and that revenue from the programs through FY2028 is already in the company’s custom target.⁴ The warrant ties payment to actual purchases, not to the win. Google received a warrant for 58,970,907 Marvell shares at $206.58. A small part, 1,360,867 shares, vests over a year. The rest vests in 240 equal tranches, one for each $500M of custom products revenue from the October quarter of FY2027 through FY2033.²⁰ That is $120B of revenue if every tranche vests, and I will not say Marvell expects it. Two things about it are clear. Each tranche pays Google about 240,000 shares for $500M of purchases, which is a rebate in stock. And the shares only have value to Google above $206.58, so Marvell is paying in a currency that costs it nothing if the stock does not rise.

Two older warrants went to customers in fiscal 2025 and 2026, for 4.2M shares at $87.77 and 1.0M shares at $87.00, vesting on revenue milestones. About 1.2M of the first had vested by August.² Accounting treats vesting as a reduction of revenue, so reported revenue is already below invoiced revenue by that amount.²

The Nvidia investment

On 31 March 2026 Nvidia paid $2.0B for 2.0M shares of convertible preferred stock, convertible into about 21.8M common shares at about $91.84, alongside a partnership to connect Marvell’s custom chips to Nvidia’s ecosystem.² It has not converted, and I count it in the share count. At today’s price the stake is worth about $3.5B C (calculated). The conversion price is 43% below today’s close.

IIThe moat

A good test of protection is what a buyer does when a cheaper option appears. Here the buyers are the largest companies on earth and they have plenty of options. They can design in house, use another design partner such as Alchip, GUC or MediaTek, or stay with Broadcom, all of which the 10-K lists as competitors.⁶ So the question is why they would stay with Marvell, and what they pay for it.

The gross margin answer is mixed. GAAP gross margin was 52.2% in the twelve months to August and 53.1% in Q2.³ It was 60.7% in FY2018, 41.3% in FY2025 and has been below 52% in eight of the ten fiscal years C (calculated). Non-GAAP gross margin, which leaves out amortization of acquired intangibles and stock pay in cost of goods, was 58.9% in Q2 against 59.4% a year earlier, and the October quarter is guided to 57.5% to 58.5%.³ Management attributes the slip to a mix shift toward custom, and expects a similar range in FY2028.⁴ That is how a supplier looks when it grows by selling more to its largest buyers: revenue rises faster than margin.

Compare the neighbours. My Broadcom page has $42.8B of operating income on $89.1B of revenue, 48%.¹¹ My NVIDIA page has $197.6B on $303.0B, 65%.²⁶ Marvell’s non-GAAP operating margin of 36.6% is respectable, and its GAAP margin of 16.8% is about a third of Broadcom’s.³ That is less a verdict on quality than a reminder of what Marvell has been doing: it bought its position, and the accounts carry the cost.

Evidence it is widening

Data center revenue grew 46% in Q2 and the October quarter is guided to about 75% growth in data center, a pace that does not happen if buyers are shopping around.³⁴ Gross margin on that growth has stayed near 59% non-GAAP. The agreements are getting longer: five years with AWS in 2024, and a Google arrangement whose warrant runs to FY2033.²³²⁰ And Marvell would not offer a warrant of about 6% of its share count unless it expected a lot of volume. I read it as a sign of how much the two sides discussed.

Evidence it is narrowing

The first sign is where Marvell ranks. Counterpoint expects Broadcom to take about 60% of the design-partner share of AI server compute chips in 2027 and Marvell 8%, one research house’s forecast quoted on my Broadcom page.¹¹ Broadcom’s AI chip revenue was $16.7B in its latest quarter. Marvell’s whole data center line was $2.17B.¹¹³ Marvell’s custom business is smaller than Broadcom’s by a large multiple, and it has to win generations to close that gap.

The second sign is how fragile the information is. In December 2025 an analyst at Benchmark Capital downgraded Marvell on a report that it had lost designs to Alchip and a large US cloud company. The analyst corrected that on 23 December and upgraded the stock on 8 March 2026, according to the board’s statement in the 2026 proxy.¹⁰ I take the correction as the board reports it, since I have not seen the notes. What I take from the episode is that a stock this large can fall on a socket rumor that turns out to be wrong, and that a real loss would look the same on the first day.

The third is price. Giving Google a warrant for about 6% of the share count is one way of paying for volume.²⁰ I do not know what the discount is on the chips themselves, and no one outside the contracts does. Hyperscalers buy in size and bargain accordingly. I show the cost to the owner as a deduction from revenue in section VI.

