How Ethernet, AI Infrastructure, and ecosystem partnerships are driving a $70B+ market shift
Introduction
Networking plays a key role in AI buildouts. From a user’s perspective, we might see this as a faster response time or better results, but behind the scenes, the network plays the vital role of connecting everything. This importance grows in the Agentic AI wave, where multiple specialized models and tools are stitched together into pipelines. Customers require choices based on the business, technical capabilities, and operating model. This blog defines the key AI network types and explains how different solution choices map to buyer segments, ultimately providing readers with a deeper understanding of AI networks and additional resources for further learning.
Definitions
AI Frontend Networks – These are the connections from the AI pod to the end-user, an organization’s data, and real-time information from the internet. Frontend networks generally resemble traditional data center leaf-spine fabrics (policy, QoS, security, observability).
AI Backend Scale-Out Networking – These are private horizontal scaling networks that connect multiple servers/racks. They leverage low-latency network fabrics, such as InfiniBand and Ethernet, where Ethernet became the majority of connections in early 2025. Scale-Out networks typically run ~10x the bandwidth per server compared to traditional frontend connectivity, measuring up to 800 Gbps per GPU and several Tbps for each server. Customers will rapidly embrace 1.6 Tbps as platforms come to market. Traffic is east-west intensive; latency and congestion management are key performance metrics.
AI Backend Scale-Up Networking – These are private vertical scaling networks that connect GPUs directly to each other, enabling GPUs to access memory directly on different GPUs. This allows many connected GPUs to act like one giant GPU, which improves end-user performance and reduces overall cost. Currently served by NVLink or PCIe, Ethernet enhancements are expected to play a greater role in Scale-Up starting in 2026. Due to their memory connectivity, Scale-Up networks typically offer order-of-magnitude higher bandwidth than Scale-Out networks, with each GPU measuring in the Tbps range.
AI Networking Becomes New $70+B Market
AI architectures are evolving so rapidly that there is no one-size-fits-all approach to building them. AI networking’s origins in HPC give the impression that it has been around for a long time. Still, the emergence of these new networks, along with their size, makes it one of the most essential networking categories and important subsegments in AI. Without AI networking, we couldn’t conduct the training or inference necessary to make AI a multi-trillion-dollar and highly disruptive market.
Open Standards and Architectures
There is a long list of standards, architectures, and Multi-Source Agreements (MSAs) to help drive the industry forward. For example, the Ultra Ethernet Consortium (UEC) is an open specification that enhances Ethernet to provide a low-latency, lossless fabric to Scale-Out and HPC networks. Future iterations are expected to enhance Ethernet for Scale-Up. Cisco was one of the founding steering members for UEC. UEC strives to meet the growing network demands of AI & HPC at scale.
Hyperscalers
These customers represent the largest and technology-leading operators, pushing the edge of what is achievable at any point in time. Cisco addresses this market with the Cisco 8000 Series Routers and Switches, Silicon One ASICs, optical systems components, and optical transceivers, with an open NOS option (e.g., SONiC) where appropriate. This customer set seeks full customization and a multi-vendor supply chain, making interoperability key. Validated interoperability at optics and fabric layers is a primary decision driver.
NeoClouds and Sovereign Clouds
These customers represent the fastest-growing operators and will make up an increasing portion of the market each year. These customers bring strong technical capabilities and are actively seeking partnerships, positioning themselves for rapid and agile growth, even without the scale of hyperscalers. Time-to-first-token and predictable Scale-Out are top priorities. Cisco addresses this market with the Nexus 9000, Cisco 8000 Series Routers and Switches, Silicon One ASICs, pluggable optics, and a mix of NX-OS, IOS XR and SONiC NOS options. Cisco’s Sovereign Critical Infrastructure portfolio addresses European digital sovereignty demands, offering configurable, air-gapped, on-premises solutions for more control, data autonomy, and EU certification compliance.
Enterprises
These customers represent the largest mix of technical capabilities, application diversity, and data gravity. These customers require simplified network operations to scale, often seeking validated designs and architectures. Private cloud at scale and unified fabric operations (brownfield + greenfield) help them move quickly. Cisco addresses this segment of the market with the Nexus 9000 and Nexus Dashboard, which allows for simplified network management. Telemetry-driven visibility and policy automation help teams correlate workload health with fabric behavior. Enterprise customers can benefit from adopting Smart switches, which enhance networks by adding embedded security capabilities.
Telco SPs
These customers play a key role, not only as consumers of AI, but also as providers of the pipes and connectivity for all other customers. Cisco addresses this segment’s internal demands with the Nexus 9000 and Nexus Dashboard and the external connectivity with the Cisco 8000 Series Routers and Switches, Cisco’s Routed Optical Networking offerings, and coherent pluggable optics. DCI architectures leveraging ZR/ZR+ coherent pluggables and metro-scale interconnects are central to rollout plans.
Intra Data Center Optics
There is a significant focus on optics in the data center to support the new AI fabric connections. For the next few years, the industry will use a mix of 400G and 800G optics, but the industry is innovating rapidly. Optics extends beyond the pluggable module to encompass deeper integration with Silicon Photonics technology, and closer integration with the switch ASIC to meet optical transceiver performance, reliability, and volume requirements while reducing power consumption and latency. At the same time, coherent technology will make its way into the data center, and co-packaged optics will bring systems and optics even closer together. These trends increase the importance of optics reliability, multi-vendor interoperability, and vertical integration for performance, yield, and availability.
DCI Optics (ZR, ZR+, Coherent Lite)
While most optics will be inside the data center, linking facilities together becomes key as operators look to multiple sites to support power requirements. This market will grow materially as customers deploy 400G, 800G, and future 1.6T coherent pluggable modules across routing and switching platforms to extend the walls of the data center beyond one physical location. Operationally, this brings optical planning (chromatic dispersion, OSNR) into the DC networking teams’ day-to-day operations.
Cisco’s NVIDIA and AMD Partnerships
The journey to AI is not something that any single vendor or offering can give the customer everything they need. Instead, partnerships help provide more options and best-of-breed solutions. Customers want vendors to collaborate to create a better solution. For example, NVIDIA and Cisco’s partnerships allow for Cisco’s networking stack to integrate with NVIDIA’s ASICs and GPUs to provide a lower TCO and an easier-to-manage solution. Similarly, Cisco and AMD collaborated on both the AI stack and a DPU to enable a Smart Switch (adding programmable data-plane functions), thereby enhancing the networking layer.
Cisco’s Approach
Cisco has taken a unique approach to the AI networking market by offering a flexible, end-to-end solution targeting systems, silicon, optics, software, and security infused into every layer. This approach allows customers to consume what they want and need. For Hyperscalers, it’s a mix of components and co-developed systems. Neoclouds, Enterprises, and Telco SPs benefit from this scale and innovation, consuming the technology through systems and validated designs to move quickly. This approach is validated by a diversified AI networking pipeline across segments and geographies.
Conclusion
Ethernet will continue to grow in AI networking, leveraging cost, scale, and interoperability to allow customers of all sizes to accelerate their AI journey. By enabling customers to reduce their time to first token or achieve higher uptime, Ethernet helps organizations improve ROI via faster job completion and reduced disruption. For readers evaluating fabrics, prioritize validated designs, multi-vendor interoperability, and telemetry-driven operations to de-risk deployment.