Marvell raises 2028 revenue outlook to US$20 billion as AI connectivity expands

Marvell's 2028 revenue anchor at $20B and 2031 target of $70-90B is not optimism—it is a repricing of the "connectivity tax" in AI. As clusters scale from 10K to 100K GPUs, interconnect cost will outstrip compute itself. Marvell is betting on that structural shift. The technical cascade centers on optical interconnect. The 800G-to-1.6T transition window hits in 2026, CPO moves from lab to fab, and Marvell's silicon photonics IP combined with custom ASIC capability positions it as the critical node in hyperscalers' "de-NVIDIA" network stack. EML laser and silicon photonics wafer capacity will be the hardest bottleneck over the next 18 months. On risk, advanced-node custom silicon is tightly coupled to foundry capacity in Taiwan, China, while Chinese vendors hold over 60% of the optical transceiver supply chain. Any geopolitical escalation hits both cost structure and delivery certainty simultaneously. Competitively, Broadcom attacks on both custom ASIC and switching, while NVIDIA internalizes connectivity revenue via Spectrum-X. Marvell's moat is holding compute-customization, switching, and optical interconnect simultaneously—a combination Broadcom has not fully closed. The 24-month make-or-break: can Marvell lock in 2026-2028 custom silicon long-term agreements with at least two hyperscalers by end of 2025? If yes, $20B is the floor. If no, Broadcom resets the pricing power in AI connectivity.