The Custom Silicon Boom: How MRVL's Data Center Micro Scales Are Reshaping AI Infrastructure

Marvell Technology is quietly becoming the go-to name for hyperscale AI operations. Here’s what’s really happening beneath the surface.

By The Numbers: MRVL’s Data Center Explosion

In Q3 FY2026, Marvell’s data center segment hit $1.52 billion in revenue—a 37.8% year-over-year jump. The real story? Custom silicon products are driving this growth. The company projects its specialized chip business will account for roughly 25% of total data center revenue next year, with projected growth of at least 20%.

That’s not just incremental improvement. We’re talking about a segment that’s fundamentally reshaping how AI infrastructure gets built.

Design Wins Paint a Clear Picture of Dominance

Marvell just announced 18 new XPU and XPU-attach socket design wins. Most critically—many are already ramping into volume production. These aren’t vaporware; they’re real silicon hitting real data centers.

The product portfolio spans:

• Full custom XPU programs
• XPU-attached chiplets
• Multi-die XPU packages with integrated electrical I/O chiplets
• Supporting silicon components

Collectively, these wins represent a pipeline exceeding $7.5 billion in lifetime revenue potential—a 10%+ contribution to the addressable market.

The Secret Sauce: Advanced Process Technology

Marvell’s custom silicon strategy has evolved from niche ASIC work into something far more strategic: multi-year, deeply embedded partnerships with hyperscalers. The company leverages advanced 5nm and 3nm manufacturing processes, including proprietary SerDes technology like 112G XSR, Long Reach SerDes, and PCIe Gen 6 configurations.

This technical depth keeps Marvell ahead in the competitive race for AI silicon customization.

The Competitive Landscape: Who’s Playing and Who’s Winning

Broadcom remains a heavyweight in custom data center silicon, with its 3.5D XDSiP packaging optimized specifically for AI XPU performance and efficiency.

AMD competes through semi-custom SoCs and Instinct Accelerators deployed across numerous data centers, plus its Alveo Adaptable Accelerator Cards for compute-intensive workloads.

Yet Marvell holds a critical advantage: NVIDIA’s backing. That partnership provides the ecosystem support and customer validation necessary to maintain competitive positioning in this rapidly consolidating space.

The Valuation Question

MRVL trades at a forward price-to-sales ratio of 7.31X versus the semiconductor industry average of 8.46X—meaningful valuation compression despite superior growth prospects.

Over the past six months, Marvell shares gained 13.1%, underperforming the Electronics-Semiconductors industry’s 22.3% advance. That gap suggests either a value opportunity or a signal worth heeding—depending on your view.

What the Estimates Say

Consensus earnings forecasts point to 81% growth in FY2026 and 26% growth in FY2027. Recent analyst revisions have trended upward across both years, signaling confidence in the custom silicon trajectory.

The Zacks consensus currently rates MRVL as a #2 Rank (Buy).

The Bottom Line

Marvell Technology’s custom silicon business micro scales across an expanding customer base precisely when hyperscalers need differentiated, cost-optimized chips for AI workloads. With 18 new design wins, advanced process capability, and institutional backing from NVIDIA, the company’s positioned to capture outsized growth in data center infrastructure.

Whether valuations justify entry at current levels remains an individual decision—but the fundamental narrative around MRVL’s competitive positioning continues to strengthen.