New York, New York
— March 2, 2026

NVIDIA Deep Brief Part 3: Products and the Networking Breakout

NVIDIA’s Q4 made one thing crystal clear: networking is no longer a side story. It’s becoming a core growth engine.

The company disclosed that networking revenue surged in Q4 (up 267% year over year) and that networking reached $31 billion for the full year, roughly 15% of NVIDIA’s $215.9B annual revenue.

So what is NVIDIA’s networking business, and why is it suddenly exploding?

What “networking” actually means in the AI era

If GPUs are the engine, networking is everything that connects those engines together so they can operate like one giant machine.

At small scale, one server with a few GPUs can run AI workloads. At real scale, training and inference are happening across racks and entire clusters. That means massive amounts of data have to move fast and reliably, or else expensive GPUs sit idle.

That “connective tissue” includes:

• NICs (network interface cards): the on-ramps and off-ramps for data flowing in and out of servers which includes NVIDIA’s ConnectX family (e.g., ConnectX-7, -8, and -9).
• Switches: the traffic controllers that route data at high speed across the network fabric such as Quantum-2 and Quantum-X800 (InfiniBand) and Spectrum-6 (Ethernet).
• Cables and optical interconnects: the physical links that carry data between machines. Basically networking gear like DAC/AOC cables and optical transceivers.

The simple takeaway: networking turns “a bunch of servers” into an AI factory.

NVIDIA’s three pillars: scale up, scale out, scale across

NVIDIA frames networking in three layers:

1) NVLink (scale up)
Connects GPUs tightly within a server or rack so they can act like a single, larger accelerator.

2) InfiniBand (scale out)
Connects racks into one cluster. This is built for ultra-low latency and high throughput. It’s a big reason large training clusters can behave like one coordinated system.

3) Ethernet (scale across)
Connects at the data-center level. Ethernet is everywhere in enterprise networking, and NVIDIA is pushing hard here as AI moves from “tech-only” clusters into broader enterprise deployments.

Why this is such a big deal for NVIDIA’s growth

There are two reasons the networking numbers matter so much:

First, it signals cluster scaling.
When networking explodes, it usually means customers are building bigger systems, not just buying a few GPUs.

Second, it expands NVIDIA’s share of the total AI bill.
Instead of selling just the compute, NVIDIA sells more of the full system: compute + networking + software. That’s how you build a moat.

NVIDIA has even stated it’s now the world’s largest networking business, which is a stunning claim if you still think of the company as “just GPUs.”

NVLink Fusion: selling the “glue,” even in mixed systems

One of the more interesting product angles is NVIDIA’s push to make parts of its interconnect story more flexible.

The key concept: in some configurations, NVIDIA can still sell the networking fabric even if the compute isn’t fully NVIDIA. That matters in a world where hyperscalers use custom chips alongside NVIDIA GPUs. It’s another example of NVIDIA trying to be the default infrastructure layer for the AI factory.

Quick compute update: Rubin is on track

On the GPU roadmap, NVIDIA reiterated that Vera Rubin (the platform after Blackwell) is tracking for deployment in the second half of 2026, and reporting indicates early samples are already in customer hands.

The cost side: OpEx is climbing

All of this product expansion isn’t free. NVIDIA’s fiscal 2026 operating expenses, which includes costs like R&D, rose to about $23.1B, up 41% year over year.

That’s not inherently bad. If the returns stay strong, higher OpEx is the price of maintaining the lead. But it’s worth tracking because if revenue growth ever plateaus, cost discipline becomes a much bigger storyline.

Bottom line: Networking is emerging as a breakout catalyst, Rubin is progressing, and NVIDIA is clearly building more than chips. They’re building the full AI factory.