Some news in semis land today, Nvida is to acquire Cumulus. We received a few questions about the rationale for this deal, and thought we should break our hiatus to touch on this.
At first glance, the deal logic is a big question. Nvidia makes chips for gaming and for data centers (read AI/ML). Cumulus makes software for networking. The history of chip companies buying software companies is not great. Intel bought and then spun off McAfee and Wind River, with little to show for either deal. Nvidia did not disclose what it paid for Cumulus, which has reportedly raised $134 million in funding.
That being said, the more we dig into Cumulus and think through this deal, the more interesting it gets. To be honest, we are not 100% clear on what motivated this deal, but there turns out to be a fairly interesting story taking place here. Or to be precise, there are three possible explanations, depending on how we view the deal.
First, the semiconductor perspective. Today, there are basically six high-volume markets for leading edge chips. Of these, the market for ‘networking’ chips is not the biggest, but it is still fairly large (~$10 billion +/- depending on how you count it). These are chips that go into switches and routers and other electronics that connect everything. Importantly, a huge chunk of these chips end up inside data centers, connecting all those racks of servers.
Recall, Nvidia acquired Mellanox last year for close to $7 billion. We wrote about that here. Nvidia generates close to a third of its revenues from “Data Centers”. This is largely selling GPUs to the big cloud players for use in AI and neural network operations. There is a big and rapidly growing market for GPUs, probably the fastest growing segment of semis today. Nvidia can now sell Mellanox products alongside its GPUs, to the same customers.
The key to understanding the Networking semis market is Broadcom. In many segments of the market, Broadcom enjoys a near-monopoly position. Mellanox tended to do best on the leading edge of products (e.g. the fastest versions of Ethernet), but Broadcom has, by far, the leading market share and port count.
At this point, no company can replicate Broadcom’s offerings. There are too many sub-segments and niches and no easy acquisitions left. Mellanox was the last major player left, nor are there many private chip companies (looking at you Sand Hill Road). The result is that any company that wants to grow a networking business needs to cobble together what it can, and Cumulus is a close-enough adjacency.
This bleeds into the second deal rationale – the common sales footrpint. Increasingly, chip sales mean providing a complete solution that bundles chips, hardware and (sometimes) software.
This is Nvidia’s view of the deal:
Cloud data centers are evolving to an architecture that is accelerated, disaggregated and software-defined to meet the exponential growth in AI and high performance computing. To build these modern data centers, HPC and networking hardware and software must go hand in hand.
We imagine the pitch internally at Nvidia goes along the lines of “you sales people are already calling on the ten cloud scale customers out there, while you are selling them GPUs and Mellanox Ethernet cards, get the customer to sign up for a license to Cumulus software.”
Too often, this kind of rationale is designed by the Corporate Development team (ahem) and then left to the sales team to make into reality. But that same Nvidia blog post points out that Cumulus has been working with Mellanox for a long time already. So it is very possible that those two companies already have a common sales playbook. Bringing Cumulus in house makes joint sales that much easier.
There are very few people today who directly interact with chips. Even the most advanced cloud builders have very small teams (relative to their total work force) who pay much attention to the chips. Semiconductors only consume abut $0.20 of every dollar spent on a rack of servers. They are the most important decision to make, but the whole bundle is where the capex actually goes. As a result, every data center sale is some kind of bundle.
And this leads to our third deal rationale – software. If you make a product that gets sold in a bundle, but your product only costs 20% of the total, and the rest of that bundle is largely commoditized bent metal, you are going to need some other leverage to close deals.
Chip companies really want to move up the stack and sell that software. As noted, this has never really turned out well. Software is a radically different core competency. No chip company has ever really been able to develop robust software as a product. However, some chip companies are very good at building software ecosystems, providing tools for their customers to run on those chips. Intel is the best example, despite their failures in buying software products, it is pretty clear that the whole data center stack owes a huge debt to Intel’s support and investment in software. The best example of this is Linux, which exists in its dominant form today in large part because Intel engineers contributed countless hours to building it. This applies as well to all the less glamorous bits of compute – BioS, firmware, tests, tools, the lists goes on. Chip companies cannot build software products, but they can build software and through that differentiate their chips.
Nvidia is actually a textbook example of that. Part of the reason that they became so dominant in GPUs for AI rests in the investment they made in an obscure piece of code called Cuda. To oversimplify, Cuda is a software layer that can run across all Nvidia GPUs. This made it much easier to write AI software. Just as the modern data center would not exist without the support Intel provided Linux, AI would not exists in its current form without Nvidia’s investment in Cuda.
By acquiring Cumulus, Nvidia may be attempting to replicate that software enabling for networking. At the very least, this may make sales of Mellanox products easier, but if done right (and they get very lucky) they may be able to differentiate their products enough to pry open other parts of the networking stack and give them an opening into Broadcom’s treasure hoard.
We do not know how the deal played out inside Nvidia, nor which of the scenarios above carried the day. Nonetheless, the deal has a clear logic to it and could provide a nice boost to Nvidia, albeit one which we may not be able to observe from the outside.