Marvell Technology’s Quiet Ascent: How a Chip Company Rode Custom Silicon to a 300% Rally

="">

A year ago, Marvell Technology was a name most generalist investors couldn’t pick out of a lineup. It wasn’t Nvidia. It wasn’t Broadcom. It was the other chip company — the one headquartered in Wilmington, Delaware, with roots in Bermuda and a CEO who kept talking about custom silicon like it was the future of everything.

Turns out, he was right.

Marvell shares have surged more than 300% from their 2024 lows, a run that has placed the company squarely alongside the biggest beneficiaries of the artificial intelligence infrastructure buildout. As Business Insider reported, the stock’s trajectory has drawn comparisons to Nvidia’s own meteoric rise — though the underlying business model is fundamentally different, and in some ways more durable than Wall Street initially understood.

The story of Marvell’s resurgence is really a story about what happens when hyperscale cloud companies decide they don’t want to be entirely dependent on one supplier for their most critical computing components. And it’s a story about timing.

The Custom Silicon Thesis Takes Hold

For years, the AI chip market was synonymous with Nvidia. Jensen Huang’s company built the GPUs that trained the large language models, powered the inference engines, and generated the breathless headlines. Nvidia’s dominance was so complete that its market capitalization briefly surpassed $3 trillion, making it the most valuable company on Earth.

But beneath that narrative, something else was happening. Amazon, Google, Microsoft, and Meta were all investing billions in designing their own custom chips — application-specific integrated circuits, or ASICs — tailored to their particular AI workloads. They didn’t want off-the-shelf GPUs for every task. They wanted silicon that did exactly what they needed, nothing more, nothing less. And they needed a partner to help design and manufacture it.

Marvell positioned itself as that partner.

Under CEO Matt Murphy, who took the helm in 2016, the company pivoted aggressively away from its legacy businesses in storage and networking semiconductors and toward custom compute. Murphy made a series of acquisitions — most notably the $10 billion purchase of Inphi in 2021 — that gave Marvell critical capabilities in electro-optics and high-speed interconnects, the plumbing that connects chips inside data centers. The bet was that as AI clusters grew larger and more complex, the connections between chips would matter as much as the chips themselves.

That bet has paid off spectacularly. Marvell’s custom compute revenue has grown at a triple-digit percentage rate over the past several quarters. The company now counts all five of the major U.S. hyperscalers as customers for its custom silicon programs, a breadth of relationships that no other merchant semiconductor company can match.

“We are in the different position of being pulled by our customers,” Murphy told analysts on the company’s most recent earnings call. He wasn’t exaggerating. Marvell’s design pipeline — the custom chip programs that have been contracted but haven’t yet reached volume production — now stretches into the tens of billions of dollars over the next several years.

The numbers tell the story. In fiscal Q4 2025 (ending February 1), Marvell reported data center revenue of $1.8 billion, up 98% year over year. That segment now represents roughly 73% of total company revenue, up from less than 40% just two years ago. The transformation has been swift and, by semiconductor industry standards, dramatic.

Why Wall Street Took So Long to Notice

One of the more interesting aspects of Marvell’s rally is how late many institutional investors were to the trade. As Business Insider noted, the stock spent much of early 2024 trading sideways while Nvidia and Broadcom captured most of the AI enthusiasm. Several factors explain the delay.

First, Marvell’s non-AI businesses were shrinking. Revenue from carrier infrastructure, enterprise networking, and consumer electronics all declined as those end markets went through inventory corrections. The headline numbers looked messy even as the AI business was inflecting sharply higher. Investors who screened on total revenue growth missed what was happening underneath.

Second, there was genuine skepticism about the custom silicon market. Nvidia bulls argued — not unreasonably — that general-purpose GPUs would always win because of their software advantages, particularly the CUDA programming framework that locks in developers. Why would hyperscalers spend billions designing their own chips when they could just buy Nvidia’s?

The answer turned out to be economics. For specific, high-volume inference workloads — the kind that power search results, recommendation feeds, and chatbot responses — custom ASICs can deliver dramatically better performance per watt and performance per dollar than GPUs. Google’s Tensor Processing Units proved the concept years ago. Now everyone wants their own version.