So I would describe the protection as strong inside programs that have won, and thin at the point where new ones are chosen. It protects margin on what Marvell already ships. It does not protect share at the next round.

IIIThe competitive advantage

Four kinds of company matter. Broadcom is the leader in custom AI chips. Nvidia sells the whole system, and now also connects to Marvell’s. Other design partners, such as Alchip, GUC and MediaTek, take custom work from the same buyers. In optics, Marvell competes with the module makers who are its customers, and with specialists such as Credo, Astera Labs and MACOM, all of which the 10-K names.⁶ The Lumentum merger changes the last of these.

Where Marvell wins

It wins on breadth and on being the second choice. A cloud company that wants a custom chip from someone other than Broadcom has few places to go, and Marvell sells the parts around the chip too: the optical engine that carries its data, the retimer that cleans its signal, the switch that links it to others. A single customer relationship can span the rack, which is what the AWS and Google agreements describe.²³²⁰ Its circuit library is also hard to copy. Its expenses are modest: non-GAAP operating expenses were $611M in Q2, or 22% of revenue, and management plans for FY2028 expenses to grow about half as fast as revenue.³⁴

Where it loses

It loses on scale. Broadcom’s AI chip revenue in one quarter is 7.7 times Marvell’s whole data center revenue in the same period C (calculated).¹¹³ Scale lets a design partner spread engineering across more programs and buy wafers and packaging in larger blocks. Marvell has just committed $8.5B to foundries and test partners, and made roughly $1B of capacity prepayments this fiscal year, so it is putting real money behind a small share.²⁴ It also loses on independence. Its largest customers can bring design in house, and some do.⁶ And with Lumentum it takes on a new problem: Lumentum sells lasers to module makers who also buy Marvell’s DSPs, and some of them may start to see Marvell as a rival. Press coverage describes Lumentum itself as a customer for Marvell’s DSPs. I could not confirm that in a filing, so I treat it as a risk and not a fact.

Finally, it loses on perception of reliability. The Benchmark episode is the clearest example.¹⁰ Broadcom’s lead means that when a cloud company diversifies, it is Broadcom’s share that is at risk, and Marvell’s that benefits. When it consolidates, the reverse. Neither move shows up in a quarter.

IVManagement

Mr. Murphy has run Marvell for ten years. The share price went from $12.52 in November 2016 to $159.80, 12.8 times, with a high of $316.43 in June.¹ Revenue grew from $2.30B in FY2017 to $8.19B in FY2026 C (calculated).¹⁵⁶ He bought Cavium, Aquantia, Inphi and Innovium, sold the automotive business at a $1.8B gain, and pushed the company from storage and networking chips into AI.¹⁰⁶ Those were big and mostly correct strategic calls. What the numbers do not show is a decade of earnings. GAAP net income summed over the ten years to FY2026 is $1.9B C (calculated).⁵¹⁸

The finance chief changed in June. Willem Meintjes, in the job since January 2023, resigned effective 15 June with no stated disagreement, and the board named one of its own directors, Daniel Durn, who had just left Adobe as its CFO.³⁰ The audit committee chair became CFO within five days of resigning from the board. He is an experienced finance chief, and the 8-K gives no reason for the change beyond Mr. Meintjes’s decision. I take no view; I note it.

How they are paid, and what that pays them to do

Mr. Murphy’s total pay was $25.06M for FY2026, of which $1.24M was salary, $3.62M annual bonus and $20.2M stock awards. That compares with $32.16M in FY2025 and $45.16M in FY2024. The ratio to the median employee’s $158,857 was 158 to 1.¹⁰ Two officers received much more in FY2026: Mr. Bharathi $48.4M and Mr. Koopmans $37.7M, nearly all of it stock awards. The proxy says they were made when both became presidents on 15 July 2025, and include units that vest on shareholder return over a five-year period.¹⁰

The annual bonus in FY2026 weighted revenue 50%, non-GAAP gross margin 15% and non-GAAP operating margin 35%, and paid 144.84% of target.¹⁰ The long-term stock units for the officers vest on Marvell’s total shareholder return against the S&P 500 over three years, capped at 200%, with a multiplier of up to 150% for non-GAAP earnings per share growth against a peer group, and an overall cap of 250%. If absolute return is negative, the payout cannot exceed 100% of target.¹⁰ The units from April 2022 paid 84% of target, because the three-year return of 30.3% was below the median.¹⁰ So the design has teeth, and I give it credit for that.

