Marvell Just Bought Its Own Ticket to the Front Row

# Marvell Just Bought Its Own Ticket to the Front Row

Marvell's Google deal has a tag on it, and the tag is worth $12 billion.

On August 19, 2026, Marvell filed an 8-K announcing an expanded custom-silicon agreement with Google. In exchange for building inference accelerators, storage and network interface controllers, and "near-memory compute" products bolted onto Google's TPU ecosystem, Marvell handed Google a warrant to buy up to 58,970,907 shares — roughly 7% of the company — at a $206.58 exercise price. Fully exercised, that warrant is worth about $12.2 billion, tied to revenue milestones that could represent up to $120 billion in cumulative procurement over roughly six years.

That's not a customer relationship. That's a hostage situation with a spreadsheet.

Here's the read most people are missing: Google is paying Marvell in its own stock to keep buying from Marvell. When the warrant gets exercised, Marvell's shares get diluted. When revenue milestones get hit, Marvell's shares get diluted more. The hyperscaler handed itself a loyalty program and handed Marvell a permanent overhang. The stock popped 12% to 14% on the disclosure because the headline was "$120 billion in procurement." Nobody at that rally was doing the math on who's actually paying for it.

Now, the business is real. Marvell reported Q2 FY2027 on August 27 — record revenue of $2.739 billion, up about 37% year-over-year, with Data Center at $2.17 billion driving 79% of the mix and up roughly 46%. Non-GAAP EPS of $0.94. They raised full-year guidance to ~$12 billion for fiscal 2027 and ~$18 billion for fiscal 2028, and pointed to custom silicon revenue more than doubling next fiscal year.

This is a company that found the exact seam in the AI trade everyone's chasing — custom accelerator silicon plus optical and memory interconnects — and stitched itself into all three major U.S. hyperscalers for the first time. The NVIDIA (~$2 billion investment, NVLink Fusion) and AMD backing aren't cosmetic. On the fundamentals, this is a top-tier story.

So why am not I convinced?

Because the tape is ahead of the ledger. Marvell closed at roughly $263.27 on September 29, up about 210% year-to-date, on a market cap near $231 billion. It trades around 80x trailing adjusted EPS and 50 to 62x forward earnings. Analyst targets cluster in the $270s to $290s — Cantor at $330, Morgan Stanley at $268, a cluster around $270, average near $289 — which means the street thinks there's maybe 8% to 25% of upside left before you account for the fact that this stock just ran 210% in eleven months. GuruFocus threw out an intrinsic value of roughly $118.67, implying the shares sit about 100% above what the math supports.

That's the whole game, and it's a brutal one to play right now.

"Priced for perfection" is the laziest phrase in equities, but it's doing real work here. Marvell guided Q3 revenue to ~$3.15 billion (±5%) with EPS around $1.10. Miss either end of that band and you don't get a "we'll explain in the call" response — you get a multiple compression that eats the next four quarters of growth in one session. Beta of 2.25 doesn't lie; this is a levered bet on the entire AI-capex complex, not just one company's execution.

And the lumpiness isn't going away. Marvell's custom-silicon book is large, concentrated, and subject to the whims of four procurement teams who can all restructure a program on a Tuesday. Broadcom reportedly holds roughly 70% of the custom AI co-design market and serves Meta, Google, and the rest of the table. Marvell is winning, but it's winning in a ring full of fighters who never went to college.

Here's where I land: Marvell is a fantastic company trading at a price that requires it to be a flawless one. The Google warrant is the part the bulls keep forgetting — it aligns your biggest customer's incentives with your share count, not against it, and it caps the upside of every milestone they celebrate together. The optical and memory story is genuine and structural; the 2nm co-packaged optics demo at ECOC in September is the kind of thing that compounds over years, not quarters.

But you don't get paid for believing the bull case. You get paid for the price you pay for it. At 80x trailing and a stock up 210%, Marvell has already been paid. The easy money was made buying the three-hyperscaler expansion. What's left is a stock where a single soft quarter, a warrant-dilution overhang, or a broad risk-off day can turn a great business into a mid-year disappointment faster than you can finish this paragraph.

My call: this is a "prove it to me" name, not a "buy and forget" one. If you own it, you're not holding shares — you're holding optionality on AI capex with a 2.25 beta and a customer who owns a piece of your downside. Size accordingly. The story's real. The price is the argument.