AI is changing not only the servers inside a data center, but also the way facilities are connected across buildings, campuses, cities and regions. GPUs, liquid cooling and 800G Ethernet are important, yet they are only the first layer of the infrastructure required to move training data, model parameters, checkpoints, storage traffic and inference requests.

Why Are AI Data Centers Becoming More Distributed?

Compute capacity, available power, enterprise data and end users are rarely concentrated in the same building. The resulting terrestrial architecture can be understood as five connected layers: Server → Campus → DCI → Metro → Long-haul. Each layer has a different role, from connecting high-speed switches inside a data hall to linking regional compute and cloud resources.

Power is becoming a location constraint

AI-optimized servers consume far more power than conventional enterprise servers and increase cooling and electrical infrastructure requirements. The International Energy Agency estimates that global data center electricity consumption could more than double by 2030 to approximately 945 TWh, with electricity demand from AI-optimized data centers increasing more than fourfold over the same period. New facilities may therefore be built where power, land and cooling are available, rather than only beside established cloud or telecom hubs.

A single building may not hold the entire AI cluster

Large AI environments can span multiple data halls, several campus buildings, separate metropolitan data centers, regional facilities and dedicated training, storage, backup or inference sites. The farther data must travel, the more propagation delay becomes part of system performance.

Training and inference have different location requirements

Training generally benefits from large, concentrated GPU clusters. Inference may need to sit closer to users, factories, hospitals, financial markets or enterprise data. A distributed AI service can therefore combine centralized training, regional storage and model repositories, distributed inference facilities, and separate backup and disaster-recovery sites.

Layer 1: Server and Data Hall Fiber

Inside the data center, AI training requires continuous communication between GPUs, switches and storage. Large clusters exchange model parameters and intermediate results throughout a training job, so slow or unstable networking can leave expensive compute resources waiting for data.

Why cable architecture depends on the optical interface

An 800G port does not automatically determine one fiber count or cabling architecture. The physical connection depends on the transceiver, single-mode or multimode fiber, parallel optics or wavelength division multiplexing, duplex or multi-fiber interface, required reach, equipment generation and the planned migration to 1.6T. Start with the optical interface and migration plan rather than selecting a fiber base architecture from Ethernet speed alone.

Layer 2: Data Center Campus Fiber

An AI data center campus may include computing buildings, storage facilities, network rooms and utility systems. Campus fiber connects GPU buildings, central network facilities, meet-me rooms, carrier entrances, backup sites and operational systems. Unlike patch cords in a controlled data hall, these cables may pass through underground ducts, utility tunnels, manholes and outdoor transition points.

Important campus design factors include fiber count, route capacity, water blocking, mechanical protection, installation tension, bend performance, indoor/outdoor transitions, fire performance, splice locations and capacity for future expansion.

Route diversity must be physical

A campus can appear to have two network connections while still depending on one physical route. For genuine resilience, check primary and backup paths for shared ducts, manholes, bridges, building entrances, cable trays, distribution rooms and carrier facilities. Different providers do not guarantee physical separation when both circuits enter through the same duct.

Layer 3: Data Center Interconnect

Data Center Interconnect (DCI) connects two or more separate facilities. Traditional applications include replication, storage synchronization, cloud connectivity, business continuity, disaster recovery and workload migration. AI adds dataset movement, checkpoint transfer, model replication, GPU resource coordination and connections between training and inference sites.

NVIDIA describes the connection of distributed AI data centers as scale-across networking: connectivity that extends beyond scale-up links inside a computing system and scale-out links inside one data center. It can coordinate facilities across a campus or metropolitan area as part of a larger AI environment.

Distance still matters

Fiber carries information at approximately five nanoseconds per meter, or about five microseconds per kilometer in one direction. Switching, routing, buffering and application processing add further delay. Long-distance fiber can move large volumes of AI data, but bandwidth alone cannot remove the latency created by distance.

AI DCI performance is more than bandwidth

High capacity alone does not ensure predictable performance. AI DCI planning needs to evaluate available rather than advertised bandwidth, round-trip latency, jitter, packet loss, congestion behavior, path length, route diversity, failure recovery, maintenance windows, network visibility and the optical power budget.

Layer 4: Metro Fiber Networks

Metro fiber connects data centers, cloud access points and enterprise facilities across a city. A metro AI network can include hyperscale facilities, colocation sites, carrier hotels, internet exchanges, cloud on-ramps, enterprise AI locations, regional inference nodes and storage or backup sites.

