Arista reported $2.71B in Q1 2026 revenue, up 35.1% year over year, and raised full-year AI networking guidance to $3.5B. The numbers are strong but not the signal. The signal is the naming: Arista is now building products around three distinct AI fabric tiers (scale-out, scale-across, and scale-up), and scale-across is the new one. It is the Ethernet switching layer inside the AI factory floor pod, connecting GPU racks to each other before they hit the spine. This is not a rebrand. It is a distinct traffic pattern requiring distinct topology.
Scale-out is the north-south fabric connecting compute, storage, and network at the data center level: high port count, asymmetric, mostly solved. Scale-up is the east-west fabric inside a compute node pod: NVLink and InfiniBand for GPU-to-GPU collectives, now with Arista's CX-8 Spectrum targeting the Ethernet entrant. Scale-across sits between them: it moves traffic among GPU racks within the same AI factory floor, at bandwidths and latencies between the other two tiers. Arista has scale-across in current deployments and has announced 1.6 Tb/sec scale-up Ethernet ports for 2027 to attack InfiniBand directly.
For teams designing AI compute infrastructure, naming this layer matters because it surfaces a hidden design decision that most teams are currently resolving ad hoc. A cluster built without a deliberate scale-across layer ends up using scale-out topology for intra-pod traffic, which adds latency and wastes port budget on traffic that should stay local. Arista announced it first; the vendors selling InfiniBand-based all-in-one pod solutions have 12-18 months before customers start treating scale-across as a separate line item on the infrastructure RFP.