Why Marvell's Optical Chipsets Could Redefine AI Infrastructure Economics

Here’s the thing about AI infrastructure that nobody talks about enough: power consumption is eating into margins. Traditional copper cables in data centers? They’re basically obsolete for next-gen AI workloads.

Marvell Technology (MRVL) just figured out how to flip the script with co-packaged optics—optical interconnects baked directly into switching ASICs. The math is brutal: copper maxes out, but Marvell’s latest 800G and 1.6T optical modules are cutting power draw and latency simultaneously while pushing further distances than legacy setups.

The Tech That Matters

Marvell’s silicon photonics light engines can hit 6.4T speeds by consolidating multiple components into modular units. Think of it as stacking optical lanes instead of fighting electrical physics. They’re replacing power-hungry DSPs and retimers, which means hyperscalers get better efficiency-per-watt—the metric that actually moves needle.

Expect the optics segment to grow double digits sequentially in Q3 fiscal 2026. Translation: this isn’t niche anymore.

The Numbers Tell a Story

• Q2 FY2026 revenue: +57.6% YoY
• FY2026 consensus estimate: $8.11B (+41% YoY)
• FY2026/2027 EPS growth: 80% / 18% respectively

But here’s the catch: Marvell stock is down 24.5% YTD while the semiconductor industry climbed 35.8%. Trading at 7.91X forward P/S (vs industry avg 7.41X), it’s priced fairly but not cheap.

The Competition Gauntlet

Broadcom and Coherent aren’t sleeping. Both have solid pluggable optics and transceivers. Broadcom owns the PAM4 DSP space. Coherent’s 800G transceivers are flying off shelves as AI/ML workloads explode. Marvell’s edge? Co-packaged optics at scale—a tighter integration that competitors haven’t fully cracked yet.

Current verdict: Zacks Rank #3 (Hold). The thesis is real, but valuation premium vs upside visibility creates friction. Watch fiscal 2026 execution—if optics hit those double-digit sequential growth numbers, MRVL could break out of this chop.

The bottleneck in AI isn’t compute anymore. It’s the pipes. And Marvell just built better pipes.