Nvidia has an underappreciated advantage as it looks to fend off rivals in the increasingly competitive chip market.
Best known as a semiconductor company, Nvidia now also claims to be “the largest networking company in the world,” according to CEO Jensen Huang’s comments on the latest earnings call. The company did $11 billion in networking revenue during the most recent quarter, up 263% from a year before and a major driver of the company’s sizable sales beat for the period.
While the networking category represents just a sliver of the overall $62.3 billion in data-center revenue that the company recently recorded, networking has seen tremendous growth in the past year because of how AI chips are changing, William Blair analyst Sebastien Naji told MarketWatch.
“It’s no longer about building the best, most-performant GPU, but it’s about how you connect all these GPUs into servers, into racks, into system-level clusters and computers,” Naji said, referring to the company’s graphics processing units. “That’s where the networking piece becomes so important.”
Nvidia’s networking business includes scale-up, scale-out and scale-across offerings, meaning the company addresses connectivity both inside and outside of the data center.
Introduced in 2016, Nvidia’s NVLink interconnect technology is used for scale-up, which involves connecting several GPUs within a server rack to work together enabling higher bandwidth and lower latency.
The company launched NVL72 with its Grace Blackwell architecture in 2024 that connects 72 Blackwell GPUs and 36 Grace central processing units to essentially act as one computer. Naji said NVLink started driving meaningful revenue with the advent of Blackwell, since the platform involved more chip clusters that needed to be connected.
Nvidia’s scale-up technology has prompted other companies in the networking space to shift attention to new opportunities for connectivity and introduce products that compete with what it has built, Naji said.
“There’s a lot more crosscurrent that has to happen between these GPUs as these systems get larger,” he said.
Thinking about Blackwell and Vera Rubin — as well as the company’s subsequent Rubin Ultra and Feynman AI platforms — Naji said it’s fair to say Nvidia’s networking segment will become more important as a revenue driver, since its chip racks will become more dense and need more connectivity.
That’s already visible in the transition from Blackwell to Rubin, he said, down to the simplest layer of switching that is needed as the number of chips doubles from one generation to the next.
Scale-out refers to connecting several racks of chips in a data center together to form a cluster of up to thousands of GPUs working together as one system, which is what InfiniBand and Spectrum-X Ethernet are used for. And now that training has expanded across chip clusters that can be multiple data-centers large, Nvidia’s Spectrum-XGS is used for scale-across networking that allows for data centers in different locations to work together as one supercomputer.
“The scale-up opportunity is the one that’s really new here, and has been introduced with Nvidia launching NVLink — that’s really what kind of opened up this market,” Naji said.
The company’s networking prowess can partly be attributed to its acquisition of Mellanox Technologies in 2020. At the time, Mellanox was a leading supplier of switches and other components based on InfiniBand, the networking communications standard for data centers running high-performance computing and AI workloads.
Nvidia is projected to do $57 billion in networking revenue this fiscal year and $70 billion next year, according to FactSet estimates.
“Looking back, what a key acquisition that was for $7 billion,” Kevin Cook, a senior stock strategist at Zacks Investment Research, told MarketWatch.
Nvidia bought the company because it realized Mellanox would not only give it the core tools for networking, but also the talent to build custom solutions, he said.
Part of the package was deep expertise on components such as SerDes, which is important for high-speed communication between chips, Naji said. Nvidia has been able to expand its networking revenue with its Spectrum-X Ethernet platform, Network Interface Cards and BlueField data processing units, he added.
All of these connectivity components are critical to the rack systems that Nvidia essentially sells as turnkey computers, according to Naji.
“A key part of why I think they have such a performance advantage over most of their competitors is this maniacal focus, probably earlier than most of the other chip companies, on networking, on connectivity as a core part of your computer,” Naji said.
And Nvidia’s networking segment is essential to Huang’s long-term vision of building AI factories, Cook said. The networking business will have to continue growing, Cook said, if Nvidia wants to position itself as the leading supplier in the data-center market that Wall Street predicts could be worth between $3 billion and $4 billion by the end of the decade.
NVLink looks like the winning technology so far, Naji said, but over the past couple of years, companies such as Astera Labs have innovated on other standards such as PCIe and UALink to do the same tasks, while companies including Broadcom and Arista Networks are looking to Ethernet, which has been dominant in scale-up technology.
In Naji’s view, Nvidia’s NVLink scale-up technology has been a sort of moat for the company, as it’s the fastest such technology on the market, and is only accessible through buying Nvidia’s full-rack systems. As long as Nvidia can keep advancing its NVLink technology, Naji expects this part of the business to be a major asset for the company going forward.
“One of Nvidia’s advantages is they have so many different connectivity elements within their portfolio,” Naji said.
But while Nvidia’s networking growth is impressive, it clearly hasn’t been enough to shift the somewhat gloomy recent attitude toward Nvidia’s stock. Networking, after all, makes up just a fraction of the company’s business, and investors have come to worry more about big-picture things, like whether overall AI spending is sustainable and whether financing deals between various AI players will come back to bite them all.
In Cook’s view, many investors think spending will peak in two years, which means Nvidia’s stock could peak now, “and the networking business isn’t going to change that.”
Comments