Cisco Wants to Make AI Chips, Too -- Are Broadcom and Marvell in Trouble?

It's been a transformational year for old networking technology giant Cisco Systems (CSCO -0.50%). An investor favorite of the 1990s and early 2000s, the internet and networking infrastructure provider enjoyed a resurgence as of late from its bets on the cloud and recurring revenue services, much like its old dot-com-era tech titan peer Oracle. For its latest trick, Cisco says it has a new family of semiconductors, the G200 and G202, used for AI applications.  

This specific domain of the semiconductor industry that Cisco entered is historically the turf of Broadcom (AVGO 3.84%) and Marvell Technology Group (MRVL 3.17%), which also announced expectations for rapid growth from AI chips in recent months. Both chip stocks ran higher on the news of renewed investor optimism.

Is Cisco's latest effort about to rain on this parade? 

Cisco wants a slice of a large and growing pie

Cisco built its empire with networking infrastructure, the equipment needed to build the internet itself, as well as organizations' own networks of connected computers and devices. Chief among this equipment are switches and routers, two types of computing units (often housed in data centers) that manage the large flow of data between devices and that power software services.  

Back in 2019, Cisco went in a new direction when it introduced Silicon One, an in-house semiconductor design segment focused primarily on key chips used in Cisco's switches and routers. The newly announced G200 and G202 products fall in this networking chip camp.  

Why are they important? AI isn't new, but recent breakthroughs in intelligent systems (like ChatGPT) have cloud computing giants scrambling to implement next-gen infrastructure. The superstars here are Nvidia (NVDA 6.18%) GPUs, which compute massive amounts of data to train these new AI systems. But computing data is only one part of an AI system. Networking chips are needed to string together thousands of these GPUs to create a massive unified supercomputer to train an AI system quickly enough that it makes economic sense. 

And that's where the Cisco G200 and G202 come in. Much like the AI networking chips Broadcom and Marvell have also introduced, Cisco's new flagship silicon can help string together up to 32,000 GPUs to tackle the largest and most complex AI training needs. While not dropping any specific names, it said five of the top six public cloud giants are testing these new Cisco Silicon One semiconductors.  

Just how big will this AI semiconductor pie get? It looks like it will be a very large market. Nvidia reported its data center segment will roughly double in just a single quarter, going from $4.1 billion in the first quarter of this year to perhaps as much as $8 billion in the second quarter. Similarly, Broadcom and Marvell both said their AI-specific sales are poised to double this year and again next year, likely thanks to the Nvidia GPU boom as all those Nvidia components needed to be hooked up to each other.

Will AI move the needle for Cisco?

The advent of new AI systems could certainly help Cisco continue its lumbering advance. The company could have lots of new switches to sell as those big cloud customers make new hardware purchases. 

Cisco's Silicon One isn't going to become some giant in the semiconductor industry anytime soon, though. More than anything, this is a way for Cisco to vertically integrate the design and distribution of its networking hardware. This could help Cisco differentiate itself from fast-growing peers in this market like Arista Networks, which also makes network switches but in an open architecture format -- meaning, in part, its customers can pick and choose which chips they want to use in their data centers, as well as what software will run on them.  

These days, a transition to cloud-based subscription revenue is what's really moving the needle for Cisco. Relying on lumpy sales growth from the introduction of new data center and networking hardware is out, and steady and recurring subscription sales are in. Cisco made numerous acquisitions to bolster its capabilities on this front, including building a cybersecurity unit that pairs with its network infrastructure. Don't expect Silicon One to be this grand disruptor of the semiconductor industry status quo.  

Should Broadcom and Marvell shareholders worry?

Ultimately, some competition from Cisco's upstart network chip design unit is a good thing for Broadcom and Marvell. Competition pushes leaders in an industry to keep innovating -- which is good for the business, and certainly good for customers. 

But bear in mind Cisco isn't the only company entering the network chip fray. Even Nvidia started its own segment here with the introduction of its Bluefield DPUs a couple of years back. Nvidia says it's also making progress in introducing these AI network designs to big customers. And Advanced Micro Devices, listed as a Cisco partner on some of its network switch products, has also been introducing some new chips of its own to try and take a swipe at this emerging AI race.

For now, Broadcom and Marvell are the clear-cut leaders in network chips and provided the most specific guidance for AI network chip sales growth -- again, doubling in sales this year compared to last, and another double next year.

However, all the competition that emerged for Broadcom and Marvell illustrates investors should curb their enthusiasm and be careful about blindly buying at any price. After the recent run-up in share price for these two semiconductor companies, it looks like at least some of the coming AI boom has already been priced in for both. Behemoth Broadcom currently trades for about 20 times trailing-12-month free cash flow, and smaller Marvell (which has lots of work ahead to improve its profit margin) trades for 48 times trailing-12-month free cash flow.  

Broadcom and Marvell are the two chip stocks to stick with for the long haul. But some caution is in order. With share prices jumping dramatically as of late and new competition (including from Cisco) now in play, investors that want to own these two top chip stocks might be best off utilizing a dollar-cost average plan to build a position over time at this juncture.