Marvell Trades Richer Than Broadcom to Be the Runner-Up in Custom AI Chips

---

Begin with the genuine strength of the business, because Marvell is riding one of the most powerful trends in technology and the critique here is about relative value and structural fragility, not quality. In its most recent quarter, Marvell reported record revenue of $2.42 billion, up 28% year over year, with record data-center revenue of roughly $1.83 billion — about 76% of the entire company — having more than doubled year over year. Its custom-silicon business reached roughly a $1.5 billion annual run rate, and the broader custom-ASIC market it serves is forecast to grow around 45% this year. It counts the largest cloud companies on Earth — Alphabet, Amazon, Microsoft — among its custom-silicon customers, and in a striking vote of confidence, Nvidia itself invested $2 billion in the company. This is a real, fast-growing, strategically important business sitting in the slipstream of the biggest capital-spending wave in history.

The custom-silicon opportunity is genuinely attractive. As the hyperscalers spend hundreds of billions of dollars on AI infrastructure, they have powerful incentives to design their own chips — tailored to their specific workloads, and free of the enormous margins Nvidia charges on its merchant GPUs. Designing a competitive AI chip is extraordinarily hard, so the hyperscalers partner with specialist firms that bring the design expertise, the intellectual property, and the relationships with the foundries that manufacture the chips. That specialist business is a wonderful place to be: high-value, deeply embedded, riding the hyperscalers' spending. And there are only two firms that do it at real scale. Marvell is one of them.

So this essay does not argue that Marvell is a bad business. It is a good one. It argues that the market has priced it in a way that ignores two uncomfortable facts: that it is the structurally smaller player in a duopoly dominated by someone else, and that its business is concentrated in a handful of customer programs that the customers themselves are increasingly capable of taking back.

Paying more for less of the market

Start with the valuation comparison, because it is genuinely odd and it frames everything. In the custom-AI-chip duopoly, Broadcom is the overwhelming leader, commanding more than 70% of the market for custom AI accelerators, with a customer roster that reportedly includes Google, Meta, OpenAI, Anthropic, and now Apple. Marvell is the challenger, targeting roughly 20% share — a real and valuable position, but a fraction of the leader's. If you were going to pay a premium multiple for one of these two companies, the intuitive choice would be the dominant leader with the larger share, the broader customer base, and the stronger position.

The market has done the opposite. After recent moves in both stocks, Marvell trades at a higher forward earnings multiple than Broadcom — the smaller, second-place company commanding a richer valuation than the dominant leader. The logic, such as it is, rests on growth: because Marvell is smaller and growing from a lower base, its percentage growth rates are higher, and the market extrapolates that catch-up into a premium. But this is precisely the error documented elsewhere in this series in the AMD-versus-Nvidia comparison — paying up for the challenger's trajectory while the incumbent's dominance is available more cheaply. The challenger's faster percentage growth is real, but it is growth toward a share target (20%) that still leaves Broadcom three-and-a-half times its size, and paying a premium to the leader for the privilege of owning the follower assumes the follower closes a gap it has not closed and that the leader is not defending. When you can buy the 70%-share leader for a lower multiple than the 20%-share challenger, the challenger has to do something genuinely extraordinary to justify the inversion — and "grow fast from a small base" is not extraordinary; it is what small bases do.

The customer concentration that defines the risk

The deeper vulnerability is structural, and it is the defining feature of the custom-silicon business: extreme customer concentration. Marvell's custom-AI-silicon revenue does not come from a broad, diversified base of thousands of customers; it comes from a small number of enormous hyperscaler programs — a chip for Amazon's AI infrastructure here, a chip for Microsoft's there, a program with Google. Each of these programs is large, which is what makes the business attractive, but the small number of them is what makes it fragile. When your revenue depends on a handful of mega-projects with a handful of mega-customers, the loss, delay, or downsizing of any single one can crater your results in a quarter.

This is not a theoretical risk; it is the central risk that Marvell's own disclosures flag. A slowdown in hyperscaler spending, a shift by a customer to in-house design, or a delay in a single chip program could feed through to results quickly and severely, precisely because the business is concentrated. Custom silicon is a project business dressed in the clothes of a product business: each major customer relationship is essentially a large, multi-year contract, and the portfolio of those contracts is small enough that the failure of one is material. A company whose value rests on a few enormous, lumpy, renegotiable programs is a fundamentally riskier and more cyclical thing than the smooth, diversified, recurring-revenue business the rich multiple implies — and the concentration cuts especially deep for the second-place player, who has fewer programs than the leader and therefore less diversification to absorb the loss of any one.

