I still catch myself introducing Nvidia as a GPU company, then watching the financials correct me. TechCrunch on March 18, 2026 put a hard number under that correction. Networking, the business Nvidia absorbed with the roughly $7 billion Mellanox acquisition in 2020, is now its second-largest revenue line after compute. Last quarter brought about $11 billion in networking revenue, up 267 percent year over year, with more than $31 billion across the full fiscal year. That is not a side quest. That is a second Nvidia growing inside the first.
From Mellanox to the AI factory's connective tissue
When Nvidia bought Mellanox in 2020 for about $7 billion, a lot of outside commentary treated it as a smart but specialized bet on InfiniBand and high-performance interconnects. In 2026 the bet looks like foresight about AI factories. Training and inference clusters are only as useful as the fabric that keeps GPUs fed and synchronized. Nvidia spent years turning that fabric into a product family with its own gravity.
The product surface cited in the coverage is broad on purpose: NVLink, InfiniBand, Spectrum-X Ethernet, and co-packaged optics. That list spans scale-up links inside a node or rack story, classic HPC-style fabrics, Ethernet aimed at AI cloud operators who want a familiar networking religion, and optical packaging that chases density and power at the faceplate. It is a portfolio built to answer "how do the accelerators talk" at every layer buyers argue about.
Second-largest revenue after compute is the organizational tell. Networking is no longer an attachment sale that shows up when a GPU deal closes. It is a primary line. In cloud terms, Nvidia is selling the roads as aggressively as the engines. Anyone still modeling the company as only a semiconductor cycle on discrete GPUs is missing half of the invoice.
I have interviewed enough infrastructure leads to know why this lands. You can win a GPU allocation and still miss training efficiency if the network is the bottleneck. Vendors that own both sides of that problem can package reference architectures that feel inevitable. Inevitability is a pricing strategy as much as an engineering one.
The year-over-year jump of 267 percent on a quarterly networking figure around $11 billion also explains the tone of March coverage. This is not steady compounding from a sleepy adapter business. It is AI buildout demand showing up in cables, switches, NICs, and the software-defined stories that make those boxes look mandatory in a tender. When the interconnect line grows that fast, the compute story stops being the only story worth budgeting for.
GTC made the stack visible, not just the GPUs
Nvidia's GTC on March 16 sat two days before the TechCrunch piece, and the messaging fit the financials. The Rubin platform, Spectrum-X Ethernet Photonics, and Inference Context Memory Storage were part of the public map. I am staying inside those named pillars rather than inventing SKU tables. The point of the keynote cluster was that Nvidia wants buyers to see a full AI factory bill of materials, not a single accelerator SKU.
Spectrum-X Ethernet Photonics is especially on-brand for the networking surge. Ethernet is where a huge fraction of cloud operators are emotionally and operationally comfortable. Photonics is how you keep pushing bandwidth and reach without drowning in power and faceplate chaos. Pairing those words is Nvidia telling Ethernet shops they do not have to become InfiniBand shops to stay in the AI lane.
Inference Context Memory Storage belongs to the same glue narrative. As inference systems hang on to longer context and more concurrent sessions, memory and storage paths next to the compute fabric become part of networking's extended family. Nvidia is not only shipping switch silicon. It is trying to define the data pathways that keep model serving from stalling.
Analysts noted that a single Nvidia networking quarter can rival Cisco's annual networking revenue. That comparison is meant to startle, and it does. It also clarifies competitive geography. Nvidia is not only fighting other GPU makers. It is occupying wallet share that used to look like traditional networking and systems vendors' home turf.
I watched the industry underestimate this for years because networking lacked the consumer brand heat of GeForce and the meme heat of AI chips. Quiet compounding is still compounding. Mellanox was the seed. AI cluster scale was the fertilizer. GTC is the part where Nvidia stops being quiet about the full-stack glue.
For cloud builders, the strategic bind is familiar. Using Nvidia end to end can simplify integration and support. It can also deepen dependency. The networking numbers suggest many customers are choosing simplification under deadline pressure. When clusters are measured in weeks to bring-up, the vendor that sells compute plus fabric plus reference designs wins meetings.
What a multibillion networking line changes
The first change is narrative. Nvidia as GPU company was always incomplete. Nvidia as AI infrastructure company is closer, and networking revenue above $31 billion in a fiscal year makes the label harder to dodge. Capital markets already priced some of this in. Procurement teams are still catching up whenever they separate "chip budget" from "network budget" and discover the same vendor logo on both.
The second change is competitive response time. Rivals can field accelerators and still lose deals if they cannot match interconnect roadmaps, optics stories, and Ethernet AI pitches. That raises the bar for AMD, custom silicon shops, and traditional networking companies alike. You do not have to beat Nvidia on every GPU benchmark if you can break the assumption that the fabric must be Nvidia too. Doing that in 2026 is harder than it was in 2020.
The third change is architectural fashion. NVLink-heavy designs, Spectrum-X Ethernet deployments, InfiniBand pockets in research and vertical HPC, and co-packaged optics experiments will keep defining what a serious AI datacenter is supposed to look like. Fashion becomes standard when the market leader's revenue says the world is already buying it.
I am writing this the day after the TechCrunch report, with GTC still in the rearview. The durable takeaway is not a single quarter's $11 billion. It is that Nvidia spent six years since Mellanox turning connectivity into a growth engine that can stand next to compute without apology. AI factories need glue. Nvidia is selling the glue by the billion.
If you only watch chip launches, you will keep missing where the money moved. The networking business is no longer the quiet cousin at the family reunion. It is the relative who bought the house next door and started charging rent for the driveway. Cloud buyers already know this from the purchase order. The rest of the industry is catching up to an $11 billion quarter that makes the interconnect stack impossible to treat as an accessory ever again.