What it pays them to do is worth saying plainly. Almost every measure is non-GAAP, which leaves out stock pay and the amortization of acquired intangibles. Revenue carries half of the bonus, and acquired revenue counts like organic revenue. Nothing rewards cash after stock pay or return on the capital spent on deals, other than a relative return measure over three years. A leader paid this way will find it natural to buy revenue in an AI cycle, and the Lumentum deal, which the companies say will lift non-GAAP earnings per share in the first year, is the kind of event the plan counts as a success.⁷ I am not saying that is why it was done. I am saying the plan does not push the other way.

Shareholders were less pleased. Support for the pay of the named officers fell from about 83% in 2025 to 55.9% of the votes cast for or against in June 2026: 327.6M for, 258.6M against.¹⁰¹² A shareholder proposal for an independent chair drew 215.9M votes for and 369.2M against, 36.9% in favour. The board opposed it and wrote that the combined chair and chief executive role is working.¹⁰¹²

How much of their own money is in

Mr. Murphy beneficially owned 412,871 shares at 31 January 2026, about $66M at $159.80 C (calculated).¹⁰ The guideline is six times salary for the chief executive, or about $7.4M.¹⁰ He is well past it, and it is small beside the $20.2M of stock he is granted in one year. The largest holders are funds: FMR, the parent of Fidelity, held 126.7M shares, 14.95%, at 31 December 2025, about $20.2B at the current price; Vanguard held 79.6M, 9.4%, at 30 September 2025; and BlackRock’s last report in the proxy, from 31 December 2023, was 60.5M, 7.1%.¹⁰

The officers are selling. Form 4 filings for the twelve months to 2 November show seven insiders sold 0.37M shares for about $57.0M in open-market sales, with no purchases by anyone.³¹

These are modest sums for a company this size, and an executive who sells after a thirteenfold decade is not sending a warning. The point is narrower. The people who know the most sold at $112 to $281 in the past year, and none bought at $49 in April 2025 or at $73.73 in February.³¹¹

What they did with capital over a full cycle

Marvell does not make much cash beyond what it spends on staying in place, so capital allocation is mostly acquisitions, and acquisitions have been the main use. From FY2019 to the first half of FY2027 it paid about $8.7B in cash for deals, issued shares on top, and ended with $13.9B of goodwill and $2.3B of acquired intangibles, together 59% of total assets.⁵¹⁸² Over the ten fiscal years to FY2026, cash after capital spending and stock pay totalled $3.5B and GAAP net income $1.9B C (calculated).⁵¹⁸ So the cash paid for deals in the past eight fiscal years is more than twice what the whole business generated in ten. Whether that was wise depends on the revenue those deals brought, and the data center growth says much of it was. But the shareholders paid for it twice, once in cash and once in shares.

Diluted shares were 509.7M in FY2018 and 921.2M in the July quarter, up 81%. Part of that is Inphi, Innovium, Celestial and XConn paid in stock, part is stock pay, and part is the Nvidia preferred, now counted as if converted.⁵¹⁸² If the Lumentum merger closes, the 4.5-for-1 exchange adds about 454M shares, and the share count is about 1,375M E (estimate).⁷ I computed that from the stated 67% and 33% split and Marvell’s 921M, not from a company figure.

The buybacks tell a different story, and a better one.

Marvell bought stock cheaply in 2025 and early 2026, including a $1.0B accelerated repurchase that settled for 11.7M shares in the fourth quarter of FY2026.⁶ That was good buying. Three things temper it. First, the buying stopped being cheap: the first-half FY2027 purchases were at $160. Second, over the same stretch diluted shares rose 8.2%, from 851M to 921M, because the company issued far more than it retired. Third, the buyback is funded in part by borrowing: debt rose from $4.5B to $5.0B in the first half of FY2027 while cash went up, and the new $3B revolver, signed on 22 October, adds room that can be used for deals.²³² On the 67% and 33% split in the 2 November release, Lumentum holders would receive about 454M new shares, by my arithmetic.⁷ The record is: good buyer of its own stock, heavy issuer of it.

