How the AI Infrastructure Boom is Reshaping Fiber Optic Network Design

By Matt Brice, Senior Technical Marketing Manager, Digital Solutions, Prysmian North America

The artificial intelligence buildout is the largest infrastructure shift the data center industry has seen in a generation, and fiber optic networks sit at the center of it. U.S. data centers consumed 176 terawatt-hours of electricity in 2023, or 4.4% of total national electricity use, according to a Lawrence Berkeley National Laboratory (LBNL) report commissioned by the U.S. Department of Energy. That figure is projected to reach between 6.7% and 12% of total U.S. electricity by 2028. Behind every megawatt of new capacity sit miles of additional fiber, deployed in fundamentally different ways than traditional networks called for.

Why AI is Changing the Fiber Equation 

Three forces are driving the need for fiber optic cable: the unique qualities that make fiber the right medium for AI in the first place, a shift in how data moves inside the building, and a sharp increase in the volume of cable needed to support it.

Why Fiber is the Only Option for AI Workloads 

Fiber delivers near-unlimited speeds, easy upgrades, low power consumption, and minimal heat output, making it superior for high-quality communication. Single-mode fiber provides the bandwidth, low latency, and long-distance signal integrity that AI training requires. The same fiber installed today supports 400G links and will carry 1.6T tomorrow with a transceiver swap, which gives operators a long runway without ripping out infrastructure. Still, there is a massive shift in the number of fibers per rack required to handle AI workloads. The more fiber that is used, the less heat is generated, and less power is consumed.

A Large Shift From North-South to East-West Traffic 

A traditional data center handles requests from the outside world. A user sends a query, a server inside answers it, and the response goes back out. That up-and-down flow is called north-south traffic, and most older network designs were built around it. 

AI training works differently. User queries become tokens, which break down information into pieces to better produce a result. Language translation models were the first examples in which random samples of text needed to be broken down and analyzed to produce a correct translation. Thousands of GPUs sit in the same building and constantly exchange data to train a model. Almost none of that traffic ever leaves the building. It moves sideways from server to server, which is called east-west traffic.

The shift matters because east-west traffic in an AI cluster is far heavier than anything a traditional network was built to carry. GPU-to-GPU connections can run at 400 Gb/s and higher, and older hierarchical network designs simply cannot keep up. That’s why AI data centers need fiber networks built around east-west traffic from the start. This approach is strictly about traffic flows within the data center, yet it has the scope of the entire network.

Where to Scale? More Directions… 

Not to confuse you with more geographic coordinate references, but the flow of traffic and capacity increases are adapting simultaneously. Scaling up, out and across are part of today’s buzzwords as well. 

Data center operators will initially “Scale Up” capacity by vertically adding servers as needed within racks. Scale-up could start as small as adding RAM to a chassis or as large as installing a whole new server. Once Scale Up reaches a performance ceiling, Scale Out connects multiple compute nodes through a high-speed network to form a large-scale compute cluster. Beyond that, Scale Across is used to connect compute, storage and network resources further across multiple data center facilities.

A Step Change in Fiber Volume 

The volume of fiber needed inside an AI facility has climbed dramatically, with bandwidth purchased for data center connectivity surging nearly 330% between 2020 and 2024, according to the Fiber Broadband Association. The buildout extends between facilities as well. To keep up with demand, the U.S. will need to grow its fiber route miles from 95,000 to 187,000 by 2029, nearly doubling the existing long-haul network.

Design Shifts Defining the Next Generation of Fiber Networks

Higher Fiber Counts in Constrained Pathways 

AI racks require many times as many fiber connections as traditional cloud racks, and at hyperscale, the totals run into the millions per campus. Traditional cloud compute racks would require 20-40 fibers. The fiber density per rack has now catapulted to nearly 30x for AI. Outside the building, the demand for high-density cables continues. Cables carrying 1,728, 3,456, and 6,912 fibers are now standard, with multiple 6,912-count cables running between buildings on a single AI site. Pathway congestion is forcing designers to rethink conduit sizing, tray loading, and patch panel density across the entire facility.