Third, Marvell’s margin profile initially underwhelmed. Custom chip programs require significant upfront engineering investment, and the gross margins on custom silicon are typically lower than on proprietary merchant chips. But as programs ramp to volume, margins improve substantially. Marvell has guided for non-GAAP gross margins to reach 62-63% as the custom compute mix scales, a level that would compare favorably with most analog and mixed-signal peers.

The skeptics have largely capitulated. Analyst price targets have ratcheted higher throughout 2025, with several firms now projecting the stock can reach $150 or above, implying further upside from current levels. Bank of America recently reiterated its buy rating, citing Marvell’s “unmatched custom silicon design capabilities” and the growing backlog of hyperscaler programs entering production.

But the stock isn’t cheap. At roughly 45 times forward earnings estimates, Marvell trades at a premium to the broader semiconductor index and at a level that demands continued execution. Any stumble — a delayed program ramp, a lost customer, a margin miss — would be punished severely. This is a stock priced for perfection in a business where perfection is rare.

The competitive dynamics are also shifting. Broadcom, which has its own massive custom silicon business anchored by Google’s TPU program, remains Marvell’s most direct rival. And new entrants are circling. Intel’s foundry services division has pitched its manufacturing capabilities to hyperscalers looking to diversify away from TSMC. Samsung is doing the same. Even smaller design houses like Alphawave Semi are trying to grab pieces of the interconnect market that Marvell currently dominates.

Murphy has acknowledged the competition but argues that Marvell’s end-to-end design capabilities — from the compute die to the SerDes (serializer/deserializer) to the optical DSP (digital signal processor) — create a systems-level integration advantage that’s extremely difficult to replicate. “You can’t just bolt these things together,” he said at a recent industry conference. “The physics of the interconnect at these speeds requires co-design from the ground up.”

He has a point. At the data rates required by next-generation AI clusters — 200 gigabits per second per lane and beyond — signal integrity becomes a first-order design challenge. The companies that can optimize across the entire signal chain, from chip to module to optical fiber, have a meaningful advantage. Marvell’s Inphi acquisition gave it exactly this capability.

What Comes Next

The next twelve months will be critical for Marvell. Several of its largest custom silicon programs are expected to enter volume production, which should drive a significant step-up in revenue and earnings. Wall Street consensus estimates call for fiscal 2026 revenue of approximately $8.5 billion, up from around $5.8 billion in fiscal 2025. If the company hits those numbers, it will have more than tripled its data center revenue in just three years.

The broader question is whether the AI infrastructure spending cycle has legs. So far, every indication is that it does. Microsoft, Google, Amazon, and Meta have all announced capital expenditure budgets for 2025 that exceed prior-year levels, in some cases by 40-50%. Much of that spending is directed at data center construction and the chips that go inside them. Marvell sits at the center of that supply chain.

There are risks. A macroeconomic slowdown could cause hyperscalers to pull back on spending. Geopolitical tensions with China could disrupt supply chains or limit Marvell’s addressable market. And the custom silicon business model carries inherent concentration risk — losing a single large customer program could create a material revenue hole.

But for now, the momentum is unmistakable. Marvell has gone from an afterthought in the AI chip conversation to one of its central characters. The stock’s 300% rally reflects a fundamental re-rating of the business, not just speculative froth. Investors are paying for what they can see in the design pipeline and betting that Murphy’s team can convert those designs into shipping silicon at scale.

It’s the kind of transformation that doesn’t happen often in the semiconductor industry. Companies rarely get to reinvent themselves this completely, this quickly. Marvell did it not by chasing the loudest trend but by positioning itself where the money was actually flowing — into the custom infrastructure that makes AI work at hyperscale.

The question now isn’t whether Marvell belongs in the AI conversation. It’s whether the stock’s valuation already reflects the best-case scenario, or whether there’s still room to run as the custom silicon market expands beyond even the most optimistic projections. For a company that spent years in Nvidia’s shadow, that’s a remarkably good problem to have.