VFinancials

Revenue was $2.74B in the July quarter, up 36.5%. GAAP operating income was $460M and GAAP earnings per share $0.33; non-GAAP operating income was $1.00B and non-GAAP earnings per share $0.94.³ The October quarter is guided to $3.15B, a non-GAAP gross margin of 57.5% to 58.5%, non-GAAP operating expenses of about $655M and non-GAAP earnings per share of $1.10, plus or minus $0.05.³ Third-party transcripts of the call put FY2027 revenue near $12B, which needs about $3.7B in the January quarter, and FY2028 near $18B, before the October investor day moved the target to about $20B.⁴²⁴

Look at the gap between the two sets of earnings. Non-GAAP net income in the first half of FY2027 was $1.58B and GAAP net income $0.34B.³ Three items explain nearly all of it: $534M of stock pay, $440M of amortization of acquired intangibles, and a $303M charge from the Celestial earnout net of its hedge.³ The first is a cost I count. The second is the leftover of past deals, which is non-cash and so left out of my owner cash. The third is the share price. When Marvell’s stock rises, the company books a loss on what it may owe Celestial’s former owners. That is a real obligation, and it is in my net debt.

What an owner could really take out

Reported free cash flow over the past year was $1.73B, operating cash flow less capital spending C (calculated).²⁶ That is before stock pay. After it, the figure is $0.90B. The business as it now stands earns more than that, but it is spending cash on growth, mostly in working capital. Here is the walk.

So the number I would call owner cash is $2.1B, not the $1.7B of reported free cash flow, and not the $2.6B of net income, which includes the automotive gain. The two figures, $0.9B and $2.1B, are both honest, and they measure different things. The first is what the company kept. The second is what the business earned before it fed its own growth. Four items belong in the gap, and none is small.

• Supplier prepayments. Prepayments for supply capacity were $487.0M on 1 August against $278.8M in January, and management plans about $1B for the year, to be credited against future purchases.²⁴ Cash goes out now and comes back as lower cost later, if the chips are bought.
• Factoring. Marvell sold $412.0M of trade receivables without recourse in the first half of FY2027.² That brings cash in early and flatters operating cash flow.
• Taxes withheld on stock. It paid $365.2M in cash in the half to cover employees’ taxes on vested shares, instead of issuing the shares. That sits in financing, not operating cash flow, and is a cost of stock pay on top of the expense.²
• Tax on the automotive sale. The 10-Q lists income tax related to the sale as a use of cash in the next twelve months. I did not find the amount, and cash taxes paid in FY2026 were only $92.1M.²⁶

Stock pay is the part I can see most clearly. It was 7.6% of revenue in the first half of FY2026 and 10.4% in the first half of FY2027, and 11.9% in Q2.³ Celestial and XConn brought $205M of unrecognized stock awards between them at closing.² If Q2’s pace holds, the past twelve months of owner cash would be about $0.5B lower. That is why I model stock pay as a falling share of revenue, from 7.9% in FY2028 to 6.5% later, and why the base case needs the revenue to come.

Quality of earnings

GAAP net income of $2.64B over the past year holds a $1.8B pre-tax gain on the automotive sale, taxed at a rate I could not isolate.⁶ Take it out and the business earned a fraction of that. Non-GAAP earnings per share of about $3.30 for the same twelve months, FY2026’s $2.84 plus $1.75 less $1.29, are the better guide to operating profit. Stock pay at the Q2 pace is $1.3B a year, about $1.42 a share on 921M shares, and non-GAAP leaves it out C (calculated).³

Two more points. The non-GAAP tax rate is 11%, helped by government incentives that depend on keeping up hiring and investment, and the 10-Q says a global minimum tax of 15% applies to most of the company’s countries.² I use 12% rising to 15%. And customer warrants reduce reported revenue as they vest, a quiet and non-cash price cut.

The balance sheet

Marvell had $3.93B of cash and $5.0B of debt on 1 August, so net debt was about $1.0B, plus the earnout liability.² Debt is a ladder of senior notes from 2028 to 2036 with coupons from 2.45% to 5.95%, and a $3B revolver, undrawn, signed on 22 October.²³² Interest expense was $61.6M in the quarter.³ The leverage is not the problem.

What is off the balance sheet is. Marvell has unconditional purchase commitments of $8.52B to foundry, test and assembly partners, $1.83B due in the rest of FY2027 and about $2.1B to $2.2B in each of FY2028, FY2029 and FY2030, plus $666M of technology services and licences and $352M of capital commitments.² Against $9.45B of trailing revenue those commitments are 0.9 times revenue. Inventory is $1.36B. Foundries allow orders to be cancelled if Marvell pays what the foundry has spent, and may charge more fees or take away priority.² In a boom this is how a fabless company secures supply. In a pause it is a bill.

