Marvell's Record Quarter Proves Custom Silicon Is Real, and Shows Exactly What It Costs

Marvell Technology delivered exactly the quarter the custom-silicon bulls had been waiting for. Investors sold it anyway — and the reason they sold is the most interesting fact in the whole report.

The chipmaker's fiscal second quarter of 2027, which ended August 1, 2026 and was reported after the close on August 27, produced record revenue of $2.739 billion, up 37% year over year and 13% sequentially, landing $39 million above the midpoint of the guidance management issued in May. Data center revenue climbed to $2.172 billion, up 46% year over year and 18% sequentially, and now accounts for 79% of everything Marvell sells, up from 74% a year earlier. Communications and other revenue was $567.8 million, up 10% year over year.

Profitability scaled with it. Non-GAAP operating income crossed the billion-dollar line for the first time at $1.003 billion, a 36.6% operating margin. Non-GAAP net income was $865.9 million, or $0.94 per diluted share. GAAP net income rose 58% year over year to $308.0 million, or $0.33 per share, on $605.5 million of operating cash flow.

“AI-related bookings remain exceptionally robust, and we expect our revenue growth to accelerate further through the remainder of fiscal 2027,” said Chairman and CEO Matt Murphy in the earnings release. On the call, Murphy raised the fiscal 2027 revenue outlook to roughly $12 billion from about $11.5 billion, and lifted fiscal 2028 to approximately $18 billion from the $16.5 billion he had given one quarter earlier — implying roughly 50% company-level growth next year, with data center revenue growing more than 60%.

Shares closed the next session down 10.28% at $216.62.

The one number that went the wrong way

Almost every figure in the release moved up. One did not. Non-GAAP gross margin came in at 58.9%, flat sequentially but down from 59.4% a year earlier, and management guided the third quarter to a range of 57.5% to 58.5%. On roughly $3.15 billion of quarterly revenue — the Q3 midpoint, which would represent about 15% sequential growth — each point of gross margin is worth roughly $30 million in quarterly gross profit.

Management did not hide the cause. “Revenue levels and product mix remain key determinants of gross margin in any given quarter,” CFO Dan Durn said on the call, pointing directly at the accelerating custom chip business, and adding that the company expects to hold that range through the fiscal fourth quarter as well.

That is the trade in one sentence. Murphy has now publicly pinned significant acceleration in the custom business to the second half of fiscal 2027, and the same mix shift that produces the acceleration is what compresses the margin. Marvell also said it is engaged in more than 50 new custom AI design opportunities across more than 10 customers, and is on pace for roughly $1 billion of supplier capacity prepayments this year.

Sitting behind the guidance raise is the warrant Marvell issued to Google on August 18, 2026, covering 58,970,907 shares at an exercise price of $206.58 — about $12.2 billion if fully exercised, and exercisable through August 2033. The warrant vests in 240 tranches, one for every $500 million of custom product revenue Marvell recognizes between Q3 FY2027 and the end of fiscal 2033. That is where the widely quoted $120 billion figure comes from. It is a ceiling, not a contract. The covered products span Google's TPU-adjacent stack, from AI inference accelerators to storage controllers, NICs and memory-interface controllers, with management flagging upside to prior custom expectations from fiscal 2029 onward.

Custom silicon versus the merchant GPU

Put Marvell's quarter next to Nvidia's, reported one day earlier, and the shape of the argument becomes clear. Nvidia's fiscal Q2 2027 data center revenue was $89.0 billion, up 117% year over year. Marvell's data center revenue was $2.172 billion. The merchant GPU business is roughly forty times larger and, this quarter, growing faster in percentage terms. Whatever the hyperscalers are doing with in-house accelerators, they have not yet dented the trajectory of the incumbent.

But the two models are not competing on the same axis. Nvidia sells a complete, differentiated system and captures the architecture rent — gross margins in the seventies. Marvell sells engineering: SerDes, packaging, IP integration and the ability to tape out a hyperscaler's own design. The customer keeps the architecture and, critically, the margin. Marvell's 58.9% non-GAAP gross margin drifting toward 58% is not a sign of weakness; it is the structural price of the ASIC model. Every dollar of custom revenue Marvell wins is a dollar of accelerator spend that flows through a merchant supplier at merchant-supplier economics rather than through Nvidia at platform economics.

Which is why the sell-off reads as repricing, not doubt. Investors were not questioning demand — bookings and the FY2028 raise argue the opposite. They were marking to market what winning these sockets actually costs, on a stock that had more than tripled from its 52-week low of $61.44 and still trades near 35 times forward earnings. The Nvidia relationship complicates the binary further: Nvidia invested $2 billion in Marvell in March and Marvell is developing custom XPUs and scale-up networking compatible with NVLink Fusion. The custom-versus-merchant framing is increasingly an “and,” not an “or.”

What to watch

The October 6 Investor Day in New York is the next real catalyst. Management explicitly deferred two things to it: a longer-term custom silicon revenue trajectory with actual ranges, and a reset of the long-term operating model. Marvell expects non-GAAP operating margin to enter its 38% to 40% target band in Q4 FY2027 and reach the upper end during FY2028.

What matters most is not the headline beat. It is whether the first Google tranches actually vest on schedule starting in Q3, and whether Marvell can hold gross margin near 58% while custom scales. If custom ramps and margin holds, the ASIC thesis works. If custom ramps and margin keeps sliding, hyperscalers really are leaving Nvidia — and taking the profit pool with them.