Marvell just printed the best quarter in its history — record revenue, raised guidance, a roadmap Wall Street had been begging for — and the stock got punished anyway, tumbling roughly 8% in a single session. That says everything about how absurdly high expectations had run, and almost nothing about the health of this business.
Here's what Marvell actually does, because it matters more than the ticker. When a hyperscaler like Google or Amazon wants an AI accelerator built to its exact spec — not an off-the-shelf GPU, but a chip designed around its own models and its own data centers — Marvell is one of only two companies on the planet that can deliver it, alongside Broadcom. That's the Google TPU partnership, the crown jewel: a custom-silicon relationship that deepened in August with a $12.2B warrant deal that effectively locks Marvell into Google's AI roadmap for years to come.
Then there's the other half of the story: every one of those chips needs to talk to thousands of others at speeds that would melt ordinary networking gear. Marvell's optical transceivers — the 800G and 1.6T modules that shuttle data between accelerators — plus its photonic fabric, scale-out switches and storage controllers are the connective tissue of every serious AI cluster on Earth. The hyperscalers aren't just Marvell's customers; they're basically its entire addressable market.
Which brings us to the quarter, and why the numbers were so good: revenue hit $2.739B, up 37% year over year and roughly $39M ahead of the company's own guidance midpoint — a record. Data center revenue, the AI heart of the business, came in at $2.17B, up 46%, and now accounts for about 79% of everything Marvell sells. Non-GAAP earnings of $0.94 a share topped estimates, gross margin held at 58.9% on a non-GAAP basis, and operating cash flow cleared $600M.
And management didn't stop there. Q3 guidance calls for $3.15B in revenue at the midpoint with non-GAAP EPS of $1.10, the full-year outlook jumps to roughly $12B — close to 50% growth — and the company raised its FY28 view as well, with AI-related bookings robust and the backlog building. On the numbers alone, this was a beat-and-raise in the purest sense.
Here's why the AI sector should care: the buildout everyone's watching has two bottlenecks — compute and connectivity. Nvidia owns the first with its GPUs; Marvell and Broadcom own the second with custom silicon, and Marvell owns the optical fabric that moves data between every node. It's also inside Nvidia's NVLink Fusion alliance, putting its optics into the next generation of Nvidia's own systems, and its controllers sit in the memory infrastructure every AI rack depends on — custom chips, optical fabric, memory: the full spine of the AI data center, and Marvell touches all of it.
So why did the stock fall? Because the bar wasn't high — it was absurd, supercharged in mid-August when news of the $12.2B Google warrant deal broke and shares gapped up hard.
By the time earnings landed, the Street wanted more than a beat; it wanted detail on exactly when Google's custom-silicon revenue starts showing up in the numbers. Marvell delivered guidance, not granularity — Barron's put it bluntly: the beat 'was not big enough' for a stock priced for perfection, and Reuters flagged the Google questions still lingering.
The de-rating actually started before the print: the stock gapped up on the day, then faded.
None of that changes the business underneath. Shares now sit about 27% below their June high — even after roughly quadrupling off last September's low, a move that was always going to invite a breather. The Street still rates MRVL a Strong Buy with a mean target north of where shares trade today.
The next catalyst is October 6, when Marvell hosts its Investor Day — the natural venue for the Google and custom-silicon detail everyone wants. That's the date to circle.
This selloff is about price and expectations, not the company.
Disclosure: The Signal holds no position in MRVL. Positions may change. This is not financial advice.