Cash is not covering the next step either. Marvell paid $1.27B in cash for Celestial and XConn, drew $1.0B of new notes, raised $2.0B from Nvidia, and repaid $0.5B of notes in the half. The Lumentum merger is all stock, so the price needs no cash. Lumentum’s own debt, including convertible notes, would come with it or be paid off, and I have not sized that. The termination terms cost cash too: Marvell owes Lumentum $3.2B if regulators block the deal, and $5.8B if its board changes its recommendation or it signs a rival deal.²⁸

The stock moves like the cycle it sells into. Over ten years there were five falls of more than 30% on daily closes: 41% from March to December 2018, 40% in early 2020, 62% from December 2021 to January 2023, 61% from January to April 2025, and now 49.5% from the $316.43 high of 4 June 2026.¹ It fell 19.1% on 27 January 2025 and 16.7% on 5 June 2026, the day after the record, and rose 32.5% on 2 June.¹ It closed at $73.73 on 4 February 2026 and $49.38 on 21 April 2025.¹

VIValuation

First, how this goes badly wrong

Before any numbers I asked what would most likely destroy this investment, and worked back. The business is sound, with modest debt. Ruin would come from the buyers, the paper and the price. Three ways stand out.

• The cloud companies pause with the parts already ordered, or move a generation of chips. Four large cloud companies spent $496.7B on capital goods in the first nine months of 2026, a pace that cannot compound forever.²⁷ Marvell has promised $8.5B to foundries and test partners and prepaid $487M of capacity, and the foundries allow cancellation only if Marvell pays for work done.² In the last downturn, revenue fell 7% in FY2024, data center fell 8% to $2.22B, and the stock lost 62% from December 2021 to January 2023.²²¹ A pause after a build this large hits revenue, margin and inventory together.
• The paper. On 2 November Marvell agreed to pay for Lumentum in stock, about 454M new shares, at a 23.9% premium, five months after its own stock closed at $316. It would fold an indium phosphide laser maker, a factory business, into a fabless company. Press coverage describes Lumentum as a customer for Marvell’s optical DSPs, which I could not confirm in a filing, and module makers who buy Marvell DSPs may start to treat Marvell as a competitor. Regulators must approve the deal, and Marvell owes Lumentum $3.2B if they block it.⁷⁸ Its shares fell 36.5% in five sessions, which tells me investors share some of these worries.¹
• A good business at the wrong price, paid for with an owner’s dilution. Cash after stock pay was $0.9B last year on $9.45B of revenue, and stock pay is running at 12% of revenue.²³ At 26 times next year’s consensus the price assumes the growth continues, and every year of it comes with new shares: diluted shares are up 81% in eight years.¹³⁵² If the growth is 17% a year and not 30%, the multiple falls as the earnings rise and the owner earns little.

The first two set my bear. The third is what the section is about. All three are in the company’s own risk factors or its own history.

What the price implies

At $159.80 the market values Marvell at $147B on 921M diluted shares, or $148B with the debt, the earnout and the cash I expect to be on the books at the end of FY2027 C (calculated).²⁹ That is 53 times the past year’s GAAP earnings and 26.2 times Nasdaq’s mean for the year to January 2028.¹³ A year later the consensus is $11.50 from five estimates, which would put the price at 13.9 times.¹³ The stock was at $316 in June. It is 49.5% lower, and the last close below today’s was $156.57 on 29 April 2026.¹

Marvell sits in the middle. It trades below Arista and Lumentum on the nearer year and above NVIDIA and Broadcom. The gap to those two is the market asking Marvell to grow faster for longer, from a base about a tenth of Broadcom’s. The consensus does grow faster, from $6.09 to $11.50, or 89%, in a year.¹³ FY2027 revenue was reported at about $9.5B in September 2025 and the company now says about $12B, so estimates have only been raised, and none has been tested by a cut.¹⁹⁴

The reverse DCF starts from the $2.06B of owner cash in section V. I solve for one growth rate for ten years. The price needs 29.3% a year, to $26.8B in FY2037 C (calculated). At the owner cash margin of my bull case that takes about $108B of revenue, nine times the FY2027 guide, and a good deal more than Broadcom’s $89.1B today E (estimate).¹¹ My base grows owner cash 21.0% a year and my bull 32.5%. The case rates are measured from the same $2.06B, though each case steps up in FY2028 on guided revenue.

The price sits between my base and my bull. An owner today pays for most of management’s plan arriving, and gets my base at 6.5% a year, 5.5 points under the 12% I require and 1.6 points over a Treasury.

