In the brutal, high-stakes arena of custom AI silicon, rumors move faster than the chips themselves. Marvell Technology CEO Matt Murphy recently found himself in the crosshairs of a market narrative that suggested the company’s foundational relationships with Amazon and Microsoft were fracturing. The specific concern was that Marvell had been displaced in the design cycles for future iterations of Amazon’s Trainium processors and Microsoft’s Maia accelerators—the very hardware meant to break the industry's expensive dependence on Nvidia.
The reality is more nuanced than a simple win or loss. Marvell is not losing its seat at the table; it is experiencing the natural, albeit painful, evolution of the "hyperscale" business model. When cloud giants like Amazon and Microsoft reach a certain scale, they no longer rely on a single partner for an entire silicon generation. They diversify. They pit vendors against each other to drive down costs. They "multi-source" to ensure their data centers never go dark because of one company’s supply chain hiccup.
Marvell is currently navigating this shift from being an exclusive architect to a competitive specialist. While analysts at firms like Benchmark have sounded alarms about "lost" designs for Trainium 3 and 4, Murphy has countered with a blunt defense: the company has the bookings, the backlog, and the firm orders to prove its relevance.
The Amazon Diversification Trap
The friction began when reports surfaced suggesting that Amazon had tapped Alchip Technologies, a Taiwanese design and packaging firm, for upcoming versions of its Trainium chips. For years, Marvell was the primary hands-on partner for Amazon’s custom silicon efforts. Losing a lead role in Trainium 3 or 4 would represent a significant hit to Marvell's custom compute revenue, which the company has spent years pitching as its primary growth engine.
However, viewing this as a binary "Marvell vs. Alchip" contest ignores how Amazon Web Services (AWS) operates. AWS is a $100 billion run-rate business that treats silicon like a utility. If Marvell’s engineering fees or lead times don't match the internal requirements of the next project, Amazon moves. But moving "away" from Marvell for a specific logic design does not mean moving "out" of the Marvell ecosystem.
Marvell still controls the plumbing. Even if another firm handles the core logic of a custom AI accelerator, that chip still needs to talk to the rest of the data center. Marvell’s dominance in optical interconnects and Digital Signal Processors (DSPs) makes them nearly impossible to extract from the AWS infrastructure. You can change the brain of the chip, but you still need Marvell’s "nervous system" to move the data.
The Microsoft Broadcom Shadow
Parallel to the Amazon noise is the growing presence of Broadcom in Microsoft’s Azure ecosystem. Reports from late 2025 and early 2026 indicate Microsoft is in deep discussions with Broadcom to co-design future Maia AI accelerators. Because Broadcom is significantly larger and has a deeper IP portfolio in high-end networking, Wall Street viewed this as a direct threat to Marvell’s tenure as Microsoft’s custom silicon partner of choice.
Microsoft’s strategy here is transparent: they are building a hedge. The Maia 100 and 200 programs have been heavily linked to Marvell’s design services. By bringing Broadcom into the fold for Maia 300 or beyond, Microsoft ensures it isn't beholden to Marvell’s roadmap.
For Marvell, this creates a "squeeze" on margins. To keep Meta, Microsoft, or Amazon from jumping ship entirely, Marvell has reportedly resorted to waiving certain up-front Non-Recurring Engineering (NRE) fees. In high-end chip design, these fees can run into the tens of millions of dollars. Waiving them is a sign of a company fighting for territory in a market where the customers—the hyperscalers—now hold all the cards.
Why the Stock Market Overreacted
The 15% to 20% swings in Marvell’s stock price following these reports were driven by a fundamental misunderstanding of "XPU" revenue. When Marvell talks about XPUs—their catch-all term for custom accelerators—they are projecting a massive ramp-up in fiscal 2027 and 2028.
Investors panicked because they assumed that if Marvell isn't the sole provider for Trainium 4, the entire growth thesis collapses. They missed the fact that the total addressable market for custom silicon is expanding so rapidly that Marvell can own a smaller "slice" of a much larger pie and still see revenue growth.
- Custom Compute Volume: Marvell has already secured 3nm wafer capacity for production starting in 2026. This capacity isn't for "potential" customers; it's for specific programs already in the pipeline.
- The Interconnect Moat: As AI models get larger, the bottleneck isn't the chip’s speed; it’s the connection between chips. Marvell’s acquisition of Celestial AI and its focus on optical I/O gives them a technological lead that Broadcom or Alchip can’t easily replicate.
The Shift to 2nm and Beyond
The real test for Matt Murphy isn't defending the business he lost or didn't lose in 2025. It’s the transition to the 2nm process node. This is where the engineering complexity becomes so high that only two or three companies on the planet can actually execute the design.
Marvell is betting that as Amazon and Microsoft push the limits of what is physically possible in a data center, they will have to come back to the specialists. The "diversification" we see today is a luxury afforded by current-generation technology. When the next wall of physics is hit at the 2nm level, the list of partners capable of delivering a working chip on the first try—without expensive "re-spins"—gets very short.
The company is currently divesting its lower-margin segments, such as its automotive Ethernet business, to focus entirely on the data center and AI. This is a "burn the boats" strategy. It signals to the market that Marvell is either going to be the premier architect for the AI era or it’s going to be a commodity component provider.
Hard Numbers and Realistic Timelines
Marvell is forecasting custom AI revenue to hit $2.5 billion and eventually scale toward $5 billion. This isn't abstract optimism; it's based on the five-year supply agreement signed with Amazon in late 2024. While the "logic" design of a chip might shift to a competitor, the physical layers, the storage controllers, and the networking interfaces remain Marvell’s territory.
The "air pocket" in revenue that bears predicted for 2026 hasn't materialized. Instead, the company is seeing a transition where older designs like Trainium 2 continue to provide a floor while newer, multi-vendor designs begin their initial ramp.
The struggle for Marvell isn't a lack of demand. It's the fact that they are doing business with the most powerful companies in the history of the world—companies that have more cash than some nation-states and zero loyalty to their suppliers. In this environment, "losing business" is often just another way of saying "the customer is getting bigger."
Would you like me to analyze the specific performance benchmarks of the Marvell 1.6T DSP versus Broadcom’s latest interconnect offerings?
