Nvidia (NVDA) reported blowout results after the bell on Wednesday, beating analysts’ fourth quarter expectations on the top and bottom lines and issuing better-than-anticipated guidance for Q1.

Wall Street wasn’t exactly excited about the announcement — Nvidia stock fell more than 4% in early trading Thursday — but that could be tied to anxiety related to the broader AI trade.

Regardless, Nvidia’s data center business drove the vast majority of its $68.1 billion in revenue, accounting for $62.3 billion in sales. Dig deeper into the details, though, and there’s another nugget worth exploring: Nvidia’s networking business.

While the bulk of data center revenue came from GPU and CPU sales, which Nvidia refers to as its compute segment, $11 billion came from the company’s networking group, up 263% year over year.

Nvidia’s networking business has become an unsung hero within its data center business over the years. And despite its relatively small size compared to the company’s computer arm, it’s gaining more attention.

“We’re … now the largest networking company in the world,” CEO Jensen Haung said during Nvidia’s earnings call.

The AI giant’s networking business can be broken down into three main categories: scale up, scale out, and scale across.

Scale up has to do with connecting multiple server blades together within a large server rack. Scale out connects multiple server racks and systems into a large cluster, allowing companies to run a data center full of servers as a single high-powered computer. Scale across refers to using networking capabilities to connect entire data centers to each other to create a massive AI computer spread across different buildings.

To do that, the company takes advantage of its NVLink, InfiniBand, and Spectrum-X technologies. NVLink is the networking that connects multiple chips and server blades in a rack, while InfiniBand and Spectrum-X connect clusters or racks, including CPUs, compute nodes, and networking. Spectrum-XGS connects multiple data centers.

And while vendors that build AI servers and racks with Nvidia’s chips are more likely to choose the company’s networking offerings, they’re not the only ones.

In December, the firm announced that Amazon’s AWS would pair its custom Trainium chips with Nvidia’s NVLink Fusion, connecting Amazon’s processors into racks that act as individual computers.

“Momentum is strong with our Spectrum-X ethernet scale-up and scale-across networking as customers work to unify distributed data centers into integrated, giga-scale AI factories,” Huang said.