美研芒格君 · 2026-10-06 · 帖子

Congrats to Marvell!! If you want to understand what the AI data center looks like in 2030, drop everything now and watch Marvell's Investor Day. Best roadmap I've seen for what comes after the GPU bottleneck: → Compute is no longer the constraint. Connectivity is. 
→ Optics keeps marching inward: scale-across → scale-out → scale-up → "scale-in," light between dies inside the package 
→ Every XPU drags 1–2 custom "attach" chips with it: NICs, memory expansion, storage 
→ $400B TAM by 2030 The tell: in their own FY31 model, interconnect is bigger than custom compute. They're pitching the nervous system of the AI factory, and they want to own every inch of it. $MRVL