NVIDIA's Networking Business: How InfiniBand and Ethernet Became AI Infrastructure

NVIDIA's networking business runs above a $13B annual rate in 2025 — roughly 12% of data-center revenue and more than triple its ~$4B level in fiscal 2023.

That's the short answer. The longer answer is more interesting: NVIDIA didn't stumble into networking. It paid $6.9B for Mellanox in 2019 because it understood something most of the market still underrates — at AI scale, the wire between the GPUs matters as much as the GPUs.

NVIDIA networking, InfiniBand, and AI infrastructure: what it actually is

NVIDIA's networking business sells the high-speed interconnect that links GPUs into a single training cluster: InfiniBand switches and adapters, Spectrum-X Ethernet, ConnectX NICs, BlueField DPUs, and NVLink. It runs above $13B annualized in 2025, about 12% of data-center revenue, and came almost entirely from the $6.9B Mellanox acquisition. Its job is to move data between tens of thousands of GPUs fast enough that they behave like one machine rather than a warehouse of idle silicon.

Training a frontier model isn't one GPU working hard. It's 25,000 to 100,000 GPUs working in lockstep, exchanging gradients every few milliseconds. If the network introduces latency or drops packets, the whole cluster stalls — every GPU waits for the slowest link. At that scale, a $40,000 GPU sitting idle because of a congested switch is the most expensive thing in the building. Networking is what keeps utilization high, and utilization is the entire economics of an AI data center.

NVIDIA networking revenue: from a rounding error to $13B+

Networking was a sleepy line item until the AI buildout. The trajectory tells the whole story — this segment grew faster than almost anything else at NVIDIA, off a base that barely registered five years ago.

Figures are approximate, drawn from NVIDIA quarterly disclosures and management commentary; NVIDIA reports on a fiscal year ending in late January, so "FY2025" largely covers calendar 2024. You can track the underlying earnings on the Big Tech Earnings dashboard.

InfiniBand vs Ethernet for AI: the comparison that defines the business

The core tension in NVIDIA networking and AI infrastructure is InfiniBand versus Ethernet. InfiniBand is the purpose-built choice — lossless, low-latency, with in-network compute. Ethernet is the incumbent everyone already runs. NVIDIA sells both, and Spectrum-X is its bet that it can win the Ethernet half too.

For years the answer was simple: if you wanted the best training performance, you ran InfiniBand. That's why xAI, Microsoft, and Meta's largest clusters were built on it. But the world's data centers are overwhelmingly Ethernet, and hyperscalers hate single-vendor lock-in. That is precisely the opening the Ultra Ethernet Consortium — backed by Broadcom, AMD, Arista, and others — is targeting. NVIDIA's counter is Spectrum-X.

The $6.9B Mellanox bet that built NVIDIA networking

In March 2019 NVIDIA agreed to buy Mellanox for $6.9B, beating Intel and others in the auction. The deal closed in April 2020. At the time it looked like a steady infrastructure tuck-in — Mellanox did roughly $1.3B in revenue in 2019. Six years later that business is doing more than $13B annualized. It is, alongside the original CUDA bet, one of the most consequential capital-allocation decisions in semiconductor history.

Mellanox handed NVIDIA three things it could not have built fast enough on its own: InfiniBand, the dominant AI training interconnect; high-speed Ethernet silicon; and the engineering that became BlueField DPUs. Combined with NVLink — NVIDIA's own chip-to-chip and rack-scale fabric, which in the GB200 NVL72 links 72 GPUs into one giant accelerator — NVIDIA now owns every layer of connectivity from inside the rack to across the data center.

Why NVIDIA networking is becoming AI infrastructure, not an accessory

Networking typically runs 10-20% of the total cost of a large AI cluster, and that share is rising. As clusters scale from 25,000 to 100,000+ GPUs, the interconnect requirements grow faster than the compute — you need more switches, more cables, more optics, and more bandwidth per GPU just to keep utilization from collapsing. On a $5B cluster, networking can be $500M-$1B of spend. That is no longer an accessory; it is a strategic line item that buyers negotiate as hard as the GPUs.

This is also why NVIDIA increasingly sells systems, not chips. The GB200 NVL72 isn't a GPU you buy — it's a rack with 72 GPUs, 36 Grace CPUs, and NVLink fabric wired together, sold as one product. When NVIDIA prices a rack-scale system, the networking is baked in, the margin is captured, and the customer can't easily swap in a competitor's switch. That bundling is the quiet reason networking margins and attach rates have held up even as competitors crowd in. For the broader picture on how this stacks against AMD and Google's TPU networking, see our breakdown of the AI hardware wars.

The threats to NVIDIA's networking moat

The networking moat is real but not unassailable. Three pressures are worth watching:

Broadcom is the most credible challenger — its merchant Ethernet silicon already powers a huge share of hyperscaler networks, and its custom-ASIC business gives it a foothold inside the same clusters NVIDIA sells into. The likely outcome isn't that NVIDIA loses networking; it's that Ethernet grows faster than InfiniBand, and NVIDIA has to defend its share with Spectrum-X rather than collect an InfiniBand toll. That is exactly why management pivoted hard to Ethernet — they read the same map.