Marvell's Record Q2 Was Real. The Market Already Paid for It.

Marvell turned in another record quarter and raised its revenue outlook again on Thursday afternoon, and the stock still slipped in extended trading. That disconnect — a clean beat met with a shrug — is the whole story of where this stock trades in late 2026. Up 184% year to date on the custom-AI-silicon narrative, Marvell had already priced in the record before the press release hit the wire. The real content of the print isn't the record. It's what the record costs.

The engine is working. Reported net revenue was $2.739 billion for the fiscal quarter ended August 1, up 37% year over year and $39 million above the midpoint of Marvell's own guidance. The engine was Data Center: $2.17 billion, up 46% from a year ago and 18% sequentially, now 79% of the company's revenue, versus 74% a year earlier. Management guided the next quarter to $3.15 billion in revenue with non-GAAP EPS of $1.10, and raised its fiscal 2027 target to about $12 billion while adding $1.5 billion more to fiscal 2028 — on the $16.5 billion number raised in May, that lands near $18 billion, roughly 50% growth stacked on top of this year's pace. CEO Matt Murphy did what CEOs do on these calls: described AI bookings as "exceptionally robust" and promised growth would "accelerate further" through the rest of fiscal 2027.

None of that is a surprise to the market, which is exactly the problem.

Growth at the expense of margin is a trade, not a triumph. Read down the income statement and the record acquires a price tag. Non-GAAP gross margin of 58.9% was down from 59.4% a year ago, and Marvell guided the next quarter to 57.5%–58.5% — compression, not expansion. The reason is the product mix: custom silicon carries lower gross margin than Marvell's merchant optical and networking parts, so every marginal dollar of custom revenue drags the gross line down. That's the structure doing what it's supposed to; the offset is operating leverage, with non-GAAP operating margin expanding to 36.6%. But the tradeoff is visible and intentional: higher revenue growth, lower unit margin.

The earnings-quality gap is where it gets harder to ignore. Non-GAAP EPS was $0.94; GAAP EPS was $0.33. The roughly $558 million difference between non-GAAP net income of $865.9 million and GAAP net income of $308 million is mostly stock-based compensation of $326.2 million — up 112% from a year ago, about three times the pace of revenue growth, and equal to nearly 12% of sales — plus $214.9 million of amortization from the Inphi and Cavium acquisitions. The diluted share count still climbed 5.8% year over year, to 921 million, even with $200 million of buybacks in the quarter. The headline non-GAAP growth rate is flattered by the exclusion of a compensation line that is growing three times faster than the business.

The custom-silicon wave is real, but customers who own the design can, in time, take it home. The reason hyperscalers route around Nvidia — which still holds roughly 70% of the AI chip market, a share that's eroding — is unit economics, not PR. At tens of millions of chips, a custom part tuned to the customer's own inference-to-training split and power budget wins on total cost per useful FLOP even after absorbing hundreds of millions of dollars in design cost. That's why Google, Amazon, and Microsoft all run custom programs, and why Broadcom — whose Google TPU supply agreement now runs through 2031 — reported $8.4 billion of AI semiconductor revenue in a single quarter, up 106%.

Marvell's share of that work is real and growing, but it is structurally concentrated. Data center is 79% of revenue, and the growth is built on custom silicon designed with a small number of hyperscale customers. In a custom program, the hyperscaler owns the architecture; Marvell sells implementation, SerDes, packaging, and the scale-out interconnect. That is precisely why the margin is lower — and why the revenue has a clock on it. Every program is a race between the customer's in-house silicon team maturing and Marvell staying indispensable. The lesson of Amazon's Graviton applies inverted: hyperscalers in-source precisely so that, over enough generations, they can take the co-designer's cut for themselves. Marvell's answer is the part of the portfolio the customer can't easily reclaim — the electro-optics, the 1.6T interconnects, the switches, and memory controllers that stitch these systems together. It announced a 102.4 Tbps AI switch in June and describes the Data Center portfolio as connectivity plus custom compute; the connectivity component is the durable claim.

That is the lens for the August Google news, which the market treated as a rocket. Marvell did not sell Google $12 billion of revenue. It issued Google a warrant to buy up to 58.97 million shares — about 6% of the company — at $206.58, worth roughly $12.2 billion only if fully exercised and only as Google meets purchase milestones tied to an expanded custom partnership that "attach[es] to the TPU ecosystem." The filing is a long-dated supply commitment; the $120 billion revenue figure circulating in coverage is aggregate potential revenue out to 2033, a projection, not a contract. The warrant is a real signal of demand stickiness — Google is locking in Marvell capacity the same way it locked in Broadcom — but it is not money in the door, and the milestones are the part worth watching.

Which returns the argument to valuation. On the day's data, Marvell closed the regular session at roughly $241, down 1.5%, at a market capitalization near $212 billion — about 24 times trailing sales and in the neighborhood of 80 times trailing EBITDA. The stock spent the past year re-rating from a $61 low to a $330 spring high on repeated multi-billion-dollar guidance raises, so by Thursday evening the raise was the base case, and a beat of about 1% on both lines didn't clear the whisper bar. Barron's headline after the close said it plainly: big expectations outweighed solid earnings.

The quarter was not the story; the price of the forward path is. Marvell can grow — the numbers are not in dispute. The question an investor has to answer is whether roughly $18 billion of fiscal-2028 revenue, produced at falling gross margins and with stock compensation consuming a rising share of the economics and concentrated in a handful of customers who own their designs, is worth the current multiple. The October 6 investor day is where management gets to show the margin structure that's supposed to back the price. Until then, the record is real — and so is everything the market already paid for it.