AI headlines usually focus on chips. The larger business story is the physical system around them: power, cooling, fiber and the telecom networks that connect data centers to the world.

The International Energy Agency expects global data-center electricity use to rise from about 485 terawatt-hours in 2025 to about 950 terawatt-hours in 2030. One terawatt-hour is one billion kilowatt-hours. That makes electricity access a deciding factor in where AI capacity can be built.[1]

Inside a training cluster, thousands of accelerators must exchange data at nearly the same time. An accelerator is a processor designed for AI calculations. When one network path is late, other processors may wait. The fiber plant therefore affects the output of some of the world’s most expensive computing systems.

AFL’s reference design for a 16,000-accelerator cluster illustrates the scale: roughly 250,000 fiber strands, about 16,000 MPO-8 terminations and about 20,000 fiber-kilometers in the back-end fabric. MPO-8 is a plug that aligns eight fibers at once. This is one design, not a universal bill of materials, but it makes the opportunity visible.[2]

Five markets worth watching

• Power infrastructure. Grid connections, substations, backup systems and on-site generation can determine whether a project opens on time.
• Liquid cooling. Higher rack power creates demand for cold plates, coolant distribution units, manifolds, pumps and maintenance.
• Dense fiber and testing. More network paths increase the value of compact connectors, clean assembly, inspection and accurate records.
• 800G and 1.6T optics. Faster Ethernet reduces the number of links needed for a given capacity, but suppliers must control power, heat and failure rates.[3]
• Telecom interconnection. International internet bandwidth reached 1,835 Tbps in 2025, up 23% in one year. Data-center growth therefore reaches metro, long-haul and subsea networks.[4]

Patents show where engineering effort is moving

Recent patent families point to hollow-core fiber, co-packaged optics, detachable high-density optical connectors and immersion cooling with heat reuse. Hollow-core fiber guides light mainly through air and may reduce delay. Co-packaged optics places optical engines close to a switch chip to reduce electrical loss. Both are promising, but manufacturability and serviceability will decide how quickly they scale.[5][6][7]

What are potentially the best business opportunities?

The strongest near-term products solve a measurable deployment problem. Examples include fibers commissioned per day, kilowatts cooled per rack, bandwidth delivered per watt, or hours saved in fault repair. The winning offer will often combine hardware, engineering, testing and lifecycle service.

For connectivity suppliers, let me bet on:

• High-density field-installable connections,
• Pre-terminated trunks,
• Inspection and digital traceability.

I would keep hollow-core fiber in controlled pilots and deploy integrated fiber sensing first on valuable routes where early fault detection has a clear financial return.

What is your opinion?

Which bottleneck is slowing your AI infrastructure project today: power, cooling, fiber availability, installation time or network operations?

Sources

1. International Energy Agency. Key Questions on Energy and AI. 2026.

2. AFL. Building AI Training Clusters at 16K Accelerators. 2026 supplied white paper. https://learn.aflglobal.com/resource-center/col/main/building-ai-training-clusters

3. IEEE Standards Association. Ethernet’s Next Bar is Now 800 Gb/s. 2024.

4. TeleGeography. International Internet Bandwidth Now Sits at 1,835 Tbps. 2025.

5. Google Patents. US12215048B2 hollow-core fiber manufacturing. 2025.

6. Google Patents. US10962728B2 co-packaged optics and transceiver. 2021.

7. Google Patents. US10667427B2 immersion cooling system for data centers. 2020.