What they said
The numbers, as reported from the room, since Marvell posted the webcast and not the deck. Fiscal 2028 revenue of approximately $20 billion, which Murphy framed as “about 67% year-over-year growth,” of which approximately $18 billion is data center. Custom revenue of more than $12 billion in fiscal 2029, up from the more than $10 billion target the company carried into the day. A fiscal 2031 framework of $70 billion to $90 billion in revenue and non-GAAP EPS above $30, the EPS figure attributed to Dan Durn. At the $80 billion midpoint: roughly $37.5 billion of interconnect, roughly $30 billion of custom compute, roughly $10 billion of switching and storage, and roughly $2.4 billion of communications and other unverified. A $400 billion addressable market by calendar 2030. Fiscal 2026, for scale, was $8.2 billion.
The long-term model. One account of the presentation carries fiscal 2031 targets of 56 to 59 percent non-GAAP gross margin and 44 to 46 percent non-GAAP operating margin unverified. I have not found a second account, Marvell has not posted the deck, and nothing has been furnished to the SEC, so those two ranges carry an unverified tag until they can be read from the source. The EPS figure above $30 appears in three accounts and I treat it as reported.
Murphy attributed the fiscal 2028 raise primarily to connectivity, naming 1.6T optical digital signal processors and scale-up optics, with data center continuing to grow faster than the company. Durn said the AI infrastructure build-out “is continuing at a massive scale and speed.”
What was announced, and why
Marvell did not come to New York with a single product launch. It came with an argument about where a hyperscaler's AI dollar goes, and three businesses it says sit in the path of that dollar.
Custom silicon, and why hyperscalers want it. Murphy's case was that the largest cloud operators increasingly prefer application-specific chips tuned to their own workloads and owned outright over renting merchant compute built for everyone. Marvell's pitch is not to out-design the merchant GPU but to supply the pieces that turn a customer's own architecture into working silicon: the ASIC design work, the high-speed SerDes that move data between dies and packages, the HBM memory interfaces, advanced packaging and chiplet integration, and access to leading-edge process nodes. The company says it works with all four of the largest hyperscalers and, per its own investor materials, has three XPU design wins and nine XPU-attach programs across them unverified. Custom was roughly $1.5 billion of revenue in fiscal 2026, close to 18 percent of sales unverified; the fiscal 2029 target is now more than eight times that derived. The Google program behind the warrant covers silicon attached to the TPU ecosystem: inference accelerators, NICs, storage controllers, memory interface controllers and near-memory compute unverified.
Interconnect, which the company now puts ahead of custom. This is the business that leads the fiscal 2031 case, at $37.5 billion against $30 billion for custom, and it is the business Murphy credited for the fiscal 2028 raise. The technology it was built on was shown two weeks earlier, on at ECOC: 2nm 400G-per-lane PAM4 optics, 2nm 800G ZR and ZR+ pluggables, 1.6T ZR on the way to 3.2T, and a 102.4-terabit co-packaged optics platform. The stated reason is power: as AI pods scale across racks, copper runs out of reach and the electrical-to-optical conversion has to move closer to the switch and the accelerator, which is what near-packaged and co-packaged optics are. Marvell had sized scale-up optics at about $300 million of fiscal 2028 revenue a quarter ago, roughly $150 million of it from the Celestial AI photonic fabric it acquired, and said in August the opportunity had grown materially unverified. The scale-up fabrics it is targeting include UALink, ESUN and NVIDIA's NVLink Fusion unverified.
Memory expansion. Management had promised a comprehensive update on CXL and memory-expansion silicon, which it says is already deployed at multiple hyperscalers in high volume unverified. I could not find an account of what was said about it on the day, and the fiscal 2031 split folds it into XPU-attach.
Why the model looks the way it does. Put the three together and the margin ranges make sense as a composition. Custom silicon ships at a lower gross margin than Marvell's corporate average, which is what the Q3 guide already showed; optics and switching ship nearer to it. A fiscal 2031 in which interconnect is the larger business is a fiscal 2031 in which the blended gross margin stops falling, which is why the range can top out at today's 58.9 percent rather than above it. The operating-margin range is a statement that research and development will not scale with revenue: roughly ten times the revenue of fiscal 2026 on a research base that grows far more slowly, and operating margin goes from the mid-thirties to the mid-forties. That is the whole model in one sentence, and it is the sentence the August piece was waiting for.
The technology, in plain English
Custom silicon versus the merchant GPU. Renting a tour bus against having a car built for your commute. The bus works for everyone and costs what the bus company charges; the car does one route and nothing else, and once it is built the per-mile cost is yours. Hyperscalers run a handful of enormous, repetitive workloads, so a chip tuned to one of them can be cheaper to run than a general-purpose accelerator. Marvell is the coachbuilder rather than the engine designer: the customer brings the architecture, and Marvell supplies the chassis work (the ASIC design), the wiring harness (SerDes, the high-speed links between dies), the fuel lines to memory (the HBM interfaces), the body shop (advanced packaging and chiplets) and a slot at the factory (access to leading-edge process nodes).
