Marvell Calls Itself AI's 'Switzerland.' The Qualcomm-Amazon Deal Just Tested That Claim.

On Tuesday, Qualcomm locked down Amazon as a custom AI-chip customer in a deal that looks less like one order and more like a decade of them. The collaboration is reported to be worth up to $60 billion of potential purchases through 2036, with Amazon receiving a warrant to buy Qualcomm stock that vests as those orders land. Qualcomm shares jumped almost ten percent on the news. And Marvell — whose CEO, Matt Murphy, has taken to calling his company the "Switzerland" of the AI era — kept climbing too.

That pairing is the setup for one of the AI trade's cleanest dividing questions, and it matters most to anyone late to a stock that has already climbed 165% this year. Is the neutrality real — does Marvell actually win regardless of which hyperscaler wins the custom-silicon race? Or is "Switzerland" a reassuring phrase doing the work of a business that is now more contested, not less? A company trading near 21 times trailing sales and up 165% year to date no longer gets the benefit of the doubt on a slogan alone.

A third player steps into Marvell's one narrow lane

To see why the Qualcomm deal tests Marvell directly, you have to look at what it contains. The two things Qualcomm and Amazon agreed to build together are silicon and high-bandwidth optical interconnects rated up to 1.6 terabits per second. Those are not Marvell-adjacent products. Custom accelerators for running trained models and the optical networking that ties AI clusters together are the two pillars of Marvell's data-center business — which is now 79% of its revenue. Amazon, meanwhile, is one of the hyperscalers Marvell already counts as a customer.

Think of the AI chip market in two stages. Training a model runs almost entirely on Nvidia, where the CUDA software moat is strongest. Running the model afterward — inference — is where latency, efficiency, and cost-per-watt decide who wins, and that is exactly the stage where hyperscalers want chips tailored to their own workloads. Custom silicon is the inference answer, and until now the outside firms a hyperscaler could hire to design those chips were, for practical purposes, two: Marvell and Broadcom. Qualcomm's entry with Amazon introduces a third, well-funded player into that narrow design-partner lane, on Marvell's own turf and at the very moment the contested share of the AI cycle is shifting toward inference.

The "Switzerland" framing Murphy has offered shares the spotlight with Nvidia's Jensen Huang on a Computex stage earlier this year is a positioning claim, and a coherent one: Marvell plays both sides, building the bespoke accelerators hyperscalers ask for while also supplying the networking fabric that ties those accelerators to Nvidia systems, so it takes a cut from whichever architecture wins. But positioning is a statement about market structure. It is not a financial result.

What the operating record actually shows

By that standard, the delivered numbers are genuinely strong. Marvell's fiscal second quarter — reported at the end of August — brought record revenue of $2.74 billion, up 37% year over year, with data center growing 46% and accounting for 79% of the total. Management raised full-year fiscal 2027 revenue to roughly $12 billion and fiscal 2028 to roughly $18 billion, with data center expected to grow about 60% and its custom-silicon line expected to more than double. On its own, this reads as a company still accelerating into the inference ramp, not one being overtaken by a rival.

The discipline, though, is to separate what has reached revenue from what is still a claim. The $2.74 billion quarter is delivered. The $18 billion fiscal 2028 figure is guidance — a forecast, not a result. Marvell has already folded the upside of an expanded commercial agreement with Google into its forecasts, with the Google-driven revenue positioned as fiscal 2029 upside. And it is committing roughly $1 billion in capacity prepayments to suppliers this year. Supply commitments of that size are a dual signal: they are evidence of demand strength, and at the same time they are rising delivery and leverage risk. If the hyperscaler orders behind those prepayments slow or slip, Marvell has already spent on capacity that has to be filled.

The cleanest way the market could have answered the question Tuesday — by dumping Marvell on Qualcomm's Amazon win — did not happen; the stock held up and kept rising into a fresh record stretch. That is the market's verdict that the AI custom-silicon pie is expanding fast enough for both firms to grow. It is a reasonable read. But there is a difference between "Marvell still grows" and "Marvell grows enough to justify what it now costs."

The judgment that matters

Here is where the valuation enters, and it changes the shape of the decision. At roughly 21 times trailing sales and around 70 times trailing EBITDA, Marvell is priced as if the entire fiscal 2028 ramp and the years after it are near-certain. Broadcom, its principal peer in custom silicon, sits near 20 times sales and 35 times EBITDA. The market has not just seen this thesis — it has fully digested it. This is precisely the situation where a crowded, fully-priced winner stops being the obvious call and becomes a question of opportunity cost.

So the honest answer to Murphy's "Switzerland" line cuts both ways. The neutrality is real in the sense that Marvell is not hostage to any single hyperscaler's fate, and its operating results currently back up the story — record revenue, accelerating data center, a clean balance sheet with roughly a billion dollars of net debt. That gives the stock a genuine floor of credibility. But being neutral and being a winner are not the same thing, and the Qualcomm-Amazon deal is evidence that the inference phase is now contested enough to deserve weighting in the decision rather than dismissal. A slogan does not become an operating result because a share price keeps climbing.

For a retail investor new to Marvell, the useful habit is to split the delivered from the priced-in. The delivered quarter is excellent. The priced-in future is enormous. When the two are that far apart and a new competitor has just landed on the exact products that future depends on, the right question is no longer whether Marvell is a real AI-cycle company — it clearly is — but whether the near-term return curve out of 21 times sales still beats the alternatives. The answer to that is not "Switzerland." It is an allocation decision, best made with the understanding that this stock's current price already assumes a great deal of the good news is going to happen.