The Numbers Are Staggering
Consider what a single large language model training run requires. GPT-4 class models train across tens of thousands of GPUs, each needing 400-800 Gbps of interconnect bandwidth. A 100,000 GPU cluster — which NVIDIA, Meta, and Microsoft are all building — generates internal network traffic measured in exabits per second.
All of that traffic flows over fiber. Every GPU-to-switch connection, every spine-leaf link, every cross-building connection in a data center campus is a fiber optic cable. A single hyperscale AI data center can contain over 10 million fiber connections.
Meta alone announced plans for a data center campus that will consume over 5 GW of power — roughly the output of five nuclear power plants. The fiber infrastructure to connect the GPUs inside that facility is measured in thousands of kilometers.
The Three Waves
Wave 1: Intra-Data Center (Happening Now) The immediate bottleneck is inside the data center. AI clusters need massive east-west bandwidth between GPU racks. This is driving explosive demand for: - 800G and 1.6T optical transceivers - Multimode fiber for short reaches - Single-mode fiber for longer campus connections - Co-packaged optics that put photonics directly on the switch silicon
Coherent optics, once reserved for long-haul networks, are now being deployed inside data centers as reach requirements grow. Companies like Ciena, Infinera, and a wave of startups are racing to meet demand.
Wave 2: Data Center Interconnect (2025-2028) AI training increasingly spans multiple buildings and even multiple campuses. When NVIDIA talks about "AI factories," they mean distributed systems connected by fiber. The bandwidth between data center buildings is growing 40-60% annually.
This is where Lumen Technologies has made a strategic bet. Their network connects major data center markets, and they've signed multi-billion dollar deals specifically for AI interconnect capacity. The company's stock has reflected this positioning, rising over 300% from its lows.
Dark fiber providers like Zayo and Crown Castle are seeing unprecedented demand for metro fiber routes between data center clusters.
Wave 3: Edge AI and Inference Networks (2027-2032) Today's AI runs mostly in centralized data centers. But the next phase pushes inference — the actual use of trained models — to the edge. Self-driving cars, augmented reality, real-time translation, and AI assistants all need low-latency access to powerful models.
This creates demand for a new layer of fiber infrastructure connecting edge data centers, cell towers, and enterprise locations. It's essentially FTTH (fiber to the home) but for AI — fiber to every place that needs to think.
Supply Chain Under Pressure
The fiber industry is struggling to keep up. Corning, the world's largest fiber manufacturer, has expanded capacity multiple times but order backlogs continue to grow. Prysmian Group, the largest cable manufacturer, reports that their fiber cable capacity is essentially sold out for the next two years.
This is driving several responses: - New manufacturing capacity: Multiple companies are building new fiber draw towers, but a new factory takes 18-24 months to come online - Alternative materials: Research into new fiber types, including multi-core fibers that carry multiple signals through a single strand - Recycling and efficiency: Better cable designs that fit more fibers into existing conduit
The Geographic Shift
AI is also reshaping where fiber goes. Traditional network topology centered on population centers. AI topology centers on: - Power availability: Data centers go where cheap, reliable electricity exists - Climate: Cold climates reduce cooling costs - Land: AI campuses need hundreds of acres
This is pulling fiber into places it's never been before. Rural areas near hydroelectric dams, wind farms, and nuclear plants are suddenly becoming endpoints for major fiber routes. The BEAD program and other broadband funding initiatives are coincidentally building fiber to many of these same areas, creating a dual benefit.
What This Means
We are in the early innings of a fiber deployment cycle that will dwarf anything that came before. The original internet buildout in the late 1990s laid about 1 billion kilometers of fiber globally. AI-driven demand could double that within a decade.
For investors, the picks-and-shovels plays are clear: fiber manufacturers (Corning, Prysmian), network operators with data center routes (Lumen, Zayo), and optical component makers (Ciena, Coherent, II-VI). But the bigger story is economic: regions and countries that build fiber infrastructure faster will attract AI workloads, and with them, the industries of the future.
The fiber optic cable — a technology invented in the 1970s — is quietly becoming the most critical infrastructure of the AI age.