What I can’t judge

Three inputs matter most, and I can’t know any of them. I can’t judge how long the cloud companies keep raising spending. Nobody has seen a build this large, and the people who supply it have every reason to be sure of it. I can’t judge how many of the custom chips Marvell will win at the next round. The company says it has won at all four large US buyers, and the Benchmark episode shows how little outsiders can see.¹⁰ And I can’t judge what the Lumentum business is worth. I have not read Lumentum’s filings in depth, so I show a range below and not a figure.

So I show the range. Growth matters more than margin. Moving non-GAAP operating margin 3 points changes value by about $12 a share. Adding 2 points of growth a year for a decade adds $17. It takes about 5.6 extra points of growth a year, every year, for my base to reach the price.

The discount rate

I use 12.0%. The capital asset pricing model starts with the 4.90% ten-year Treasury yield and Damodaran’s 4.20% implied equity risk premium of 1 October.³³³⁴ My beta on ten years of month-end price returns is 1.69, which gives 12.0%; on five years, 2.32, which gives 14.6%; on three years, 2.68, which gives 16.2% C (calculated).¹³⁵ The shorter windows are mostly the AI trade, and I don’t lean on them, but they are all higher than the 10-year figure and I want the reader to see that. I take the ten-year result, 12.0%, which is also the rate on my Micron page and a point and a half above Broadcom’s. One point on the rate moves my base by 11% to 15%. Even at 10.0% and 4% growth after year ten the base is $157, still below the price.

Three cases

I build each case from revenue, a non-GAAP operating margin, stock pay and a deduction for customer warrants. I start from the FY2027 guide of about $12B, where data center is about $9.9B and communications about $2.15B.⁴ From non-GAAP operating margin I subtract stock pay as a cost, so the result is after stock pay. Tax is 12%, rising to 15% by FY2030 because most of the company’s foreign earnings face the 15% global minimum tax. Capital spending starts at 4.5% of revenue and falls to 3.6%, against depreciation of 3.0%.² Working capital takes 18% of each new revenue dollar and is not released when revenue falls, because receivables and supplier prepayments have grown faster than sales. The customer warrant is a revenue deduction of 1.0% of revenue from FY2029 in my base, 0.4% in the bear and 2.5% in the bull, where more tranches vest.²⁰ I add no buybacks and no Lumentum.

My base is not cautious in the near term. It has revenue of $19B in FY2028, which is 95% of the $20B the company is reported to have set, and non-GAAP earnings of about $6.96 a share, above the mean of $6.09 E (estimate).¹³²⁴ For FY2029 I have $9.59 against $11.50. I am above the Street for next year and below it for the year after, and the value is still well under the price. The gap has less to do with the next two years than with years five to ten.

Bear

The cloud companies pause in FY2030 and one of them moves a generation of chips to another design partner. Programs the company has talked about slip, so the Google revenue arrives late and small. Revenue is $14.5B in FY2028, falls to $14.0B in FY2030, and reaches only $19B in FY2037. Non-GAAP operating margin drops to 30% in FY2030 and stock pay stays near 10% of revenue. Marvell still grows, but as a mid-sized chip supplier with the Lumentum shares outstanding.

Base

FY2028 comes in at $19B, about 95% of the $20B goal, and FY2029 at $27B. Growth slows to 17% in FY2031, when revenue reaches $40.5B, about 58% of the low end of management’s $70B to $90B target. One digestion year follows, down 2.5%, and growth then fades from 11% to 5%. Non-GAAP operating margin settles at 37%, stock pay at 6.5% of revenue, and Marvell ends FY2037 as a $59B company.

Bull

The company’s plan arrives close to on time. Revenue is $20.5B in FY2028 and $75B in FY2031, inside the target range, and reaches $138B in FY2037, about 1.5 times Broadcom’s revenue today. Custom chips, optics and scale-up all work. Non-GAAP operating margin stays near 39%, stock pay falls to 5.5%, and customer warrants cost 2.5% of revenue.

Weighted 25%, 50% and 25%, the three cases give about $121 a share E (estimate). The price is 32% above that. About half of each value sits after FY2037, so all of them rest on what the business looks like in ten years. No one can see that clearly, which is one reason not to pay for it in full now.

The skew at $159.80 is −82% to my bear and +51% to my bull C (calculated).

The Lumentum merger

The price already carries this deal, and my values do not, so it needs its own section. Here are the terms.

