The Showcase: Marvell Technology (MRVL)

The AI trade may begin with compute, but it compounds through connectivity

MARVELL TECHNOLOGY

The Order Book Spoke First

Jensen Huang did not make Marvell Technology a trillion-dollar baby from a Computex stage. What he did was more useful. He confirmed, in public, what the order book and price action had already begun to say in private.

When Huang gestured toward Marvell CEO Matt Murphy and called Marvell “the next trillion-dollar company,” the market reacted violently. Marvell added roughly $30 billion in market value in the following twenty-four hours and climbed more than 50% before the week was over. That kind of move is easy to dismiss as keynote theater and over-romanticize. Huang is not a neutral analyst. Nvidia has strategic reasons to elevate Marvell, and every public endorsement from the most powerful semiconductor CEO on earth should be read with that conflict in mind.

Marvell had already banked eighteen hyperscaler custom silicon design wins. Management had already announced a $75 billion lifetime revenue funnel. Interconnect growth guidance had already been raised from 50% to 70% inside a single quarter. Revenue expectations had already moved higher. Wall Street earnings estimates already implied profit base could expand dramatically by the end of the decade.

The market did not discover Marvell at Computex. Computex gave investors permission to believe what the data had been broadcasting for months.

The Quant Was Early

Before the keynote, before the market repriced the story, Marvell was already behaving like an institutional accumulation name.

In the AlphaApes relative strength system, Marvell spent much of April and May inside the top ten to fifteen names out of more than 3,500 U.S.-listed stocks. In several sessions, it ranked inside the top three. Its RS Price score, which measures price return momentum across daily, weekly, monthly, and quarterly horizons, currently sits in the 99th percentile of the universe.

High relative strength over a sustained period is rarely random. It does not prove the thesis, and it certainly does not eliminate valuation risk. But it tells you sophisticated capital was positioning before the public narrative caught up.

This is the proper framing for Huang’s comment. He was not creating the thesis. He was confirming the tape.

Three Layers, One Moat

Most investors who recognize Marvell know it as a custom chip designer. That is true, but too narrow. The bull case rests on three layers of AI infrastructure, each reinforcing the others.

The first layer is custom silicon. Marvell designs application-specific chips, often called XPUs, for hyperscalers that want better economics than a general-purpose GPU can provide. These programs are not simple component sales. They are multi-year co-design engagements built around advanced process nodes, high-speed SerDes, and deep integration with a customer’s infrastructure roadmap.

Eighteen wins across more than ten customers is not proof of inevitability, but it is evidence that Marvell is not being treated as a replaceable supplier. The custom silicon business has already scaled to a meaningful run rate, and the $75 billion lifetime revenue funnel gives the company a credible path to much larger scale if those programs ramp as expected.

The second layer is networking and interconnect. Similar to how a band is only as good as its drummer, AI systems are only as useful as the data movement architecture around them. Chips have to communicate with chips, servers with servers, clusters with clusters, and data centers with other data centers. Marvell’s Ethernet switch ASICs, DSPs, and interconnect products sit inside that problem.

This is the more durable part of the business because it is less dependent on which compute architecture wins. A data center built around Nvidia GPUs, Broadcom custom silicon, or Marvell-designed XPUs still needs high-performance networking. The AI trade may begin with compute, but it compounds through connectivity.

The third layer is silicon photonics. This is the least understood and potentially most important part of the story. At AI data center scale, copper becomes a constraint. It consumes power, generates heat, and struggles with bandwidth over distance. Optical connectivity addresses that bottleneck by moving data with light rather than electrons.

Marvell’s role in digital signal processors and co-packaged optics puts it near one of the defining constraints of the next AI infrastructure cycle: not whether companies can buy enough accelerators, but whether they can move enough data at tolerable power cost.

The partnership is not simply about attaching Marvell to Nvidia’s halo. It places Marvell inside the architecture Nvidia is building for semi-custom, rack-scale AI infrastructure.

Why Inference Changes the Economics

The strongest version of the Marvell bull case does not require Nvidia to lose. It requires inference to scale.

Training large AI models is compute-intensive, capital-intensive, and still dominated by Nvidia GPUs. That is Nvidia’s castle, and it is not falling tomorrow. Inference is different. Once models are trained, they have to be run constantly across billions of queries. At that scale, cost per query, power consumption, and workload-specific efficiency become decisive.

This is where custom silicon becomes more compelling. A chip designed for a specific workload can produce better infrastructure economics than a general-purpose accelerator built to handle everything. As inference grows, hyperscalers have stronger incentives to design specialized silicon into their own infrastructure.

Marvell benefits from that shift without needing to displace Nvidia outright. The company can sit inside a hybrid architecture where Nvidia remains central to the AI stack while custom chips handle workloads where specialization improves cost and efficiency.

However, Marvell is not necessarily the anti-Nvidia trade. It may be one of the ways Nvidia’s ecosystem adapts to the next phase of AI infrastructure.

What the Nvidia Investment Actually Means

Nvidia’s investment in Marvell is not charity, and it is not a neutral endorsement. It is strategy.

