What Is Scale-Across Networking in AI Infrastructure?

1. What Is Scale-Across Networking?

Scale-across networking is an AI infrastructure architecture that connects multiple geographically separated data centers or AI computing sites. It extends connectivity beyond the boundaries of one facility, allowing distributed AI resources to participate in a larger computing environment.

Traditional data center networking focuses on connecting servers, switches, and storage within a facility. Scale-across extends this connectivity over campus, metropolitan, regional, or longer distances, depending on the application and network design.

Unlike conventional data center interconnect (DCI), which has historically focused on connecting front-end networks, storage, and enterprise services between facilities, AI-oriented scale-across architectures may also connect the back-end compute fabrics. This enables distributed AI workloads to exchange accelerator-related traffic across sites.

2. Why Is AI Infrastructure Moving Beyond One Data Center?

Large AI facilities require substantial electrical power, cooling capacity, physical space, and network infrastructure. A single location may not be able to provide all the resources needed for future AI expansion.

Building additional data centers in different locations allows operators to distribute infrastructure investment and access new power and facility capacity. However, the sites must be connected with networks capable of supporting the communication requirements of AI workloads.

Scale-across is therefore not simply about connecting two data centers with a high-speed link. It requires coordinated design of compute resources, network transport, routing, congestion management, security, and workload scheduling.

3. Scale-Up, Scale-Out, and Scale-Across: Three Networking Dimensions

AI infrastructure uses different networking layers to connect accelerators at different physical and logical distances. Scale-up, scale-out, and scale-across form complementary parts of a larger architecture.

These layers have different performance requirements. Scale-up emphasizes very high bandwidth and low latency over short physical distances. Scale-out connects many servers within a facility. Scale-across must account for much longer distances, optical transport, propagation delay, and the operational boundaries between sites.

4. How Does a Scale-Across Network Work?

A scale-across architecture connects independent AI data center networks through a dedicated inter-site network. Each location typically contains its own compute infrastructure and internal scale-up and scale-out fabrics.

The inter-site network provides the transport path between these environments. Depending on the design, it may include routers, optical transport equipment, coherent transceivers, WDM systems, and high-capacity fiber links.

When a workload requires communication between sites, data travels from the source data center's network fabric to the inter-site transport network. The destination site receives the traffic and delivers it to the appropriate compute or storage systems.

The network may carry several types of traffic:

• Model distribution: Moving model files, checkpoints, and updated model versions between locations.

• Data synchronization: Coordinating datasets, metadata, and application state.

• Inference traffic: Distributing requests and exchanging information between inference services.

• Cache synchronization: Sharing selected cache or state information between distributed serving environments.

• Training communication: Carrying collective or synchronization traffic when a training job is deliberately distributed across sites.

Not every deployment requires all these traffic types to share one network. Separating front-end services, storage traffic, and back-end compute communication can make performance, security, and operational management easier.

5. Data Center Interconnect: The Foundation of Scale-Across

Data Center Interconnect (DCI) refers to the technologies and network architectures used to connect separate data centers. It can use optical fiber, coherent transmission, wavelength-division multiplexing, routers, and optical transport systems.

Traditional DCI commonly supports storage replication, data backup, application connectivity, disaster recovery, and communication between enterprise or cloud facilities. AI scale-across introduces additional requirements when the inter-site network carries back-end accelerator traffic.

Front-End DCI

Front-end DCI connects services and application-facing networks between data centers. Typical traffic includes user requests, application data, storage access, and selected synchronization operations.

These workloads may tolerate more latency than tightly synchronized GPU communication, although their requirements vary according to the application.

Back-End Scale-Across Connectivity

Back-end scale-across connects compute-oriented network fabrics between sites. It may carry distributed inference traffic or communication associated with a training job that spans multiple data centers.

This is more demanding than simply connecting front-end networks. If GPUs in different locations participate in synchronized collective operations, network latency, packet loss, jitter, and congestion can directly affect accelerator utilization.

6. Why Latency Becomes a Major Challenge

Physical distance introduces propagation delay. Even when a fiber link provides very high bandwidth, information still requires time to travel between locations.

In optical fiber, propagation delay is approximately 5 microseconds per kilometer in one direction. A 100 km fiber path therefore introduces roughly 500 microseconds of one-way propagation delay, or about 1 millisecond of round-trip propagation delay, before adding equipment and processing delays.

These are simplified propagation estimates based on fiber distance. Actual end-to-end latency also includes route length, optical equipment, switching, routing, queuing, and processing.

Impact on Distributed Training

Distributed training often uses collective operations to synchronize data across GPUs. When communication spans multiple sites, the additional propagation delay can increase the time required for synchronization.

Some workloads can overlap communication with computation, reducing the impact of network delay. Others depend on frequent synchronization and are more sensitive to distance.

