MRVL

Three-leg structural read:

- Custom AI ASIC - Marvell co-designs multi-generation AI accelerator silicon for hyperscalers outside NVDA. AWS Trainium successor + Microsoft Maia successor are the named programs; the unnamed third hyperscaler is the upside scenario. Revenue per program ramps over 3-4 generations.

- Coherent-optical DSP - Inphi acquisition gave Marvell the dominant share of DSPs inside every 400G/800G/1.6T optical pluggable. As AI scaleout drives DCI bandwidth, every additional pluggable shipped includes a Marvell DSP.

- DPU + networking SoCs - DPU is the smaller leg but pairs with the hyperscaler custom-silicon strategy. Networking SoCs are the legacy switch + Wi-Fi business - steady but not the growth driver.

Bear:

- Hyperscaler custom-silicon programs are inherently customer-concentrated. The top 3 customers are a majority of AI-silicon revenue. Re-architecture risk if a hyperscaler pulls a program in-house.

- Competitive set on coherent-optical DSP includes $AVGO's DSP line + the in-house designs at hyperscalers themselves long-term.

- Forward PE has expanded with the rally. Multiple compression risk if any of the three legs prints below cadence.