The customers are also the competition

Here is the subtlest and most important risk, the one that makes custom silicon a structurally precarious business no matter how well Marvell executes: Marvell's customers are also, increasingly, its potential competitors. The whole reason a hyperscaler designs custom silicon is to internalize capability and reduce its dependence on outside suppliers — first on Nvidia, but ultimately on anyone, including the custom-silicon partner itself. Amazon, Google, and Microsoft are hiring chip-design talent aggressively and building deep in-house silicon teams, and as those teams mature, the hyperscalers need their design partners less, not more. The same impulse that creates Marvell's business — the desire to own one's own chips — is the impulse that, taken to its conclusion, eliminates the need for a Marvell.

This is the existential tension of the custom-silicon model. Marvell is helping its customers build the capability to eventually not need Marvell. In the near term, the hyperscalers need the partner's expertise and the work is enormous; in the longer term, every chip program a hyperscaler completes with Marvell teaches that hyperscaler more about how to do the next one itself. A company whose customers are actively working to develop the ability to replace it is in a fundamentally weaker position than its current growth suggests — and again, this risk falls hardest on the runner-up, because a hyperscaler choosing to in-source is more likely to keep a relationship with the dominant, hardest-to-replicate leader than with the smaller second source. Marvell's chips also compete, ultimately, with Nvidia's merchant GPUs, which means a shift in the build-versus-buy calculus at any hyperscaler — toward buying Nvidia, or toward building entirely in-house — squeezes Marvell from both sides.

Why is Nvidia funding its own competition?

The $2 billion investment Nvidia made in Marvell deserves a harder look than the bullish framing gives it, because on its face it is strange: Nvidia is the company whose dominance custom silicon exists to erode, and Marvell builds the very chips that help hyperscalers buy fewer Nvidia GPUs. Why would Nvidia write a nine-figure check to a company working to reduce demand for Nvidia's products?

There are a few readings, and none of them is purely the ringing endorsement the bulls present. The charitable version is that Nvidia sees custom silicon as inevitable and would rather have a stake and a relationship in it than be shut out — and that Marvell's optical-interconnect technology, which Nvidia needs for its own systems, is genuinely valuable to it. That is plausible. But the more forensic version notes that this is one more strand in the increasingly circular web of cross-investments binding the entire AI industry together — the same pattern seen in AMD handing OpenAI equity warrants, in Nvidia's tangle of investments across the labs and suppliers it also sells to, in the hyperscalers funding the startups that fund their cloud revenue. When the dominant chipmaker invests in a custom-silicon rival, it is partly buying influence, partly hedging, and partly participating in the mutual-validation machine in which everyone invests in everyone and every investment is cited as proof that the boom is real. A strategic investment from Nvidia is a meaningful data point, but it is not the same as durable demand from end customers, and treating a competitor's hedge as confirmation of the thesis is exactly the kind of circular reasoning this AI cycle runs on. The check is real. What it proves about Marvell's long-term independence from the same customers and competitors it is entangled with is considerably less than the headline suggests.

What the bulls genuinely get right

In fairness, the bull case is strong and Marvell's position is genuinely valuable — the debate is the price and the structural risks, not the quality. Several points cut in Marvell's favor. The custom-silicon market is real, enormous, and growing fast (forecast around 45% this year), and a strong number two in a booming duopoly is a very good place to be — there is more than enough demand for both Broadcom and Marvell to thrive for years. Marvell's data-center growth of 109% is extraordinary and reflects genuine design wins and real customer demand, not financial engineering. The technology is real and broad: beyond custom AI silicon, Marvell has a strong position in optical interconnect — the high-speed links that connect chips and servers in AI data centers — which is its own large and growing opportunity. Nvidia's $2 billion investment, whatever its strategic motivation, is a meaningful endorsement from the most important company in AI. And the hyperscalers' incentive to design custom chips is durable and powerful, which underpins demand for design partners for the foreseeable future. Marvell is a high-quality company executing well in a great market.

