Share with your CTO
Marvell Technology is positioning itself as the infrastructure layer beneath the GPU layer, designing custom AI chips (ASICs, or application-specific integrated circuits built for one job rather than general use) for hyperscalers including Amazon, Microsoft, Alphabet, and Meta, while also supplying the optical and networking components that connect massive GPU clusters. At its investor day, management projected $70 to $90 billion in annual revenue by fiscal 2031, up from $8.2 billion today. A reported acquisition of optical components maker Lumentum, with an $84 billion market cap, would extend that vertical integration play into photonics.
What this means for your business
The AI infrastructure market is splitting into two procurement realities. Hyperscalers are betting heavily on custom silicon tailored to their own model architectures, which means the GPU duopoly is losing its grip at the top of the market. If you’re running AI workloads at a scale where chip efficiency and interconnect bandwidth actually show up in your cost structure, the vendor landscape you plan around today looks materially different from the one you’ll face in 2028.
The Lumentum angle matters more than Wall Street’s initial 8% sell-off suggests. The real bottleneck in large-scale AI inference isn’t raw compute anymore, it’s moving data fast enough between processors without melting your power budget. Copper interconnects hit physical limits at the distances and bandwidths next-generation clusters require. Optical interconnects solve that problem, and a Marvell that owns both the custom ASIC design and the optical plumbing becomes a single-throat-to-choke supplier across multiple layers of the stack. That’s a structurally different vendor than one selling you chips alone.
The projection of $30 in adjusted EPS by fiscal 2031 comes from Marvell’s own management, and Motley Fool, writing for retail investors bullish on AI hardware names, has every reason to present the bull case cleanly. The tenfold revenue growth assumption embedded in that forecast requires hyperscalers to keep expanding custom silicon programs at a pace that isn’t guaranteed if AI capital expenditure cycles compress or model efficiency gains reduce compute demand. The number to watch isn’t the 2031 forecast, it’s Marvell’s reported design wins with named hyperscalers over the next four quarters. If those wins are accelerating, the thesis holds. If they stall, the $80 billion revenue target is a slide deck, not a roadmap.
Based on reporting from Not Nvidia. Not AMD. This Custom AI Chip Designer Could Outperform Both by 2030., originally published 2026-10-29 16:30:00.
