Marvell Technology, Inc. MRVL

Key data

The business

Marvell sells the silicon that hyperscale data centers run on, and two things matter. First, custom compute: it co-designs bespoke AI accelerators, the XPUs, for a handful of cloud giants who want their own chip instead of Nvidia's, then hands the manufacturing to TSMC. Second, interconnect: the high-speed optical DSPs, retimers and ethernet silicon that move data between thousands of those chips without choking. Data center is the large majority of revenue now, and the old lines, carrier, enterprise networking, storage, are along for the ride. The moat is the design socket: once Marvell wins a hyperscaler's custom chip it is embedded for that chip's multi-year life, and the SerDes know-how behind fast interconnect sits with only two or three firms. It is a clear number two to Broadcom in custom silicon. In August it folded in Celestial AI, a photonics startup bought to push interconnect onto light inside the rack, the deal now showing as $2.8 billion of fresh goodwill, so the connectivity move flagged a week ago has closed and is in the numbers. China is 42% of where product ships, up from 29%, a licensing headache waiting to happen.

The numbers

Revenue is the clean line; GAAP profit is noise until you strip the one-offs.

Sales rose every quarter, up 37% year over year to a record. The $1.90B in Q3 FY26 is a one-time gain from selling the automotive ethernet unit to Infineon; Q1 FY27's collapse to $35 million is amortization and deal charges eating operating income. The gauge that holds still is non-GAAP earnings, which the company prints steadily: $0.76, $0.80, $0.80, $0.94 across the last four quarters, each a hair above what analysts modeled.

Four straight years of GAAP losses through FY2025, then FY2026 revenue jumped 42% and the company crossed into profit as AI orders landed. The half already at $5.16B annualizes past $10B, another 26% up. That is the whole story: a slow-growth networking supplier that became an AI compounder in eighteen months. Consensus puts non-GAAP earnings near $10 by the year ending January 2029, roughly triple FY2026. Pay $223 today and you sit at 22 times that number if it arrives, 74 times what was actually earned last year if you doubt it, more than any full year in its measurable history. The market prices the custom-silicon ramp as near-certain; what settles it is whether data center revenue holds a 35%-plus growth rate while Broadcom fights for the same sockets.

Where the product ships tells the export story in one glance.

Management

Insiders sold and none bought: twelve sales worth $19.4M over the year, CEO Matt Murphy taking $2.2M in June and CFO Chris Koopmans two lots of $2.8M and $2.1M, plan status not disclosed on any of them. Small against a $196B company, but one-way. Capital allocation reads better. They bought back $2.04B of stock in FY2026 when shares were far cheaper than today, then eased off in the first half of fiscal 2027 as the price ran, which is the right instinct. Debt rose to $4.96B to help fund Celestial, against $3.93B cash, so net leverage stays under a third of a year's cash earnings. R&D runs 26% of revenue, the real rent on staying in the socket.

How it fails or surprises you

Broadcom takes the next socket. Custom-silicon revenue leans on a few hyperscaler programs, and the biggest buyers already dual-source. Lose one XPU at its next node, a decision made 18 months before it shows in revenue, and the growth line snaps toward flat. Watch for any hyperscaler naming a second vendor on a chip Marvell holds today.

Light in the rack (right tail). If Celestial's photonics and Marvell's optical DSPs make co-packaged optics the standard way to wire an AI cluster, the interconnect business expands into a market barely priced now. The tell is 1.6T optical design wins and a hyperscaler committing to optics inside the rack, not just between racks, within a year.

The 74x is never earned. GAAP profit was $343 million for the half while the stock is $196B. Non-GAAP closes that gap only if you accept adding back the better part of a billion a year in stock pay plus steady intangible amortization. If that gap never narrows, real earnings never justify the multiple, and the 42% crash to $63 in the last year is what that looks like.

Closing thoughts

The evidence shows a business genuinely inflecting, priced as if the inflection is guaranteed. The market already knows the story: 32 analysts model the ramp, the multiple discounts triple earnings by FY2029, and whoever sells to you is underwriting the same custom-silicon wins you are. The edge is not the story, it is the slope, whether data center compounds a few points above or below a high bar. The left tail is fatter than the right, because one lost socket or an export clamp on the 42% shipping to China removes years of growth at once, while the optical upside builds slowly. Break the downside linchpin and you lose more than half; land the right tail and you make a multiple over years, not next quarter.

The bet is still that Amazon, Microsoft and Google keep hiring Marvell to design their AI chips and buy its parts to wire the racks, and keep choosing it over doing the work themselves or calling Broadcom. What breaks it is a socket lost or a China license denied, and the one pair of numbers that warns you first is the data center revenue growth rate set against that 42% China exposure. You are paying a near-record multiple for a company whose GAAP earnings barely exist at scale, wagering the franchise is durable enough to grow into it. If data center growth drops below 20% while China stays two-fifths of shipments, the read is wrong.