Arista posts $3.04 billion quarter as AI networks scale up, out and across data centres
The three axes of AI networking explain why the links between GPUs, racks and regions are now design decisions rather than plumbing.
Arista Networks reported second-quarter 2026 results that The Next Platform, in a report published on 10 August, framed around three directions in which AI networks now grow. Scale-up networks tie together the memory of GPUs and other accelerators inside a node or rack-scale system. Scale-out networks loosely couple many of those nodes or racks so they can share a job. Scale-across networks link entire data centres and regions.
For the quarter, Arista's product revenue reached $2.61 billion, up 38.8 percent, and services revenue reached $430.5 million, up 31.3 percent, for total revenue of $3.04 billion, a rise of 37.7 percent. Operating income was $1.38 billion, up 39.7 percent, and net income was $1.21 billion, up 36.5 percent. Cash and equivalents stood at $13.34 billion, up 50.9 percent.
Two forward-looking figures drew attention. Deferred revenue, money already collected for products and services not yet delivered, rose 69 percent to $6.87 billion. Customer purchase commitments climbed to $9.7 billion, roughly 2.7 times the prior level. On the strength of that pipeline, the company lifted its full-year 2026 revenue guidance by $1.1 billion to $12.6 billion, which would represent 44 percent growth.
Chief executive Jayshree Ullal put the company's AI networking revenue forecast for 2026 at $3.6 billion, about 2.1 times the prior year. She said scale-across networking would account for roughly 30 percent of that AI revenue, about $1.08 billion, with scale-up and scale-out together making up the remaining 70 percent, about $2.52 billion. On the scale-up tier, where Nvidia's NVLink dominates today, Ullal said it would take time for customers to move from a proprietary interconnect to the alternatives.
Why it matters
Networking used to be a rounding error in a compute budget; these numbers show it has become a line item that moves with accelerator spending. The three-tier framing is useful because each tier has different latency, bandwidth and reliability demands, and the products serving them are diverging. Scale-across growing to nearly a third of AI revenue signals that power and site limits are pushing training and inference across multiple buildings and regions. For anyone learning infrastructure, knowing where a packet crosses a tier boundary is now as important as knowing the GPU it feeds.
Networking used to be a rounding error in a compute budget; these numbers show it has become a line item that moves with accelerator spending.
Scale-across: site to site
Links whole data centres and regions; latency set by distance; built for bulk bandwidth.
Scale-out: rack to rack
Loosely couples many nodes or rack-scale systems so they can share one job; microsecond latency.
Scale-up: accelerator to accelerator
Ties GPU and XPU memory together inside a node or rack; today dominated by proprietary links such as NVLink.
What you can learn from this
- Scale-up versus scale-out. A scale-up fabric connects accelerators so closely that one GPU can read another's memory almost as if it were local, which demands very high bandwidth and latency in the range of hundreds of nanoseconds. A scale-out fabric connects whole servers or racks and moves messages rather than memory operations, tolerating microseconds of latency in exchange for reaching thousands of nodes. The split exists because physics limits how far you can run a memory-semantic link before signal integrity and latency break the illusion of shared memory. Training frameworks are written to keep the tightest communication inside the scale-up domain and the looser synchronisation across the scale-out one.
- Proprietary interconnects and open Ethernet. NVLink is a vendor-specific interconnect that pairs tightly with Nvidia's GPUs and switches, while Ethernet is a standardised technology that many vendors implement. Proprietary links can be tuned end to end and often lead on raw performance, but they tie the buyer to one supplier. Open standards let operators mix hardware and negotiate on price, which is why efforts to make Ethernet suitable for scale-up traffic attract so much industry attention. The transition is slow because software stacks and reliability track records take years to build.
- Collective operations and tail latency. Distributed training relies on collective operations such as all-reduce, where every accelerator contributes a slice of data and receives the combined result. A collective finishes only when its slowest participant finishes, so the job's speed is set by the worst path through the network rather than the average. That is why AI fabrics obsess over congestion control, lossless delivery and predictable latency instead of raw throughput. A single overloaded link can idle thousands of GPUs, which converts a networking flaw directly into wasted capital.
- Scale-across and the speed of light. Linking data centres across a metro area or a continent introduces latency dictated by distance, roughly five microseconds per kilometre of fibre, which no switch can remove. Workloads that span sites therefore split at coarse boundaries, such as running separate training stages or independent inference pools in each location. Operators accept this because a single site is capped by the power it can draw and the land it sits on. Scale-across networks are built for enormous bandwidth and encryption rather than for the microsecond precision of a scale-out fabric.
- Front-end and back-end networks. In an AI cluster the back-end network carries accelerator-to-accelerator traffic, while the front-end network handles storage, orchestration, user requests and the internet. They are often physically separate so that a burst of training traffic cannot starve storage reads or management access. Knowing which network a packet belongs to tells you which failure would stop a job, which would slow it, and which would merely inconvenience a user.
We teach this
How to use this in practice
- Draw the three tiers of an AI cluster. On one page, sketch a rack of eight accelerators joined by a scale-up fabric, a pod of many racks joined by a scale-out fabric, and two data centres joined by a scale-across link. Beside each tier, write the order of magnitude of latency and bandwidth you would expect and the protocol most often used there. Add a note on which tier a single failed link would hurt most. Done looks like a diagram where you can explain why a training job places each type of communication where it does.
- Measure latency at three distances. Run
pingandiperf3between two devices on your own local network and record round-trip time and throughput. Then ping a cloud endpoint in a nearby region and one on another continent, and compare the results with the fibre distance using the rule of roughly five microseconds per kilometre. Done looks like a table with three rows showing how latency grows with distance and how much of it is unavoidable. - Trace the path a packet takes. Run
tracerton Windows ortracerouteon Linux and macOS to a cloud service and count the hops. Note where the latency jumps most, and identify which hops are inside your provider, which are transit and which are inside the destination's network. Done looks like an annotated copy of the output with each segment labelled and one sentence on where a scale-across link would sit if you were the operator. - Time a collective operation on one machine. Using PyTorch's
torch.distributedwith the Gloo backend, start two or four processes on your laptop and run an all-reduce on tensors of one megabyte and one hundred megabytes, timing each. Then introduce a deliberate delay in one process and observe how the whole operation waits. Done looks like a timing table and a short note explaining why the slowest participant sets the pace for everyone.
Sources
- As AI Networks Scale In Three Directions, So Does Arista — The Next Platform
Our reporting is an original summary; full coverage is at the links above.
Don't just read about it — build it.
Square 1 teaches the skills behind the headlines, with every line of your work graded by AI. Find your starting point in 3 minutes.
Get your free skill report