Impact on Distributed Inference

Distributed inference can involve multiple stages or model components running at different locations. If a request needs information from a remote site, the additional network round trip can increase response time.

For this reason, inference systems often benefit from placing tightly coupled model components close together and using scale-across links for workloads that can tolerate the associated distance.

7. Optical Technologies Used in Scale-Across Networks

Optical connectivity is fundamental to scale-across because fiber provides the transmission capacity and reach needed to connect separate data centers. Different optical technologies are used according to distance, bandwidth, fiber availability, and network architecture.

Coherent Optical Transceivers

Coherent optics use advanced modulation and digital signal processing to transmit high-capacity data over optical fiber. They are widely used in metro, regional, and long-haul optical transport networks.

Coherent technology is particularly relevant when scale-across links need high bandwidth over distances beyond the practical reach of many short-reach direct-detect interfaces.

Wavelength-Division Multiplexing

Wavelength-Division Multiplexing (WDM) allows multiple optical channels to share a fiber by transmitting them at different wavelengths. This increases the capacity available from existing fiber infrastructure.

In scale-across networks, WDM can help operators expand inter-site capacity without requiring a separate fiber pair for every individual optical channel.

Optical Transport Systems and ROADMs

Optical transport systems provide the infrastructure needed to amplify, multiplex, route, and manage optical signals over longer distances. Reconfigurable Optical Add-Drop Multiplexers (ROADMs) allow wavelengths to be added, dropped, or redirected at different points in an optical network.

These systems are important when multiple data centers must be connected through a metro or regional optical network.

Pluggable Coherent Optics

Coherent optical technology is also available in pluggable form factors for compatible routers and switches. These modules can simplify some DCI deployments by integrating optical transmission capabilities into network equipment ports.

The appropriate coherent module depends on the required reach, line rate, optical power budget, fiber route, and host equipment support.

8. 400G, 800G, and 1.6T in Multi-Data-Center AI Networks

As AI clusters grow, the amount of traffic exchanged between sites can increase significantly. High-capacity optical interfaces help provide the bandwidth needed for distributed compute and inference environments.

400G Optical Connectivity

400G interfaces can support high-capacity connections between compatible routers, switches, and optical transport equipment. They are relevant to DCI deployments where the required capacity and distance match the available optical technology.

800G Optical Connectivity

800G interfaces increase the aggregate capacity available through a network connection. They can help reduce the number of physical links needed for a given bandwidth target, subject to equipment and optical system support.

1.6T Optical Connectivity

1.6T interfaces represent a further increase in port capacity for next-generation network platforms. They may be relevant to future AI infrastructure that requires higher-density connectivity between large network systems.

However, the data rate printed on an optical module does not establish its suitability for every DCI distance. Many long-distance scale-across links require coherent optical transport systems, while short-reach pluggable modules are designed for different applications.

9. Network Design Challenges in Scale-Across Architecture

Connecting data centers over long distances introduces challenges that are less prominent in a conventional single-site network.

Propagation Delay and Jitter

Distance creates unavoidable propagation delay. Variations in queuing and network processing can add jitter, which may affect synchronized workloads and latency-sensitive inference services.

Congestion and Buffering

Multiple high-bandwidth traffic flows may share inter-site links. When traffic arrives faster than the available egress capacity, queues build up. Buffering, traffic engineering, congestion control, and adaptive routing are important for maintaining stable performance.

Packet Loss and Recovery

Packet loss can trigger retransmissions and slow down communication-intensive applications. For tightly synchronized distributed training, a delayed or missing transfer can hold up other participants. Network resilience and reliable transport behavior are therefore important.

Fiber Availability and Capacity

High-capacity links require suitable fiber routes and optical infrastructure. Some locations may have limited fiber availability or high costs for additional capacity. WDM and higher-capacity optical systems can help make better use of available fiber.

Security and Data Governance

Inter-site networks must protect data moving between facilities. Encryption, access controls, traffic segmentation, monitoring, and data-residency policies should be considered as part of the architecture.

Operational Complexity

Multiple sites introduce separate power systems, network equipment, maintenance schedules, and failure domains. Centralized monitoring and coordinated network management help operators understand performance across the entire distributed environment.

10. Scale-Across Applications in AI Infrastructure

Not every AI workload needs to span multiple data centers. Scale-across is most useful when the benefits of distributed infrastructure justify the additional network complexity and latency.

For tightly coupled training, the distance between sites must be evaluated carefully. In some cases, distributing inference services or transferring model artifacts between facilities is more practical than extending every GPU collective operation across a regional network.

11. C-LIGHT Connectivity Products for AI Infrastructure

Scale-across networks require high-capacity connectivity at several layers. Optical transceivers and cable assemblies can support connections within data centers, while longer-distance inter-site links may require coherent optics and dedicated optical transport systems.

C-LIGHT provides high-speed connectivity products for data center and AI infrastructure applications, including:

12. Conclusion