A fixed exchange ratio means Lumentum holders take whatever Marvell’s stock is worth. At today’s price the offer is worth $159.80 × 4.5 = $719.10 a share, and Lumentum closed at $626.75, 12.8% lower C (calculated).¹²⁹ A gap that wide may reflect real odds that the deal does not close, or does not close on these terms, and the 18 months to a close in 2028. Both stocks fell: Marvell by 36.5% from its 28 October close, in steps of 13.2% on the day talks were reported, 10.0% the next day and 12.7% on the announcement, and Lumentum by 31.4% even though it is the company being paid a premium.¹²⁹ Other chip and AI stocks fell 4% to 19%, so some of it is the market. Most is Marvell’s own, and the timing points to the deal, though I can’t separate it from other news.

The premium works out to about 23.9% on a Lumentum value of $92B at its 28 October close, so roughly $22B handed over by Marvell’s holders, paid in their shares C (calculated).⁷²⁹ The synergies promised are $1B a year of cost savings, mostly in cost of goods. Taxed at 15% and capitalized at 15 times, that is about $13B E (estimate), more than half the premium but not all of it. The first-year accretion promise is mostly arithmetic: Marvell pays in stock valued at 26 times next year’s earnings for a business bought at about 22 times Lumentum’s June 2028 consensus, so earnings per share rise whether or not the price is right C (calculated).¹³²⁹ Before the announcement, Raymond James was reported to have said earnings per share could be flat by 2028 or lower by more than 10% depending on structure.³⁷ I can’t reconcile that with the company’s claim, and the S-4 will say more.

What does it do to a Marvell owner? A simple test, without a model of Lumentum. Marvell owners keep 67% of a bigger company. They are no worse off than before when what Lumentum adds, with synergies, is worth at least 33/67, or 49%, of what Marvell is worth alone. On my values that break-even is $13B in the bear, $48B in the base and $110B in the bull. Lumentum’s market value is $63B at its close.

The table shows the plain lesson. Where Marvell is worth less than the market says, as in my base and bear, buying Lumentum with Marvell’s stock is cheap in my terms. Where I take Marvell to be worth $241, the break-even is $110B, above what the market now pays for Lumentum, so the deal gives away a third of the upside in exchange for cover on the downside. That is the real trade. It is a reasonable trade for a holder who thinks AI optics are the next big market. It takes value from a holder who already thought Marvell would do well on its own. I have not tried to improve on the market’s view of Lumentum, so these figures cover only the share exchange.

Other uses of the same dollar

A dollar put into Marvell today has other places to go. The hurdle isn’t only a Treasury. It’s the best use of the same dollar in the same theme, and four of them are on this site.

On my own numbers, a dollar in Broadcom, NVIDIA or Micron buys more cash per dollar of price than Marvell does, and all three have larger positions in the same build. My base return for Marvell, 6.5%, is below Broadcom’s 7.5%, NVIDIA’s 9.7% and Micron’s 7.1% at their late-October prices, and above Arista’s 3.7%. Micron has since fallen 10%, which would raise its figure. I would place Marvell near the middle of this group on price, with a wider range of outcomes.

The margin of safety

I set the margin-of-safety price from the bear case: the price at which my $29 bear value is still at least three quarters of what was paid. That is $38, 76% below the price and 64% below my base. That is below the two-year low close of $49.38 in April 2025 and the $73.73 of February 2026, and the last close at or below $38 was $36.91 on 25 April 2023, before the AI build lifted the stock. So it is a price from a different cycle.

Two different questions

How good is the business? Good, and improving. Revenue rose 37% in the latest quarter to $2.74B, the largest cloud buyers have chosen it for multi-year programs, and net debt is only about $1B. It is unproven in the thing the price pays for, a decade of winning generation after generation against Broadcom and in-house teams. How good is the price? That is a separate question, and today it looks like a good business at a full price, with a deal on top that nobody can yet price. The price is 50% above my base and 34% below my bull. Marvell avoids being a poor purchase only if it reaches something close to its own plan.

