⚠ Customer Concentration and the Re-Compete TreadmillHigh threat

The central risk in Marvell is structural rather than operational. Its data-center revenue — roughly three-quarters of the company¹ — comes from a small number of hyperscalers, each of whom re-competes each generation of silicon, and none of whom is contractually obliged to continue. There is no installed base that renews, no software ecosystem, and no switching cost beyond the inconvenience of changing partners.

This is not hypothetical. The later generations of Amazon's Trainium were reported to have moved to Alchip² — a program that had been among Marvell's most prominent, lost to a cheaper design house once the architecture matured. The pattern is intrinsic to the model: the first generation of a custom chip requires the deepest engineering partnership, and each subsequent one requires less, which steadily converts a technical relationship into a price negotiation.

The mitigations are genuine. Marvell now has 18 cloud design wins and more than 50 active opportunities across over 10 customers³, and the Google agreement adds the one major hyperscaler it lacked. Breadth converts a catastrophic single loss into a manageable one, and the electro-optics business sells to many module makers rather than to one buyer.

What would falsify the bull case is simple to state: a second flagship program lost at a generational boundary, particularly if it went to a lower-cost house rather than to Broadcom. That would establish the Trainium outcome as the rule rather than the exception, and would mean Marvell is selling design labour into a commoditising market at a valuation that assumes it is selling irreplaceable technology. Watch program retention across node transitions above every other disclosure.