Marvell targets up to $90bn revenue by 2031 as custom AI silicon outlook explodes

Marvell Technology, Inc. (Nasdaq: MRVL) has laid out one of the semiconductor industry’s most aggressive long-term growth forecasts, targeting fiscal 2031 revenue of approximately $70 billion to $90 billion after generating $8.2 billion in fiscal 2026. The company also lifted its fiscal 2028 outlook to roughly $20 billion, above its previous expectations and above the Wall Street consensus cited by Reuters, as cloud customers expand custom accelerator programmes alongside rapidly growing optical, networking and memory connectivity requirements. Marvell shares jumped sharply following the investor-day presentation, while Broadcom Inc. also gained as investors interpreted the forecasts as further evidence that custom silicon is becoming one of the largest markets created by artificial intelligence infrastructure.

The scale of the forecast is difficult to overstate because the midpoint of Marvell’s fiscal 2031 range would be almost ten times its fiscal 2026 revenue. Management expects custom silicon to become one of the largest contributors, with programmes involving hyperscale customers such as Google, Amazon Web Services and Microsoft supporting a pipeline that could expand dramatically after 2029. Marvell is simultaneously projecting major growth from interconnect products, giving the company exposure to both the processors doing artificial intelligence work and the high-speed infrastructure moving data between those processors.

How realistic is Marvell’s jump from $8.2bn to as much as $90bn of revenue?

The first discipline is to treat the fiscal 2031 range as management’s long-term target rather than guaranteed revenue. Marvell would need extraordinary execution across several semiconductor cycles to move from approximately $8.2 billion in fiscal 2026 to even the bottom of the $70 billion range. Customers must deploy programmes on schedule, manufacturing partners must produce advanced chips at scale and artificial intelligence capital expenditure needs to remain strong enough to support years of continuing infrastructure expansion.

The target nevertheless has more foundation than a purely speculative total-addressable-market slide. Marvell already generated 42% fiscal 2026 revenue growth, and management subsequently reported robust artificial intelligence bookings and accelerating demand across custom silicon, electro-optical connectivity, switching and storage products. The company’s fiscal 2027 trajectory is therefore being supported by existing programmes before the larger hyperscaler designs expected later in the decade enter full production.

The unusual characteristic of Marvell’s opportunity is that it does not rely on defeating Nvidia Corporation in general-purpose graphics processors. Hyperscalers are increasingly designing proprietary accelerators optimised around their own workloads, while outsourcing significant engineering and semiconductor implementation work to specialised partners. Marvell participates in that process, allowing it to benefit when Google, Amazon Web Services or Microsoft decides a custom chip is economically preferable for part of its infrastructure.

Why are hyperscalers spending so aggressively on custom AI accelerators?

Nvidia’s processors remain extraordinarily capable, but dependence on a single external supplier can create both strategic and economic problems for the largest cloud companies. Custom chips can reduce cost for predictable workloads, allow customers to optimise memory and networking around internal software, and provide negotiating leverage when external accelerator prices remain high. The larger a hyperscaler becomes, the easier it is to justify the enormous engineering cost associated with developing silicon specifically for its own infrastructure.

Google has already operated its Tensor Processing Unit architecture for several generations, Amazon Web Services has expanded Trainium and Inferentia, and Microsoft is developing proprietary accelerator programmes. Those systems do not necessarily replace Nvidia across entire data centres, but they can capture portions of workloads where economics favour specialised architectures. Every percentage point of hyperscale computing shifted into proprietary silicon can translate into enormous semiconductor volumes because the underlying data-centre budgets now run into tens of billions of dollars.

Marvell’s role resembles that of an engineering and platform partner rather than a conventional merchant-chip vendor selling the same product to every customer. That model can produce deeper relationships and higher switching costs because designs become embedded in multi-year architecture road maps. It also creates concentration and execution risks, since the loss or delay of an individual hyperscale programme can materially affect revenue when each customer represents very large potential volumes.

How should investors interpret Marvell’s Google agreement and the $120bn figure?

The Google relationship demonstrates why the post-draft distinction between potential and committed revenue is crucial. Marvell issued Google a warrant structure in August under which most shares vest as qualifying custom-product revenue is generated, with the full performance-linked framework mathematically corresponding to approximately $120 billion of cumulative revenue through fiscal 2033. That does not mean Google has committed to purchase $120 billion of Marvell products, because vesting depends on future purchasing activity and the programme structure gives the customer substantial flexibility.

