✦ The Future BetsNarrow moat

Marvell's future bets are unusually concrete for a company at this multiple: they are named programs with named customers and dated ramps, rather than markets that might exist. That is the nature of custom silicon — the revenue of fiscal 2028 is being designed now.

The most consequential arrived in August 2026, when Marvell disclosed a commercial agreement with Google to design custom semiconductors attaching to its TPU ecosystem — inference accelerators, storage controllers and network interface controllers¹. Google had been Broadcom's flagship account for seven TPU generations, so a Marvell engagement there completes a sweep of the major hyperscalers and validates the second-source logic at the most demanding customer in the market.

The second bet is Microsoft's Maia line, where management has guided to a substantial step-up in custom accelerator revenue in fiscal 2028 tied to the next-generation part². The third and fourth are technical rather than commercial: the move to 1.6-terabit optics, where each speed grade thins the field of companies able to ship a working signal processor, and the emerging scale-up interconnect inside the rack, where the industry is standardising on Ethernet and Marvell's serial-interface IP applies directly.

Grade all four against a single number: custom silicon revenue against its roughly $1.5 billion run rate. The programs are real, the ramps are dated, and the risk is not that they fail to arrive but that they arrive alongside losses elsewhere. This is a company whose bookings and cancellations happen in the same conversation with the same handful of buyers.