Why optics, and why co-packaged. Inside an AI data center the chips talk to each other constantly, and the wires that carry that traffic are the bottleneck. Copper is a garden hose: fine over a short run, but it loses pressure with distance and the pump that pushes signal through it burns power. Fiber is a pipe that carries light and does not care about distance. Today the converter that turns electrical signal into light sits at the edge of the rack in a pluggable module. Co-packaged optics put the converter on the chip's own package, which is the difference between a kettle in the kitchen and a kettle on your desk: less walking, less wasted heat. That is what "scale-up" optics means here, connecting accelerators to each other inside a pod, and it is why a 2nm optical chip is news: the smaller the converter, the less power per bit.
Gross margin against operating margin. Gross margin is what is left of a sale after paying for the ingredients; operating margin is what is left after also paying the chefs. A custom chip is a lower-margin dish because the customer negotiated the price knowing the recipe was theirs. The recipe, which is the research and development, is written once and then sold for years, so the kitchen's payroll grows far slower than its sales. That is the entire reason the company can model a gross margin that goes nowhere and an operating margin that goes up eight points at once.
The warrant. A loyalty card that pays Google in Marvell shares, one tranche for every $500 million of custom orders. The more the stock is worth, the more each tranche is worth to Google and the more it costs Marvell's other shareholders. It is a rebate that gets more generous as the thesis works.
Read against the view
The piece did not dispute the quarter and did not dispute the raise. Its claim was that the tape was repricing what a Marvell revenue dollar is worth: the third-quarter guide implied a 51.9 percent incremental gross margin against the 58.9 percent just reported, the fiscal 2027 raise dropped through at roughly 31.9 percent against a 38 to 40 percent model, and the Google warrant vests in 240 tranches of $500 million of custom revenue, so the better the stock does, the more each tranche costs. The falsifier was dated: wrong if Q3 FY27 non-GAAP gross margin prints at or above 58.5 percent, right if at or below 57.5 percent.
Today raised the numerator again and, for the first time, priced the ratio. Fiscal 2028 has now gone from approximately $16.5 billion in May to approximately $18 billion in August to approximately $20 billion in October, $3.5 billion in five months derived, and the share of it that is data center, approximately $18 billion of $20 billion, is 90 percent derived against 79 percent of revenue in the quarter just reported. The mix the August piece worried about is the mix the company is guiding toward, faster. And the long-term gross-margin range, if the one account is right, has a midpoint of 57.5 percent, 140 basis points below the quarter just reported derived. The company's own five-year model does not recover the margin it just printed. That is the August read, stated by Marvell.
The tell is the operating-margin range. A 44 to 46 percent target against 36.6 percent reported is an 840-basis-point step at the midpoint derived, delivered entirely below the gross line. Marvell is telling you the incremental dollar earns less gross profit and more operating profit, because the research that produces a custom program is paid once and the program ships for years. If that is right, the August piece's drop-through arithmetic was correct about the quarter and wrong about the destination: the 31.9 percent that the fiscal 2027 raise fell through at is the cost of starting programs, not the economics of running them.
The warrant is still the warrant, and today made it bigger. More than $12 billion of custom revenue in fiscal 2029 is 24 tranches of $500 million derived if every custom dollar were Google's, which it is not, so read 24 as the ceiling on how fast the ladder can vest in a year. At 240,041.83 shares a tranche that ceiling is about 5.76 million shares derived. The number the warrant costs per dollar of revenue still rises with the share price, and the share price rose today.
What changed in the view, and what did not
Unchanged: the December print is the test and the thresholds are 58.5 percent and 57.5 percent. A long-term range of 56 to 59 percent does not tell you where the third quarter lands inside it.
Changed: the question is no longer whether Marvell's gross margin comes down with custom. The company has modeled that it does, or at best holds. The question is whether operating leverage pays for it, and the company has put a number on that too. The honest reading of the August piece is that it was right about the gross line and did not give enough weight to the opex line, and the day corrected the weighting. The interconnect-led fiscal 2031 split is the mechanism: the custom ramp dilutes, the optics ramp does not, and the research base grows slower than both.
Left unresolved, stated plainly: the margin ranges rest on one account, Marvell has not posted the deck, and nothing has been furnished to the SEC. If the presentation is filed, this page will be reread against it. Also unresolved: what was said about memory expansion, and whether the fiscal 2031 split is a forecast or, as the company called it, a scenario.
- Wrong if Q3 FY27 non-GAAP gross margin, due in , prints at or above 58.5 percent. That would mean the custom ramp is not diluting at the rate the Q3 guide implied, and the August read was wrong about the incremental dollar.
- Right if it prints at or below 57.5 percent and the fiscal 2028 guide holds at approximately $20 billion, which is the raise outrunning the gross margin for a third consecutive quarter, with the operating-margin range now the thing the company has to prove.
The tape, for the record: $271.25 on , $287.19 at today's close after a session that ran from $267.26 to $301.27, against the $216.595 the piece was written at. The stock has repriced the revenue dollar upward since August. The print that says what the dollar earns is still ahead.