Earlier this week we attended a Marvell Industry Analyst event where they reviewed their businesses and laid out a new framework for how to think of the company. For all the talent at the company, one of Marvell’s long-standing challenges has been how to define it. Boil it all down, the company has now gone all-in on supplying infrastructure to the hyperscalers. According to the company’s latest data, they now generate over 70% of revenue from sales to the large Internet companies who run most of the world’s data centers. For better or worse.

Part of the company’s messaging problem stems from the fact that they produce products across such a wide range of areas – processors, interconnect, memory controllers, optical, storage and probably a few more. On the one hand, this makes them hard to categorize and describe. On the other, it demonstrates the wealth of technical talent they have built. Probably most interesting this week was their announcement of a custom stack for HBM, the memory critical in so much of AI compute today. This is important and also hard to do.

Another big element of Marvell’s story of late is the growing importance of its custom silicon business. Marvell provides critical support in helping the hyperscalers design all their internal silicon. Most prominently, they recently announced a five-year agreement with Amazon. This goes a long way to assuaging our concerns about the project-by-project nature of custom silicon work. We still have a lot of questions about this business, but it is now fairly clear that Marvell is competing with Broadcom for some fairly high value-add work, clearly differentiated from lower level design services available from companies like GUC and Alchip in Taiwan. Marvell has much more capabilities than those companies, and more importantly has areas where they can truly challenge giant Broadcom.

That all being said, we are still a bit uneasy with Marvell’s reliance on a handful of big customers. The semis industry is littered with the wreckage of companies that depended on giant, concentrated companies. Admittedly, Marvell is far larger than most of the recent examples we could name, but the risk remains. Selling into the hyeprscalers is a difficult business. They have immense resources, and the negotiating leverage that comes with that. They are demanding in terms of support requirements, contract terms, and IP protections. And like all giant customers deep down they give very little thought to the viability of their suppliers. The hyperscalers give, and sometimes, often, they take away.

Marvell is probably fairly insulated from this. The breath of their offerings makes them important enough that the hyperscalers cannot easily replace them (except with Broadcom, sometimes). But this level of customer concentration is an issue for any company. So while we really like what we see technically coming from the company, we worry that there is a lot of volatility ahead for them.