The honest synthesis is that Marvell is a fine business that the market has priced as something better than its structural position warrants — a premium-to-the-leader valuation on a second-place player whose revenue is concentrated in a few customer programs that the customers themselves are learning to replace. None of that means Marvell stumbles soon; the AI capex wave is powerful and the demand is real, and Marvell could grow handsomely for years. It means the price embeds an assumption — that the runner-up deserves a premium to the leader and that the concentration and in-sourcing risks will not bite — that the structure of the business does not support. The bull is right that Marvell is a great way to play custom silicon. The question the valuation poses is why you would pay more for the smaller player than the market is asking for the larger one.

The shared chokepoint and the cyclical truth

There are two final structural facts worth naming, because they apply to the whole custom-silicon trade and the premium multiple ignores both. The first is the foundry chokepoint: both Marvell and Broadcom depend entirely on TSMC to actually manufacture their chips at the leading edge, which means they share a single critical supplier whose capacity is contested by every advanced-chip company on Earth, including Nvidia and Apple. Neither design firm controls its own manufacturing, and in a world of constrained leading-edge capacity, the design partners are at the mercy of the foundry's allocation and pricing — a dependency that caps how much value they can capture and exposes them to a bottleneck entirely outside their control.

The second is cyclicality. The custom-silicon boom is a function of the hyperscaler capital-spending cycle, and that cycle, however powerful right now, is a cycle. The same hyperscalers documented elsewhere in this series — spending $200 billion, $190 billion, hundreds of billions collectively on AI infrastructure, with free cash flow already strained — will not spend at this pace forever, and when their capex growth decelerates, the lumpy, concentrated, project-based revenue of their silicon partners decelerates with it, amplified by the leverage of concentration. Marvell's 109% data-center growth is a peak-of-cycle number riding a peak-of-cycle spending wave, and a premium multiple applied to peak-cycle growth is the classic setup for the double disappointment of falling earnings and a falling multiple when the cycle turns. The business is excellent. The price assumes the cycle, the share gap, the concentration, and the in-sourcing risk all resolve in Marvell's favor — a stack of optimistic assumptions sitting on top of a genuinely good but genuinely second-place company.

The optical business is the real hedge — and it's not what's being priced

If there is a part of Marvell that genuinely deserves a premium and that the bull case under-emphasizes, it is the optical-interconnect business rather than the custom-silicon one. As AI data centers scale, the chips have to talk to each other at enormous speeds, and the optical components and connectivity that link them are a large, growing, and more diversified market than the concentrated custom-ASIC programs. Marvell has a genuinely strong franchise here, and it is arguably a higher-quality, less customer-concentrated, more defensible business than the custom-silicon work that gets all the attention.

The irony is that the market's enthusiasm — and the premium multiple — is fixated on the custom-silicon story, which is the more concentrated and more competitively exposed half of the company, while the steadier optical business is treated almost as a footnote. An investor paying a premium-to-Broadcom valuation for Marvell is paying it primarily for the custom-AI-silicon narrative, which is the part most vulnerable to the concentration and in-sourcing risks described above. A more sober valuation would weight the durable optical franchise more heavily and the lumpy, contested custom-silicon programs less — but that is not the valuation the market has assigned, because narratives, not franchises, set multiples at the top of a boom. Marvell is a better business than the custom-silicon hype alone would make it; it is also a more expensive one than even that better business justifies.

The kicker

Marvell is a real winner in one of the best markets in technology, and nothing here disputes that it will likely keep growing as the hyperscalers keep designing their own chips. But the market has made a curious decision about how to price it: it has decided that the second-place company in a duopoly — smaller, more concentrated, more exposed to its customers turning into competitors — deserves to cost more than the dominant leader. That is the sort of inversion that happens when a market falls in love with growth rates and stops asking what the growth is worth, or how durable it is, or who actually controls the customers and the manufacturing. Marvell at a discount to Broadcom would be an easy thing to like. Marvell at a premium to Broadcom is a bet that being the runner-up, in a business where the customers are building the capability to need no one, is somehow the better place to be — and the history of paying up for second place, in this industry and every other, is not a happy one.

The hyperscalers hired Marvell to help them stop depending on Nvidia, and Marvell built a wonderful business doing it — for now. But the same customers are learning, program by program, to do the work themselves, the leader owns more than three times the share at a lower price than the follower, and a single delayed or cancelled program can take an entire quarter down with it. The market looked at all of that and decided the runner-up was worth a premium to the champion, which is the kind of decision that makes sense only as long as the music, and the capex, keep playing — and the one certainty about capital-spending cycles is that, eventually, the music stops, and the second source, carrying the premium it was somehow bought at, is always the first to feel the silence.