Marvell Technology, Inc. (MRVL)

For decades, the semiconductor industry was visible in consumer products — the processors in computers and phones, the memory in laptops. Marvell Technology operates almost entirely out of sight, designing chips that sit at the backbone of data centres, cloud providers, and storage systems where companies stash their photographs, emails, and records. The business is unglamorous but essential, and its fate is tethered to how much computing infrastructure the world decides to build.

The obscure foundation of modern computing

When you upload a photograph to the cloud or access a document from anywhere in the world, a chain of invisible infrastructure has to route, store, and retrieve that data. Marvell sits inside that chain. The company designs chips that control how data moves through storage systems, how servers connect to networks, and how the enormous processor farms that run artificial intelligence and large-scale data processing talk to each other. It does not make the chips itself — none of the modern fabless design companies do — but it owns the intellectual property, sets the specifications, and licenses the design to manufacturers like TSMC and Samsung.

Sehat Sutardja founded Marvell in 1995 in Santa Clara, in the heart of Silicon Valley, with the idea of building high-performance controllers for hard disk drives. At that moment, hard drives were multiplying in computers and data centres, and the controller chip — the piece that orchestrated how data moved on and off the spinning platters — was a bottleneck. Marvell’s early success came from solving that problem. But a smart company does not stay wedded to a single product when the underlying market shifts beneath it. As solid-state drives began to replace hard drives, Marvell pivoted. As data-centre operators began buying not just individual servers but entire systems, Marvell shifted from disk controllers toward the broader infrastructure that made those systems work: networking chips, storage-area-network controllers, and processors designed for specific jobs inside the data centre.

That sequence of pivots is what has kept Marvell relevant and valuable across three decades of radical hardware change.

From storage to infrastructure at scale

Marvell’s business is now organised around three main segments: data centre and infrastructure, storage, and client.

The data centre and infrastructure segment is the crown jewel. This is where the company designs chips that manage how servers talk to each other, how they connect to storage, and how they move vast quantities of data. Cloud providers like Amazon Web Services, Microsoft Azure, and Google Cloud — companies that have built their entire business on the ability to run compute at scale — are the largest customers for these chips. So are enterprise data-centre operators who run their own internal cloud systems. As companies shift more of their IT to cloud platforms and as artificial intelligence workloads demand denser and faster connectivity between processors, the value of having a specialised chip that can handle those connections grows.

The storage segment serves the market for controllers that sit inside storage systems — the machines that hold terabytes or petabytes of data in data centres. A bank’s mortgage records, a streaming service’s film library, a social-media company’s user photographs: all of it sits on systems where Marvell chips orchestrate the flow of reads and writes.

The client segment is the smallest, addressing applications like personal computers and mobile devices — markets where Marvell competes against larger rivals and where the margin pressure is higher.

The composition of Marvell’s revenue reflects the company’s migration away from consumer-facing markets toward the infrastructure that serves those markets from behind the scenes. Data centre and infrastructure is the largest and fastest-growing segment. Margins are wider in infrastructure chips because the customer — a cloud provider or data-centre operator — is buying performance and reliability, not checking a consumer price tag.

Why this business is capital-intensive without owning a factory

Marvell does not own a semiconductor fabrication plant. Instead, it designs chips at its offices in California and other locations, then contracts the actual manufacturing to foundries — primarily Taiwan Semiconductor Manufacturing Company (TSMC) and Samsung. This fabless model is standard for many chip-design companies because building and maintaining a fab is enormously expensive and requires continuous heavy investment just to stay current with the latest manufacturing technology.

But fabless does not mean capital-light. Designing the chips that customers demand requires hundreds of engineers, spending on research and development year after year, and the ability to take on risk when you commission a large production run and discover that the chips have a flaw or that the customer’s product roadmap changed. Marvell has invested billions in acquiring other chip companies to expand its product lines — acquiring Cavium in 2017, acquiring Inphi in 2021 — which accelerated its shift from storage-focused toward a broader infrastructure player.

Those acquisitions have also introduced integration risk. Marvell has had to consolidate product lines, ensure that the acquired companies’ customers are retained, and avoid losing talented engineers who might be poached by competitors. The 2021 acquisition of Inphi, a maker of infrastructure chips for data centres and hyperscalers, was particularly significant because it doubled down on the company’s direction toward cloud and data-centre markets. But the complexity of integrating that large acquisition — and of shipping new, combined products to major customers who expect high reliability — has been a drag on execution and margins.

How customers drive the roadmap

Marvell’s fortunes move with the capex cycles of its largest customers. When cloud providers and data-centre operators are in expansion mode, building new regions and upgrading infrastructure to handle more compute, they buy new systems and the chips inside them. When they are in pause mode, ratcheting back spending and trying to wring more value from existing hardware, Marvell’s orders soften. This is not Marvell’s fault — it is the nature of selling infrastructure to customers whose spending depends on their own growth expectations.

The second driver is the march of Moore’s Law and the demands it places on chip design. Every couple of years, the semiconductor industry shifts to a smaller manufacturing process node, which lets designers fit more transistors on a chip and build faster, lower-power designs. But each node shrink requires new masks, new designs, and new engineering. Marvell has to keep pace with TSMC’s roadmap and with its own customer roadmaps, or risk falling behind competitors like Broadcom and Cavium (before the Marvell acquisition). This perpetual race is expensive, which is why a fabless company’s survival depends on a steady stream of large, wealthy customers who will fund development.

Pressures and the path forward

The largest cloud providers — Amazon, Microsoft, Google — have begun designing their own infrastructure chips tailored to their specific workloads, which has eroded some of Marvell’s addressable market. This move by hyperscalers is natural: if you control your own data centres at the scale that Amazon does, designing your own processor for your own use case can give you an edge. But it means Marvell is selling not to all cloud providers but to those who choose to buy rather than build.

A second pressure is geopolitical. Marvell depends entirely on TSMC, which operates in Taiwan, a company at the centre of US-China tensions over semiconductor manufacturing. Any disruption to Taiwan or any restriction on US companies’ access to TSMC’s advanced nodes would threaten Marvell’s ability to manufacture its latest designs. This risk is existential for any US fabless company that relies on Taiwan.

Finally, the transition from artificial-intelligence inference (running trained models) to training (building those models in the first place) is reshaping the data-centre market. Training workloads are less forgiving about latency and more demanding of raw compute density, which favours certain chip architectures over others. If Marvell’s designs do not track this shift, its share of wallet with key customers could decline.

Reading the company

Anyone studying Marvell should start with the annual 10-K filing (SEC CIK 0001835632), which breaks down revenue by customer segment and by geography and explains the company’s dependence on large cloud providers. The earnings calls are the place to watch for: the health of orders from major customers, the trajectory of margin as the company digests acquisitions, and commentary on the company’s competitive position in new workloads like artificial intelligence.

The price-to-earnings ratio shows how the market is valuing Marvell against other semiconductor companies and against its own history. Any multiple compression often signals concerns about customer spending or market share. Revenue growth in the data-centre and infrastructure segment is the single most important metric, because that is the strategic focus and the highest-margin business. And pay attention to gross margin, which is vulnerable to customer mix and to manufacturing costs — if TSMC’s prices rise, that pressure flows down to Marvell.