Marvell’s AI Rally Looks Priced for Perfection — And That’s Dangerous

Marvell Technology has become Wall Street’s favorite AI infrastructure play not named Nvidia. The stock has more than doubled since early 2023, fueled by a narrative that custom silicon for hyperscalers like Google and Amazon will drive decades of growth. But after the latest earnings beat and raised guidance, the real question isn’t whether Marvell benefits from AI — it’s whether anyone buying at 45 times forward earnings has a margin of safety, or just momentum.

The AI Tailwind Is Real, But Priced In

Let’s give credit where it’s due. Marvell’s data center revenue hit $816 million last quarter, up 87% year over year. The company’s electro-optics and custom compute chips are genuinely critical for the AI clusters that Microsoft, Meta, and others are frantically building. CEO Matt Murphy said on the call that AI revenue will hit $2.5 billion this fiscal year — double the prior target. That’s not hype. That’s purchase orders.

But the market didn’t just price in $2.5 billion. It priced in $5 billion, then $10 billion, with perfect execution and zero competition. The stock trades at a premium to Broadcom, its closest custom-silicon peer, despite Broadcom’s larger scale, broader portfolio, and proven VMware integration. Marvell’s enterprise networking and carrier infrastructure segments — still 40% of revenue — are flat to declining. You’re paying AI multiples for a mixed business.

Custom Silicon vs. Merchant Market Dynamics

The bull case rests on Marvell winning successive generations of custom ASIC programs at Google, Amazon, and maybe Microsoft. That’s a great business — high margins, sticky customers, multi-year visibility. But it’s also a concentrated one. Three customers. Three design wins. If Google decides its next TPU generation stays in-house, or if Amazon’s Trainium roadmap shifts, Marvell loses a $1 billion-plus revenue stream overnight.

Contrast that with the merchant market, where Nvidia sells H100s to anyone with a purchase order. Marvell’s custom model locks in volume but locks out optionality. And the switching costs cut both ways: hyperscalers design their own chips precisely to avoid vendor lock-in. They’ll keep Marvell honest on pricing. They’ll also keep developing internal alternatives. Google’s TPU team didn’t hire thousands of engineers to outsource forever.

The Valuation Math Doesn’t Add Up

At $75 billion market cap, Marvell trades at 12 times trailing revenue and 45 times forward EPS. Assume the $2.5 billion AI target hits. Assume it grows 30% annually for three years — aggressive, but let’s play along. That’s $5.5 billion in AI revenue by FY2027. Add a generous 60% gross margin and 30% operating margin on that increment. You get maybe $1.6 billion in incremental operating income. On a 25% tax rate, that’s $1.2 billion net. Capitalize at 25x — rich for a cyclical semi — and you get $30 billion of value just from AI growth.

The rest of the business? Enterprise networking, carrier, automotive, consumer — call it $3.5 billion revenue, low-single-digit growth, 20% operating margins. Worth maybe $15 billion. Total: $45 billion. Today’s price: $75 billion. You need AI revenue to hit $8-10 billion *and* sustain 30%+ growth *and* the rest of the business to re-accelerate just to justify today’s multiple. That’s not investing. That’s hoping.

What Could Go Right (And Wrong)

Fair’s fair: the bull case has teeth. Marvell could win the next Microsoft Maia generation. It could capture Merckle interconnect share as clusters scale beyond 100k GPUs. The 1.6T and 3.2T DSP roadmap could become the de facto standard for optical connectivity. And if the enterprise networking cycle finally turns — Cisco just guided up, Arista’s beating — the “other 40%” becomes a tailwind instead of an anchor.

But the risks are asymmetric. Export controls could cut off China revenue (15% of sales). A hyperscaler pause — remember Meta’s 2023 “year of efficiency”? — would crush custom ASIC demand. And if Nvidia’s NVLink ecosystem makes merchant accelerators more attractive than custom silicon, the whole thesis inverts. Marvell’s CFO recently sold $2.3 million in shares. Insiders don’t usually sell at the bottom.

None of this makes Marvell a bad company. It’s a very good company in a very hot sector. But the stock price assumes perfection in an industry famous for cycles, competition, and sudden technology shifts. If you own it, consider trimming. If you’re thinking of buying, wait for the inevitable pullback when the next quarter’s guidance is “only” 20% growth instead of 30%. The AI revolution is real. Marvell’s valuation? That’s the bubble talking.

Source: original report