The Transition to Reduced-Diameter Fibers 

200 µm fibers are replacing 250 µm fibers in dense environments. The smaller coating allows designers to fit significantly more strands into the same duct without sacrificing optical performance. For projects already constrained by existing pathways, the shift can extend the useful life of legacy infrastructure. Expect to see diameters of fiber coatings shrink, allowing high-density cables to set new records for fiber counts within a given diameter. The litmus test for compatibility of these reduced-size fibers will be compatibility if the ever-present 125 µm cladding standard for communications fibers is also reduced.

Ribbon Splicing as the New Standard 

Mass fusion splicing twelve fibers at once is no longer optional. Labor shortages and compressed project timelines have made ribbon-based architectures the default for hyperscale and AI builds.  Crews can now terminate in hours what would have taken days using single-fiber methods. And now, 16 fiber ribbons are here to stay, supporting Base-16 architectures inside the data center and on campus. Due to the high density now pushing into racks and rows, pigtail splicing is common with new 16F, 24F, and soon-to-be 32F Very Small Form Factor (VSFF) connectors, enabling reliable terminations at the GPU.

Next Generation Optical Fibers 

Hollow Core Fiber (HCF) and Multi-Core Fiber are two exciting developments in the industry, poised to meet the growing bandwidth and lower-latency demands of artificial intelligence. 

For HCF, Double Nested Anti-Resonant Nodeless (DNANF) hollow core fiber technology surpassed Photonic Band Gap (PGB) to achieve an attenuation below 0.10 dB/km at 1550nm while reducing latency by 50%. This technology is only now being industrialized and deployed in proof-of-concept projects, so current adoption will be slow. 

Similar to HCF, Multi-Core fiber is still in the early stages of development. For middle-mile applications, quad-core designs seem best able to offset non-linear effects, but shorter distances, such as within the data center or campus, could support a higher number of cores. So the determination of the number of cores and their profiles has stalled development. Challenges with splicing in the field may persist as this technology develops, and standardization is lagging behind market demand. Still, Multi-core fibers will play a role in addressing the data center market.

Convergence of Indoor and Outdoor Cable Designs 

AI campuses often span multiple buildings on a single site. Indoor/outdoor-rated cables that run from outside plant directly into the data hall reduce splice points and shorten commissioning timelines. Eliminating transition splices also removes one of the more common sources of long-term signal degradation.

Strategic Considerations for AI-Ready Network Design

Planning for Higher Rack Densities 

Traditional racks operated at 5 to 10 kW. AI workloads have pushed that figure dramatically higher, with Dell'Oro Group reporting current densities of 60 to 120 kW per rack and next-generation systems projected to exceed 600 kW. Fiber pathways and patching schemes need to anticipate these thermal and physical loads, since cable management choices made today will live next to liquid cooling lines and high-current bus bars for the next decade.

Planning for 800G and 1.6T Speeds 

400G links are already deploying at scale, and the next generation is moving even faster. The Ethernet transceiver market grew 93% in 2024 and is expected to grow 65% in 2026, driven largely by 800G shipments and the start of 1.6T volume production. Emerging fiber types, such as hollow-core fiber, may also extend the reach and speed of inter-data-center links, opening new options for site selection. Fiber decisions made in 2026 should still perform well in 2032, which means specifying fiber that supports both current transceivers and the optical budgets coming next.

Bend Performance as a Baseline Requirement 

Tight cabinet routing, dense patch panels, and the realities of field installation all penalize fibers that lose signal under stress. Specifying bend-insensitive fiber at the design stage prevents costly rework later and gives installers more flexibility on the floor.

Evaluating Total Cost of Deployment 

According to the Fiber Broadband Association, labor accounts for approximately 60% to 80% of total fiber deployment costs. Splice time, reel logistics, and rework all add to that figure once a project reaches the field. Cable choices that ship on longer reels or splice faster pay for themselves before the project closes out.

Prysmian Optical Fiber, Engineered for the AI Era 

Prysmian has manufactured optical fiber in North America since the 1980s, and that experience shapes every cable we ship today. Our optical fiber solutions are engineered for the density, performance, and reliability that AI data centers demand, with bend-insensitive single-mode designs, ultra-high-count ribbon cables, and indoor/outdoor options manufactured in North America. You can count on Prysmian to drive further innovation, such as Hollow Core Fiber, reduced diameter fiber coatings and high-density cable constructions. We work alongside designers, contractors, and owners to specify networks that perform tomorrow as well as they do today. 

Contact our team to learn how Prysmian’s optical fiber solutions can support your next build.