If inference workloads increasingly require semi-custom infrastructure, Nvidia has two choices. It can resist the transition and risk watching hyperscalers pull more of the stack in-house, or it can build an ecosystem that keeps Nvidia relevant even as custom silicon becomes more prominent.

NVLink Fusion points to the second path. By bringing Marvell into Nvidia’s rack-scale architecture, Nvidia gets a partner capable of supplying custom XPUs, scale-up networking, and optical connectivity within a broader Nvidia-controlled framework.

That is why Huang’s endorsement should not be read as prophecy. It should be read as positioning. Nvidia is acknowledging that the next phase of AI infrastructure will not be built on GPUs alone.

For Marvell, that is a major validation. It does not guarantee execution. It does not make the valuation safe. But it does place the company in the center lane of a structural transition that the largest player in semiconductors appears to be taking seriously.

The Math Behind the Trillion

The trillion-dollar claim is possible, not casual.

For Marvell to reach that level, several things have to happen at once. Revenue has to compound aggressively as custom silicon programs move from design wins into volume production. Data center revenue has to remain the dominant growth engine. Operating margins have to expand meaningfully as AI mix improves. The share count has to remain reasonably controlled. And the market has to keep awarding Marvell a premium multiple long after the current excitement cools. That is a demanding bridge.

The optimistic version is straightforward. Marvell grows into something closer to a scaled AI infrastructure franchise, earnings expand dramatically, and investors eventually value the company less like a cyclical semiconductor supplier and more like a strategic infrastructure platform. Under that scenario, the path to a trillion-dollar market cap exists.

But the phrase “path exists” is doing real work. This is not a valuation you stumble into. It requires years of execution and very little tolerance for disappointment.

That is the right way to understand Huang’s comment. He was not saying Marvell is worth a trillion dollars today. He was saying the order book contains a version of the future where it could be.

The Contrast Next Door

Broadcom is the obvious comparison, and it is also the obvious objection.

Broadcom is already what Marvell wants to be: scaled, profitable, deeply embedded with hyperscalers, and dominant in custom silicon. Its AI semiconductor growth remains extraordinary, and its operating model is far more mature. Any argument that Marvell is the better forward opportunity has to begin by admitting that Broadcom is the better business today. That is not a concession. It is the setup.

Broadcom is the incumbent. Marvell is the accelerating challenger. Broadcom has the margin structure, the installed base, and the dominant market position. Marvell has the sharper earnings revision story, more visible margin runway, and a partnership with Nvidia that makes its strategic relevance harder to dismiss.

The market reaction around the two companies’ recent earnings capture the difference. Broadcom delivered strength but held its forward AI forecast steady. In a market priced for constant upward revision, steady can look like deceleration. Marvell, by contrast, raised the relevant forward markers and received the most valuable public endorsement available in semiconductors.

Both businesses are real. Broadcom is the castle. Marvell is the army at the gate. The question for investors is not which company is better. It’s which company has more positive surprise left.

Where the Bull Case Can Break

The biggest risk is not that the story is fake, but that it’s real and already priced too aggressively.

At a valuation above 70 times forward earnings and more than 20 times forward sales, Marvell has very little room for ordinary execution. A delayed hyperscaler ramp, a margin disappointment, weaker interconnect growth, or a single large customer pulling more work in-house could reset the multiple quickly.

Customer concentration remains an honest problem. Eighteen design wins across more than ten customers sounds diversified, but revenue weight is not evenly distributed. If Amazon or another anchor customer accounts for a disproportionate share of the custom silicon ramp, the headline design-win count may overstate the durability of the revenue base.

Vertical integration is the longer-term threat. Google, Amazon, Microsoft, and Meta all have the engineering talent and capital incentive to internalize more of their silicon stack over time. Marvell’s co-design model gives it a strong position today, but hyperscalers do not like permanent dependency when the economics are large enough to justify control.

The networking and photonics businesses are more defensible against that risk. Even if hyperscalers build more compute silicon internally, they still need connectivity, switching, optical infrastructure, and power-efficient data movement. But custom silicon is the valuation accelerant. If that engine slows, the trillion-dollar math gets much harder.

The stock is not cheap just because the company is important–importance and upside are not the same thing.

The Trade

Marvell is not a clean value story. It’s not a hidden asset. It’s not early in the public narrative anymore.

It’s a high-expectation AI infrastructure compounder whose order book, customer base, and relative strength profile all point in the same direction. That combination deserves attention. It also deserves discipline.

The cleanest version of the thesis is this: Marvell sits at the intersection of three problems hyperscalers cannot avoid — custom compute, data movement, and optical efficiency. Inference makes those problems larger. Nvidia’s partnership makes Marvell more strategically visible.

But the stock now has to earn the story. At this valuation, the burden of proof has shifted from “is Marvell real?” to “can Marvell execute fast enough to justify what the market has already started to price?”

Huang described the destination. The order book explains why it’s plausible. The quant saw the market moving before the crowd had language for it. The distance still has to be traveled.