The $120 Billion Ceiling That Isn't a Floor

Marvell's stock fell 8% in premarket trading on Tuesday. The company had just beaten earnings, raised its outlook, and its CEO had spent the week calling a Google partnership "game-changing." None of that mattered because investors had already priced in a version of the future that even the company's own guidance couldn't confirm.

That disconnect between what the Marvell narrative promises and what its numbers commit to is the real story — not the headline, not the warrant, and not the $120 billion figure that's been recirculated as though it were a signed contract.

Here's what actually happened. On August 19, Marvell disclosed an expanded commercial agreement with Google covering custom silicon products — inference accelerators, storage controllers, network interface controllers, memory interface controllers, and near-memory compute — attached to Google's TPU ecosystem. Attached to the deal is a warrant giving Google the right to buy up to 58.97 million Marvell shares at $206.58 each. Fully exercised, that stake would be worth $12.2 billion and represent up to 7% of Marvell's total shares.

The warrant doesn't vest on a schedule. It vests on Google's spending.Only 1.36 million shares — about 2.3% of the total — vest automatically in quarterly installments over the first year. The remaining 57.6 million shares unlock in 240 tranches — roughly 240,000 shares for every $500 million in qualifying chip purchases, stretched through fiscal 2033. For all 240 tranches to vest, Google would need to spend a cumulative $120 billion with Marvell.

The SEC filing describes Google's purchases as discretionary.

This isn't a $120 billion order book. It's a ceiling — the maximum amount of revenue Marvell would need to deliver to Google over six years for the warrant to fully vest. Every tranche between now and then is Google making a buying decision, not a contract obligation. For Marvell, it's an equity-linked volume rebate: if Google directs more spending its way, Google gets equity upside. The demand risk sits with the supplier, not the buyer.

Marvell CEO Matt Murphy wouldn't leave it at that. On the recent earnings call, he said the total opportunity is "massive for Marvell and game-changing at the sort of peak performance of what could be achieved now over the next six and a half years". He went further: "You should assume in that timeframe that on the custom side, these numbers would be a lot larger than overall custom than anybody has been modeling so far".

Executive language about what the market hasn't modeled is usually the moment to check what the market has already priced.

Marvell's shares are up 162% year-to-date and 260% over the trailing twelve months. The stock, at roughly $223 today, carries a 74x trailing P/E multiple and a market cap of $195 billion. For context, Broadcom — Google's primary TPU partner with a long-term deal through 2031 — trades at a 61x P/E despite being a $1.78 trillion company with $8.4 billion in quarterly AI revenue alone. Marvell's gross margin is 52%, operating margin 17%, and return on invested capital just 5.5%. The premium isn't being justified by current profitability. It's being justified by the arc of growth the market believes is coming.

So what are those numbers? Marvell's fiscal 2026 revenue was $8.2 billion, up 42% year-over-year. The latest quarter — Q2 of fiscal 2027 — came in at a record $2.74 billion, with data center revenue climbing 46% to $2.17 billion, now 79% of total revenue. The company raised its FY2028 revenue forecast from $16.5 billion to approximately $18 billion, implying 50% growth. Custom chip revenue is projected to top $10 billion by FY2029.

Those are solid numbers. They're just not the numbers that got the stock here.

The 8% post-earnings drop came precisely because the $18 billion FY2028 raise — higher than previously modeled — didn't go far enough for a market that had already absorbed the "game-changing" thesis. Goldman Sachs analysts described investor expectations as "elevated" heading into the quarter, driven by the Google partnership and "exceptionally robust" AI bookings. The raised outlook underwhelmed because the market's internal model ran larger than what Marvell could commit to on its own call.

There's a structural reason to question whether that $120 billion ceiling is even a realistic target for a company doing $8.2 billion in annual revenue today. To hit $120 billion in cumulative revenue over six years, Marvell would need Google alone to average $20 billion per year — more than double the company's entire current revenue base. Even if you factor in that the $120 billion covers all Marvell products within the Google TPU ecosystem (not just custom ASICs), it still requires Marvell to more than double its addressable share of Google's silicon spending from today's level.

That's not impossible. It's just an enormous ask.

The deal does something structurally useful, though: it converts what was previously an invisible pipeline — custom chip programs that investors could only guess about — into a measurable framework. Every time a tranche vests, Google's warrant exercise reveals how much qualifying revenue has accumulated. It's a window into Google's actual purchasing behavior that didn't exist before. The warrant gives Marvell a customer success story and gives Google an equity hedge that profits if the supplier succeeds. It's a smart structure for both parties, but it's not a revenue guarantee.

The competitive context matters too. Marvell's products sit in the "scaffolding" around Google's in-house TPU, which is designed by Google and fabricated by TSMC alongside Broadcom. Marvell isn't building the main AI accelerator. It's building the components that connect, store, and feed data to it. Broadcom, by contrast, is developing and supplying future generations of Google's custom AI chips through a long-term agreement through 2031. When Marvell announced its Google deal, Broadcom's stock fell more than 5% — but the Morningstar analyst assessment was probably right: this is Google diversifying its supply chain, not Marvell replacing Broadcom. There's room for both if the pie keeps growing.

The pie is growing. US hyperscalers are expected to spend more than $700 billion on AI infrastructure this year, up from roughly $400 billion in 2025. Marvell has custom engagements with all major US hyperscalers and is riding that wave. Data center revenue alone grew 46% year-over-year last quarter.

But the stock is already pricing a very specific version of how this plays out — and the margin for disappointment is narrowing.

The question for investors isn't whether Marvell is well-positioned. The data center trajectory, the expanding custom chip business, and the Google warrant structure all point to a company at the right place in the AI infrastructure build-out. The question is whether $195 billion is the right price for a growth path that hasn't been proven at this scale yet, for a company whose CEO says the market has been "modeling" numbers too small, even as the market's own model just proved too large for the company's raised guidance.

The warrant tranches will provide the evidence. Until Google starts vesting shares at a rate that confirms the $120 billion ceiling is a reachable target, not just a theoretical maximum, the story remains a story. Marvell's investors are being asked to hold the line on a 74x P/E multiple while the engineering reality — which tranche vests next, how much revenue it represents, whether the custom chip margin profile holds at scale — catches up to the executive narrative.