Nvidia quietly builds multibillion-dollar behemoth to rival chip business

Nvidia CEO Jensen Huang anticipated the market by over a decade, initiating work on AI-specific chips in 2010, long before the current AI boom. A parallel strategic move in 2020—bolstering data center networking through a key acquisition—has created one of Nvidia’s fastest-growing and most profitable divisions, though it has flown under the radar.

Within just a few years, Nvidia’s networking division, which connects data centers, has become the company’s second-largest revenue source, trailing only compute. According to Nvidia’s latest earnings report, this segment generated $11 billion in the last quarter, representing a 267% year-over-year increase, and contributed over $31 billion for the full year.

Fueled by AI processing demands, this division encompasses technologies such as NVLink, which facilitates GPU communication within data center racks; Nvidia InfiniBand Switches, an in-network computing platform; Spectrum-X, the Ethernet platform for AI networking; and co-packaged optics switches.

Collectively, Nvidia’s networking business provides the essential technology required to build an “AI factory,” a data center optimized for training AI models.

Kevin Cook, a senior equity strategist at Zacks Investment Research, described Nvidia’s networking segment as one of the company’s most remarkable new developments. “[Nvidia’s networking business] reported $11 billion for the quarter; that figure exceeds Cisco’s entire networking business and nearly matches full-year estimates,” Cook noted, adding that this single quarter’s performance matches what Cisco achieves in a year.

Despite its success, this segment receives less attention than Nvidia’s larger chip business. It also garners less publicity than the company’s gaming division, its original core business, which is nearly three times smaller.

Nvidia’s networking origins trace back to Mellanox, an Israeli networking firm founded in 1999 that Nvidia acquired in 2020 for $7 billion.

Kevin Deierling, Nvidia’s senior vice president of networking, joined the company via the Mellanox acquisition. Deierling admitted to TechCrunch that public unawareness of Nvidia’s networking capabilities might stem from inadequate marketing on his part, though he dislikes that explanation.

“People often view networking merely as connecting a printer,” Deierling explained. “Jensen stated on the first day of the acquisition that the data center is the new unit of computing. Networking involves far more than transferring small amounts of data between compute nodes; it serves as the foundational infrastructure.”

While Deierling initially did not understand Huang’s rationale for the acquisition, he now sees the logic. Maintaining a networking division alongside GPUs allows Nvidia to sell chips paired with the technology they perform best with.

“When Jensen acquired Mellanox in 2020, he recognized it as the missing piece to create a complete GPU package,” said Cook, the Zacks analyst.

Deierling emphasized that another factor in Nvidia’s networking success is its exclusive focus on selling full-stack solutions rather than individual components, and its reliance on partners for distribution rather than direct sales.

“I cannot think of other companies possessing [the] full-stack capabilities we offer,” Deierling stated. “We are distinct. We develop the entire compute stack, fully integrated, and then bring it to market through our partners.”

Nvidia recently unveiled numerous updates to its networking system during Huang’s keynote address on March 16 at the company’s annual GTC technology conference. The company introduced the Nvidia Rubin platform, featuring six new chips designed to power an “AI supercomputer.” Additionally, Nvidia announced a new Inference Context Memory Storage platform and more efficient Spectrum-X Ethernet Photonics switches, among other products.

“Networking is no longer a peripheral for connecting printers or slow I/O devices,” Deierling remarked. “It is fundamental to computing. In the past, we had what was called the backplane inside the computer. Today, the network serves as the backplane of the AI factory, making it critically important.”