NVIDIA Networking Business Explained: NVLink, Spectrum-X, InfiniBand and Why AI Factories Need More Than GPUs

Summary

Calling NVIDIA “a GPU company” is becoming less useful with every generation of AI infrastructure.

A modern AI factory can contain thousands—or eventually hundreds of thousands—of accelerators. At that scale, the GPUs are only valuable if they can move data and synchronize work fast enough to behave like one computing system.

That makes networking part of the product, not an accessory.

NVIDIA's fiscal 2026 Data Center networking revenue reached $31.38 billion, up from $12.99 billion a year earlier. The company attributed the 142% increase to NVLink compute fabric, Ethernet and InfiniBand growth.

The important question is no longer simply “How many GPUs did NVIDIA sell?”

It is increasingly:

How much of the entire AI factory can NVIDIA supply?

One GPU Is Not an AI Factory

A single accelerator can run AI workloads.

A frontier-scale model is different.

Thousands of GPUs need to exchange parameters, memory and intermediate results constantly. If networking becomes congested, expensive GPUs spend time waiting instead of computing.

That destroys utilization.

NVIDIA's current networking architecture addresses several different scales of that problem.

NVLink: Connecting GPUs at Scale-Up Speed

NVLink is NVIDIA's high-speed interconnect for tightly coupling GPUs and related processors.

Think of it as the mechanism that helps many accelerators inside a rack-scale system behave less like isolated chips and more like one large computing engine.

NVIDIA's fiscal Q1 2027 filing specifically cited demand for NVLink, InfiniBand and Spectrum-X Ethernet as contributors to Data Center growth.

InfiniBand: A Dedicated High-Performance Fabric

InfiniBand has long been important in high-performance computing.

In AI clusters, it is designed for extremely high throughput and low latency between compute nodes.

For customers building dedicated training infrastructure where maximum network performance is the priority, NVIDIA's Quantum InfiniBand family remains an important part of the stack.

Spectrum-X: NVIDIA's AI-Native Ethernet Bet

Many data centers prefer Ethernet because it is already deeply embedded in cloud infrastructure.

The problem is that conventional Ethernet was not originally designed for highly synchronized AI traffic.

NVIDIA developed Spectrum-X as an Ethernet platform optimized for large AI fabrics. It combines purpose-built switches with SuperNICs and congestion-management technology. NVIDIA says the platform can deliver higher effective AI network performance than off-the-shelf Ethernet configurations.

The strategic significance is obvious:

NVIDIA does not want customers to choose between “NVIDIA GPUs” and “someone else's network.”

It wants to sell an integrated system.

Spectrum-XGS: When One Data Center Is No Longer Enough

The newest AI deployments can become too large for one building.

Spectrum-XGS is designed to connect separate data centers so they can operate more like one distributed AI infrastructure system. NVIDIA says it uses topology-aware congestion control and other techniques to improve multi-site collective performance.

This creates another addressable market:

scale across campuses, not merely scale within one rack.

2026 Added Another Networking Layer: Scale-In

In August 2026, NVIDIA described Scale-In as a new networking domain built around BlueField-4 and DOCA.

The goal is to accelerate the infrastructure traffic surrounding AI compute—storage, security, application access and data movement—without consuming host CPU resources.

This gives NVIDIA an increasingly broad networking stack:

Scale Up → NVLink

Scale Out → InfiniBand / Spectrum-X

Scale Across → Spectrum-XGS

Scale In → BlueField / DOCA infrastructure

Why Networking Revenue Can Grow Faster Than GPU Revenue

As clusters get larger, the number and complexity of connections rises rapidly.

A 100,000-GPU system requires much more than 100 times the networking complexity of a small test cluster.

Customers need:

• switches;
• NICs and SuperNICs;
• interconnect fabrics;
• optical links;
• congestion control;
• management software.

That creates an opportunity for networking content per GPU to rise as systems become larger.

Spectrum-6 Shows Where the Architecture Is Going

The Vera Rubin generation adds Spectrum-6, NVIDIA's 102.4 Tb/s Ethernet switch system.

NVIDIA says Spectrum-6 is designed for gigascale AI factories and can connect very large GPU populations as part of the broader Spectrum-X architecture.

The latest Q2 FY2027 earnings release confirmed Spectrum-6 systems are arriving across gigascale AI-factory deployments.

This is a product-cycle development, but it is more useful to see it as evidence of a broader transition:

NVIDIA is selling systems, not isolated accelerators.

Why Photonics Is Becoming Important

At extreme scale, traditional pluggable optical transceivers consume significant power.

NVIDIA's Spectrum-X Ethernet Photonics approach integrates optics more closely with the switch.

The company says its new architecture improves power efficiency and network resilience while enabling very high bandwidth for future AI clusters.

As power becomes a binding constraint in AI data centers, every watt used by networking is a watt unavailable to computation.

Why Investors Should Watch Networking Separately

A growing networking business can change NVIDIA's economics in several ways.

First, it raises the amount of NVIDIA content in each AI factory.

Second, it increases switching costs because customers deploy more of NVIDIA's integrated architecture.

Third, it broadens competition. NVIDIA is no longer competing only with other GPU designers; it is also participating in markets historically associated with networking companies.

The Risk: Customers May Not Want One Vendor Everywhere

Vertical integration is powerful, but customers often prefer multi-vendor infrastructure.

Large cloud providers also design internal networking technologies and accelerators.

NVIDIA itself acknowledges intense competition across high-performance interconnects and notes that customers are developing their own specialized silicon.

The networking bull case therefore depends on more than technical performance.

NVIDIA has to convince customers that the productivity gain from an integrated stack outweighs the desire for vendor diversity.

What MEXC Has Already Covered — and What This Article Adds

MEXC's earlier NVIDIA earnings preview already noted rapid networking growth and the contribution from NVLink, Spectrum-X and InfiniBand.

That article was primarily an earnings setup.

The more durable investment question is different:

Can networking turn NVIDIA from a dominant accelerator supplier into the default architecture for the entire AI data center?

That is the thesis worth following across multiple product generations.

How This Connects to NVDAON

NVDAON does not represent NVIDIA's networking division separately.

Networking matters because it can change NVIDIA's:

• revenue;
• margins;
• competitive moat;
• customer spending per deployment.

Those factors influence NVDA, which then influences NVDAON.

For token mechanics, see What Is NVDAON?.

FAQ

Is NVIDIA only a GPU company?

No. Networking has become a large part of its Data Center platform.

How large was NVIDIA Data Center networking revenue in fiscal 2026?

$31.38 billion, up 142% year over year.

What is NVLink?

It is NVIDIA's high-speed scale-up interconnect for tightly coupling accelerators and processors.

What is Spectrum-X?

It is NVIDIA's AI-optimized Ethernet networking platform.

Does NVIDIA still use InfiniBand?

Yes. InfiniBand remains part of NVIDIA's scale-out networking portfolio.

Why does networking matter to NVDA stock?

It increases NVIDIA's potential revenue per AI deployment and strengthens its full-system strategy.

Risk Disclaimer

NVIDIA's networking growth depends on AI infrastructure spending, competitive performance, customer adoption and the continued expansion of large-scale computing. Product-performance claims from NVIDIA should not be interpreted as guaranteed commercial outcomes.