Why AI increases metro fiber demand

Multi-site data center markets, location-sensitive inference and higher resilience expectations all raise metro connectivity needs. Architectures may require diverse entrances, ring or mesh topologies, high-fiber-count routes, dark fiber availability, multiple interconnection sites, rapid fault localization and sufficient duct capacity for expansion.

Dark fiber versus managed connectivity

Dark fiber gives an operator direct control over optical equipment and a fiber pair, but also requires optical engineering, monitoring and maintenance capabilities. Managed wavelength or connectivity services reduce operational responsibility, while capacity, route visibility and upgrade options depend on the provider. The fit depends on distance, required capacity, available routes, growth expectations, service-level needs and operating expertise.

Layer 5: Long-Haul Fiber Networks

Power-driven site selection means a new compute facility may be located near an energy source but far from a traditional data center hub. Long-haul fiber reconnects those facilities with major cities, cloud regions, enterprise data sources, regional inference sites, backup facilities and other AI computing locations.

What AI traffic moves over long-haul fiber?

Long-haul links are generally suited to traffic that can tolerate regional latency: dataset migration, model distribution, checkpoint replication, storage backup, disaster recovery, regional cloud connectivity, inference delivery and capacity balancing. They are less suitable for conventional tightly synchronized GPU communication.

Capacity is only part of the design

Long-haul planning also considers route distance, fiber attenuation, chromatic dispersion, optical transmission format, amplification, regeneration locations, splice count, repair accessibility, deployment method, natural-disaster exposure, construction risk and true route diversity. Physical-route auditing is essential because two logical lines can still share a bridge, railway crossing, utility corridor or aggregation site.

AI Fiber Highways as an Infrastructure Investment Theme

AI infrastructure spending is no longer limited to processors and servers. Capital is also moving into data center land, electricity supply, cooling, interconnection and regional networks. J.P. Morgan estimates that hyperscaler capital expenditure could reach approximately $697 billion in 2026. Although that figure covers more than connectivity, it illustrates the scale of infrastructure required to support AI growth.

Lightstorm provides one example of terrestrial fiber becoming part of this investment cycle. In August 2026, the company announced a ₹25 billion long-term debt facility to expand an AI-oriented connectivity platform. Lightstorm describes a terrestrial network spanning more than 30,000 kilometers and connecting more than 100 data centers, with an emphasis on capacity, low jitter and predictable performance between AI and cloud infrastructure locations. The broader significance is that connectivity is increasingly treated as a necessary AI infrastructure layer rather than a secondary utility.

The Complete AI Data Center Fiber Architecture

The five layers do not replace one another; they form a connected physical system. More AI compute creates demand for more internal bandwidth. More data halls create campus fiber requirements. Distributed facilities generate DCI traffic. Regional inference and cloud access increase metro connectivity, while power-driven siting creates demand for long-haul routes.

Seven Questions for AI Fiber Infrastructure Planning

1. What is the actual link distance? A data-hall connection has different requirements from a campus, metro or regional route.
2. What traffic will the link carry? GPU synchronization, replication, backup and inference traffic have different latency and availability needs.
3. How much capacity will be required in the future? Fiber count and route decisions should account for additional data halls, higher-speed interfaces and interconnection sites.
4. Are backup paths physically independent? Logical redundancy is not sufficient when both links share a duct, entrance or regional corridor.
5. What environment will the route cross? Indoor rooms, underground ducts, direct-buried routes and aerial infrastructure have different mechanical and environmental requirements.
6. Has the complete optical loss budget been calculated? Include fiber length, connectors, adapters, splices, distribution points and engineering margin.
7. Is the infrastructure fully documented? Useful records include route maps, fiber numbering, splice locations, polarity, insertion-loss results, OTDR traces and maintenance history.

Frequently Asked Questions

Conclusion

AI infrastructure does not end at the server rack. High-density optical links inside a data hall are only the first layer. As facilities expand, fiber must connect buildings, campuses, metropolitan compute zones and regional infrastructure. Tightly synchronized computing is most practical over short distances, while datasets, models, backup traffic and AI services can move across metro and long-haul networks. Scalable, predictable and physically resilient fiber is therefore central to connecting compute, data and service locations.

References

1. International Energy Agency, “AI is set to drive surging electricity demand from data centres,” April 10, 2025.
2. NVIDIA Developer Blog, “How to Connect Distributed Data Centers Into Large AI Factories With Scale-Across Networking,” September 9, 2025.
3. J.P. Morgan, “Financing AI Infrastructure and U.S. Data Centers.”
4. Lightstorm, “Lightstorm Launches India’s AI Superhighway; Secures ₹2,500 Crore Debt Facility,” August 11, 2026.