Scale-up networking ties many accelerators into one logical computing unit with very high-bandwidth, very low-latency interconnect such as NVLink. Scale-out networking connects many of those units into a cluster, typically over Ethernet. Scale-across networking connects separate AI sites or whole data centers so they can share one training job. The three layers use different technologies, different reach, and different economics, and every AI factory is built from some mix of all three.
Scale-Up or Scale-Out: Which Does Your Build Need?
It depends on the size of the models and the site. If your training job fits inside one GPU domain, scale-up interconnect does the work. If you are assembling hundreds of accelerators across racks in one building, scale-out Ethernet carries the traffic between domains. If your accelerators sit in multiple buildings, campuses or regions, scale-across links carry gradients over distance. Most large AI factories run all three layers at once, and the mix you choose drives the switch, optics and cabling bill.
What Is a GPU Domain?
NVIDIA grew the scale-up domain from 8 GPUs to 72 as of July 2026 and targets 576, then 1,152 in future generations (see our fireside chat with NVIDIA SVP of Networking Gilad Shainer). Inside the domain, NVLink supplies extreme bandwidth and ultra-low latency. As domains grow, the reach and power limits of copper push the domain boundary toward optics and toward Ethernet-based scale-up fabrics, which is where the merchant switch silicon story begins (see What Is Merchant Silicon?).
How Big Does Scale-Out Get?
Scale-out Ethernet is where the money is. AI scale-out Ethernet revenue is set to surge past $100B by 2030, and the 800G generation of switching and NICs alone exceeds $50B in 26/27 with nearly 200M ports shipped, as of July 2026. We published the trajectory in In the AI Era, Ethernet Set to Surge in Scale Out and Ramp in Scale Up, and the switch silicon underneath climbs from 51.2 Tbps parts to 102.4 Tbps platforms emerging in 2H 2026, aimed at clusters beyond 100,000 accelerators (see our OCP Summit slate analysis).
What Is Scale-Across?
Scale-across connects AI sites so that no single building caps the training run. It leans on data center interconnect, long-reach coherent optics and multi-site scheduling. We set out the architecture in Today’s Distributed Hybrid AI Requires Investments in AI Fabrics, DCI WAN and Security. The 448 Gbps SerDes generation now in standards work is the electrical lane underneath the next round of this build-out (see our notes from the OIF’s 448 Gbps workshop).
How Was This Calculated?
The figures in this Explainer come from 650 Group’s published quarterly research on AI networking, data center switching and interconnect markets. Each figure links to its publication, and the full catalog is on our research reports page.
What to read next: What Are 400G, 800G and 1.6T Ports? for the port-speed ladder, What Is Data Center Networking? for the base layer, and Views from GTC 2025 for how the vendors frame the three layers. New quarterly data lands in our data center category.