Wall Street is focused on Nvidia's (NVDA) graphics processing units (GPUs). The faster-growing story is sitting beside them. Nvidia has quietly built a networking operation that generated nearly $15 billion last quarter, up from roughly $3 billion per quarter two years ago. Wall Street expects that figure to approach $17 billion when the company reports earnings on Wednesday.
Compute still dwarfs networking inside Nvidia's Data Center business. But networking has been growing faster, up nearly 200% year over year in the last quarter compared with 77% for compute.
That pace is slowing. Analysts expect networking growth of roughly 134% in the quarter about to be reported, still ahead of compute at about 100%.
The change goes straight to one of the biggest questions hanging over Nvidia.
Everyone is designing AI chips these days. Alphabet (GOOG, GOOGL), Amazon (AMZN), Microsoft (MSFT), Meta Platforms (META), and other hyperscalers are spending billions developing or commissioning custom processors that can handle some workloads without an Nvidia GPU.
But those chips still have to talk to one another.
As AI data centers connect thousands of accelerators into giant computing systems, moving data among those processors becomes a challenge in its own right. That has pushed the AI trade deeper into networking, memory, storage, and other infrastructure.
Jessica Inskip, director of investor research at StockBrokers.com, described Nvidia's strategy on Yahoo Finance's Opening Bid on Tuesday as "investing in the bottlenecks that are the AI infrastructure."
Networking is one of them.
Nvidia increasingly sells the plumbing that keeps huge AI systems running. Its Spectrum-X products connect servers over high-speed Ethernet, while InfiniBand and NVLink help move enormous amounts of data between processors quickly enough to keep expensive chips from sitting idle.
More recently, Nvidia has pushed NVLink Fusion, which lets companies plug their own custom processors into Nvidia's broader system of high-speed connections, networking, power, and cooling. Think of it as designing an entire server rack to work like one giant computer, rather than selling the chips one by one.
That changes the competitive equation.
Nvidia's stock has already been acting differently from much of the chip industry.
Since the semiconductor trade peaked around June 22, Nvidia stock has gone essentially nowhere, while the iShares Semiconductor ETF (SOXX) has fallen roughly 20%. As Tuesday's Chart of the Day showed, Nvidia largely avoided the chip bear market that swallowed most of the group.