The long-term investor-day forecast nevertheless suggests Marvell now expects several custom-chip programmes to scale enough for enormous revenue potential to move from theoretical architecture discussions into formal planning. The company is no longer presenting custom silicon as an adjacent experiment around its networking portfolio. It is describing the category as one of the central engines capable of reshaping the entire income statement over the second half of the decade.

That makes programme conversion the most important metric. Investors should watch design wins entering production, customer diversification and actual recognised custom revenue rather than extrapolating every maximum contractual threshold. Semiconductor history contains many promising programmes that were delayed, redesigned or abandoned before reaching the volumes originally imagined.

Why could interconnect revenue become almost as important as the chips themselves?

Artificial intelligence clusters are fundamentally distributed computing systems. A processor can operate extremely quickly, but an AI workload involving thousands or hundreds of thousands of accelerators requires enormous amounts of data to move between chips, racks, storage systems and networks without creating delays. The faster compute performance becomes, the more valuable high-speed optical and electrical connectivity becomes because bottlenecks migrate away from processors toward the infrastructure connecting them.

Marvell has spent years building exposure to this layer through optical digital signal processors, switching, networking and other connectivity products. Management now sees an interconnect opportunity measured in tens of billions of dollars, potentially providing a second growth engine alongside custom accelerators. This matters strategically because the interconnect business can benefit regardless of whether a data centre uses Nvidia processors, proprietary hyperscaler chips or a mixture of both.

That supplier neutrality can reduce some of the concentration inherent in custom silicon. If the overall number of accelerators continues increasing, the network connecting those devices also needs to grow, often at increasingly sophisticated speeds. Marvell consequently has exposure to a classic infrastructure multiplier in which each additional unit of compute creates surrounding demand for connectivity.

What would need to happen for Marvell to earn more than $30 a share by 2031?

Management’s long-range plan reportedly includes adjusted earnings above $30 per share and gross margins between approximately 56% and 59%. Reaching that level would require not only enormous revenue growth but substantial operating leverage because semiconductor research and development expenses need to expand more slowly than the resulting sales base. The company would also have to manage equity dilution, acquisitions and manufacturing costs while maintaining pricing power across both custom and connectivity portfolios.

Custom silicon economics can be attractive at scale because a platform may remain in production for several years after the major engineering work has been completed. However, hyperscalers are sophisticated buyers with enormous purchasing power and will negotiate aggressively over price. Marvell therefore needs the technical value of its design capabilities to remain high enough that revenue expansion does not come at the expense of margins.

Advanced manufacturing is another dependency. Marvell is fabless and therefore relies on foundry partners for leading-edge production, creating exposure to wafer availability, packaging capacity and process-node execution. Artificial intelligence customers will not tolerate significant delivery failures simply because demand is strong, particularly when multiple semiconductor suppliers are competing for future designs.

Why did Marvell shares rally while Broadcom also gained?

Marvell’s stock surge makes intuitive sense because management dramatically increased the scale of the future earnings opportunity. The rise in Broadcom shares is more interesting because the two companies compete in several custom silicon and connectivity markets, yet investors treated Marvell’s forecast as evidence that the addressable market may be expanding fast enough to support multiple winners. Artificial intelligence infrastructure spending can therefore create a situation in which a competitor’s strong guidance validates the entire category rather than automatically threatening every rival.

That interpretation may persist only while hyperscaler capital expenditure continues rising. If data-centre spending eventually slows, investors are likely to become much more sensitive to which supplier owns the strongest designs and highest-margin customer relationships. Marvell’s current valuation increasingly assumes the company will convert extraordinary AI infrastructure demand into durable programmes rather than simply participate in a temporary construction wave.

The most useful way to read the investor day is consequently not that $90 billion of fiscal 2031 revenue is destined to arrive. It is that Marvell believes the combination of custom silicon and interconnect has expanded its reachable market enough for a company that generated $8.2 billion last year to plan for an entirely different scale. The market’s positive reaction shows investors are willing to entertain that possibility, but the next five years will be measured in actual tape-outs, production ramps and revenue rather than presentation targets.