The Real Story Behind Marvell (MRVL)
Marvell (MRVL) has become one of the biggest winners from AI infrastructure spending. The story matters less for the chips themselves and more for moving data around massive AI clusters.
These clusters now run on racks of computers. The unit of compute used to be the chip. Now it is the whole rack of 72 chips. Getting racks to talk within themselves and to each other is the biggest limiting factor on AI growth, after power and infrastructure. Making chips talk to each other with the least heat and least energy is one of the gates on how fast AI can grow.
Marvell as a Bridge Between Vendors
Marvell (MRVL) lets chipsets from many vendors - Nvidia (NVDA), AMD (AMD), Amazon (AMZN) or others - all talk to each other. This means you do not have to buy your entire data center from one company for it to work. You can buy from many suppliers and use a company like Marvell to make those different technologies work together.
The Nvidia (NVDA)-AWS (AMZN) announcement is a good example. Everyone focused on the 2 million GPUs, which is a huge number worth focusing on. Buried in it was a lot of detail about Nvidia and AWS working together on interconnect - the communication inside the racks. That announcement is not bad for Marvell at all. It shows where the engineering effort is going. Marvell is placed to benefit from both the Nvidia stack and the growing open stack that serves as an alternative. Marvell can sit on both sides, and that should be a big part of its growth.
For Nvidia (NVDA) to hit its targets, everything around its compute must scale at the same rate. Anything that could slow AI growth is a concern. Some see Jensen Huang and Nvidia as both the kings and the king makers, given the giant funding around Arm (ARM) tied to Nvidia.
Optical Interconnect vs. Custom Silicon
Optical interconnect is the "steak"; custom silicon is the "sizzle." The key question for the earnings call is how management talks about optical interconnect, especially after the Nvidia (NVDA)-AWS announcement. This interconnect piece is the story for this quarter. Three quarters of Marvell's (MRVL) revenue comes from the data center, so interconnect will be the story for the next few quarters.
Custom silicon is a separate, longer-term conversation. Marvell (MRVL) bought companies like Celestial that helped it a lot in interconnect.
Google's TPU Split and the Opening for Marvell
Google (GOOGL) split the design of its eighth-generation TPUs. Part of the design went to Broadcom (AVGO), likely for the training TPUs, and part went to MediaTek. This raises a point: the custom ASIC business is so large that even a company the size of Broadcom and a customer like Google may decide to use several partners. That creates a sizable opening for a smaller player like Marvell (MRVL).
Marvell vs. Broadcom (AVGO)
Broadcom (AVGO) is the closest comparison, though it is a much bigger company. Marvell (MRVL) can look at Broadcom's success and the success of others doing custom ASIC design and decide that interconnect is a strong business and that future growth for it lies in those areas.
Marvell (MRVL) and Broadcom (AVGO) have become very different investments. Broadcom is in so many places; Marvell has a tighter book of business, especially on the custom ASIC side. That gives a clearer view of where its revenue and value come from. When the Nvidia (NVDA)-AWS deal appears and you remember Amazon (AMZN) is one of Marvell's biggest customers, you can trace what it means for the custom ASIC design business more directly than inside a larger company like Broadcom.
It is tempting to cast Broadcom (AVGO) as the giant and Marvell (MRVL) as the scrappy challenger. Broadcom is historically very nimble too, so this is not about nimbleness. On scale, Broadcom is the story. On a company branching out and pivoting into the custom ASIC business, you are looking at Marvell.
How Far Into the Future Can They See?
What Nvidia (NVDA) did differently last time was how far into the future it was willing to talk. That raises the question of whether forward visibility becomes the new measuring stick for the whole space.
With Nvidia (NVDA) products like Vera Rubin, the investments are enormous - hundreds of millions of dollars for those systems. So Nvidia mainly looks at its largest customers, the hyperscalers and others that can afford that level of investment. A company like Marvell (MRVL), which provides the inside architecture that helps technologies from different providers talk to each other, may not be able to forecast as far out. It can say directionally that the rising tide will lift its boat too. Holding other companies to Nvidia's far-out forecasting may not be the right measuring stick for the rest of the space.