VIIThe case for the price

What most people miss

Three things cut in Marvell’s favour. First, the Google warrant vests only on purchases: 240 tranches, one for every $500M of custom products Google and its affiliates buy through FY2033, and nothing on a design win.²⁰ It costs Marvell nothing if Google buys nothing, and its $206.58 strike is 29% above today’s price.²⁰¹ Second, consensus for the year to January 2029 is $11.50 from five analysts, which puts today’s price at 13.9 times earnings if it comes true.¹³ Third, the stock is back where it traded on 29 April, before management raised its FY2028 target three more times.¹²⁴

The bull case

In plain words: AI buyers want their own chips and few firms can design them, so the two leaders get a long run. Marvell is the second, and the only broad one, with the optics every AI cluster also needs. Mr. Murphy said in the 27 August release that “AI-related bookings remain exceptionally robust,” and the company has raised its FY2028 revenue target from about $13B in December 2025 to about $20B on 6 October, with a first FY2031 target of $70B to $90B.³²⁴ On 2 November he called the Lumentum deal an opportunity that “comes along once in a generation.”⁷ Piper Sandler’s David O’Connor kept a $400 target, reported on 31 October.³⁷ Barclays raised its target to $150 on 9 April, when the stock was near $120.¹⁹¹ Nvidia put in $2.0B at $91.84, and FMR held 14.95% on 31 December.²¹⁰ I saw reports of the notes, not the notes, and I attribute no view to funds I did not check.

Against reality

Today’s numbers support the near term: revenue rose 36.5% in the July quarter, the October quarter is guided about 15% higher, and data center about 75% on a year.³⁴ If FY2028 ships at $19B to $20B with non-GAAP operating margin of 38% to 40%, my FY2028 figures hold.⁴

The price needs more. If Marvell reaches $75B by FY2031 with a 39% margin and stock pay down to 5.5%, $241 follows. Today’s price needs about $108B of revenue by FY2037 at the bull’s cash margin. Marvell has $9.45B of trailing revenue, $0.9B of cash after stock pay, and custom revenue of about $1.5B in FY2026.¹⁹

Multiply it out. Cloud spending keeps rising to FY2031 without a bust, 60%. Marvell wins enough custom sockets for $10B or more in FY2029, 50%. Optics and scale-up reach a line worth more than $10B, 50%. Margin holds at 38% to 40% while custom grows, 60%. Lumentum closes and integrates without hurting the DSP business, 70%. These are my guesses and not independent. Multiplied they give 6.3% E (estimate); allowing for overlap, perhaps 10% to 15%. I still give the bull 25%, as on my other pages, which is generous.

What a 10x would take. Ten times $159.80 is $1,598, about $1.47T on today’s shares and $2.2T after the merger. At 25 times earnings that is $59B of net income, and at the 26% net margin of my bull about $223B of revenue: 19 times the FY2027 guide, 2.5 times Broadcom’s trailing $89.1B, and most of the way to NVIDIA’s $303.0B, from 3% of it.¹¹²⁶ By FY2037 that means growth of about 34% a year for ten years C (calculated). Marvell rose 12.8 times in ten years, from about $6.4B of market value, with the share count up 81%.¹⁵ A tenfold move is close to impossible, and I would not build a case on it.

Where I come down

The repeated target raises and the Google warrant tied to actual purchases moved me: they are why FY2028 in my base sits above the Street. The thin cash after stock pay, the share count and an all-stock deal near the top of the cycle did not, because they sit at the centre of what an owner receives. Custom revenue of $10B in FY2029, non-GAAP operating margin at 38% and flat shares before the merger would raise my values most.

VIIIRisks, and what would change my view

The three ways this goes badly wrong are at the start of Valuation. These are the rest.

• Supply and one fab region. Marvell depends on a limited number of suppliers, including TSMC for advanced wafers, and most of its qualified foundries are in the same part of Taiwan.²⁶ It prepaid $487M for capacity and plans about $1B this year, which only pays off if the chips ship.²⁴
• China and export rules. Shipments to China were 42% of Q2 revenue, mostly for non-Chinese customers with factories there, and the 10-Q lists tariffs and export licences as risks.² China approval is also needed for the merger.
• Customers who build their own. The 10-K says some customers have developed certain chips internally and that this may spread, and it lists the risk of losing or gaining design wins at key customers.⁶
• Accounting that moves with the share price. The Celestial earnout is marked to market and cost $434M in the first half. A rising stock raises the bill, though a falling one lowers it.²
• Governance. One person is chairman and chief executive, and 37% of votes cast for an independent chair. Pay approval was 56% in June, and the new CFO joined from the board in June.¹²³⁰ The merger will add three directors from Lumentum to a board of seven.⁷
• The Nvidia link. Nvidia holds preferred stock that converts to about 21.8M shares and a partnership that connects Marvell’s custom chips to its ecosystem. It is also the largest company in the market Marvell sells into, so the interests overlap without being the same.²
• Legal and technical. The 10-K says Marvell is sometimes sued over patents and may not prevail, and that AI could change its business model.⁶ I cannot size either.