AINEWS
Arista, the cloud networking hardware company, just put a $3.5B price tag on the switches that wire GPUs together inside the data centers of the largest cloud operators.
Arista Networks spent most of its life as a footnote in cloud data centers: the switch vendor that beat Cisco on speed and price, mostly invisible to anyone outside networking. Its fiscal 2026 numbers, released earlier this quarter, make the company hard to miss.
Revenue in the first quarter reached $2.71 billion, up 35.1% year over year. Diluted earnings per share came in at $0.87, up 31.8%. Full-year revenue guidance climbed to $11.5 billion, implying roughly 27.7% annual growth. The number doing the most structural work, though, is the AI line: Arista now projects $3.5 billion in AI-related revenue for fiscal 2026, more than double the prior year. It is the first company-sized disclosure of what the networking tier of the AI buildout is actually worth.
A switch is the rack-mounted hardware that moves data between thousands of GPUs inside a single data center. As training and inference clusters have grown from dozens of accelerators to tens of thousands, the bandwidth between them has become its own bottleneck. The dominant answer inside hyperscalers, the largest cloud operators, has been to standardize on faster Ethernet fabrics rather than the InfiniBand alternative that NVIDIA promotes for its own GPU systems. Arista, which built its business on high-speed Ethernet switching for cloud customers, sits inside most of those builds. The $3.5 billion AI figure is the visible residue.
A piece republished by Yahoo Finance characterized the same numbers using a proprietary inflow signal as a framing device. The underlying fact is structural: AI capex has a stack, and the networking layer is now a tracked, multi-billion-dollar line on a public company's income statement, addressable on its own rather than buried inside a GPU or server line.
Three risks come attached, and the next earnings print on Aug. 4 will test each one.
The $3.5 billion AI line is a hyperscaler line: a small group of cloud operators, primarily Meta, accounts for the bulk of those orders. Arista discloses the top-customer share but does not name the buyers. If any one of them slows an AI build, the networking order book moves with it.
Networking hardware demand is downstream of hyperscaler AI capex, not upstream of it. Pullbacks at the GPU layer, whether from chip shortages, model-training saturation, or revenue disappointment at the model labs, arrive at Arista a quarter or two later in the form of softer orders.
The structural story is real but already partly priced in at current levels. Aug. 4 is the next date the market gets to check whether the AI mix is still expanding and whether full-year guidance holds.
The reader takeaway is bigger than Arista. For two years the AI capex narrative has run on GPUs and the small set of model labs that buy them. The networking tier was a fraction of its current size a few years ago. It is now sized to exceed $3.5 billion in a single fiscal year. If that line keeps doubling, the next stage of the buildout is a discrete, addressable market in its own right. If it does not, the GPU story is still the story, and Arista rides the same wave as the labs it sells to.
The Aug. 4 report will not answer the question. It will move the reader one data point closer to knowing whether the networking layer of AI is a category or a quarterly spike.