Spending on data center switches deployed in AI back-end networks will surpass $100 billion by 2030, driven by rapid expansion across scale-up, scale-out, and scale-across architectures, according to a new report from Dell’Oro Group. The forecast points to Ethernet emerging as the dominant technology across both scale-up and scale-out segments as AI workloads place new demands on interconnect bandwidth and topology.
The January 2026 Data Center Switch – AI Back-end Networks report highlights a shift in how AI infrastructure connects GPUs and memory. As agentic and physical AI workloads increase compute intensity, operators increasingly deploy scale-up fabrics to tightly couple accelerators within shared high-bandwidth domains, complementing traditional scale-out networking across racks and clusters.
Dell’Oro expects proprietary fabrics to coexist with open alternatives, but sees Ethernet gaining momentum alongside emerging options such as UALink. While NVLink remains prevalent in near-term scale-up designs, Ethernet’s economics, ecosystem depth, and roadmap position it as the long-term winner across AI back-end architectures, including deployments led by vendors such as NVIDIA.
- Spending on AI back-end switches is forecast to exceed $100 billion by 2030, spanning scale-up, scale-out, and scale-across domains.
- The majority of AI back-end switch ports have shifted to 800 Gbps, with transitions to 1600 Gbps by 2027 and 3200 Gbps by 2030.
- Co-packaged optics adoption is expected to accelerate during the forecast period, with early deployments led by NVIDIA.
- Neo Cloud providers are projected to be the fastest-growing customer segment over the next five years.
“The next wave of the AI journey—driven by agentic and physical AI applications—is placing unprecedented pressure on compute demand,” said Sameh Boujelbene, Vice President at Dell’Oro Group. “While we predict a strong adoption of UALink, we expect Ethernet to emerge as the long-term winner across both scale-up and scale-out architectures.”
🌐 Analysis
The forecast underscores a broader industry convergence around Ethernet as AI clusters scale in size and complexity, with switch silicon roadmaps pushing 1.6 Tbps and beyond per port over the next few years. At the same time, growing interest in scale-up fabrics reflects hyperscalers’ efforts to balance performance, power efficiency, and cost as GPU counts per node and per rack continue to rise.
