How AI is reshaping power, cabling, and connectivity in modern Data Centres

AI isn’t only transforming what happens inside data centres - it’s reshaping the facilities themselves. Surging power consumption, soaring rack densities, new cooling demands, ultra-high-speed networking, and the shift toward 800G and 1.6T infrastructure are redefining how data centres are designed and operated. This blog dives into how AI workloads are driving unprecedented changes in power, cabling, and connectivity requirements, and what organisations should prioritise when building an AI-ready backbone.

The AI Era: A New Baseline for Power and Infrastructure

Artificial intelligence is pushing data centres into a fundamentally new design paradigm. As training and inference workloads explode, electricity consumption is rising sharply. Industry forecasts indicate that global data-centre power usage could more than double by 2030 -  with AI as the primary accelerator.

At the rack level, the shift is dramatic. Densities that previously hovered in the single-digit kilowatt range are now commonly 10–30 kW, while cutting-edge AI clusters exceed this and require liquid cooling as a standard. This reality is forcing operators, utilities, and planners to rethink everything from grid connections and cooling topologies to the physical design of cabling and connectivity systems.

Balancing AI’s Energy Demands with Efficiency

Over the past decade, data centres have significantly boosted energy efficiency. Average PUE has improved from around 2.20 in 2010 to roughly 1.55 in 2022, with AI-centric facilities now targeting sub-1.3 figures. Yet AI introduces a paradox: while it promises solutions to global sustainability challenges, it increases energy demand today.

Large-scale AI training is famously resource-intensive. The IEA projects data-centre electricity consumption could double between 2022 and 2026, driven heavily by AI. A University of Massachusetts Amherst study found that training a single large AI model can emit over 626,000 pounds (284,000 kg) of CO₂ — more than the lifetime emissions of five cars.

This introduces a classic rebound effect: efficiency gains make AI cheaper to operate, which then encourages more use, ultimately increasing total energy consumption.

The central challenge emerges clearly: how can the industry harness AI’s potential without escalating environmental impact?

Strategies Already Underway

To address this tension, data centres are deploying or developing several complementary approaches:

• Renewable energy integration: Transitioning to carbon-free power sources such as small modular nuclear reactors or large-scale solar/wind.
• Optimised AI models: Algorithms designed to balance performance and energy efficiency.
• Energy-efficient hardware: Purpose-built chips and systems that cut total energy draw.
• Advanced cooling: Liquid cooling, immersion cooling, and next-generation heat-extraction systems reduce energy spent maintaining safe temperatures.

Cabling and Connectivity: The Next Big Bottleneck

The explosion of large AI model deployments has revealed a new constraint: network performance. AI clusters rely heavily on GPU interconnects, often using RDMA to minimise CPU involvement and maximise throughput.

Traditional networking is no longer adequate:

• Ethernet alone cannot meet large AI cluster requirements.

• Even advanced standards such as InfiniBand and RoCE are pushing their limits.

• 800G is beginning to bottleneck — the industry is accelerating toward 1.6T.

Why AI Data Centre Networks Are Different

AI-centric environments require:

1. Extremely high-speed networking

2. Ultra-low latency across massive GPU clusters

Networks are evolving from 400G → 800G → 1.6T at unprecedented speed. This shift directly impacts structured cabling requirements (ISO/IEC 11801-5, TIA-942), making high-density MPO systems essential.

The Technology Behind the Upgrade: SerDes and Channel Speed

Bandwidth increases rely on two elements:

• SerDes channel speed (25G → 50G → 100G → 200G per lane)

• Number of channels used

For example:

• 800G can be achieved with 16×50G, 8×100G, or 4×200G solutions.

• A 16-channel setup needs two MPO-16/24 connectors.

• A 4-channel solution needs only one MPO-8/12 connector — simplifying infrastructure, lowering costs, and improving energy efficiency per Gbps.

As channel speeds increase, cabling becomes less complex and more cost-effective.

Latency: The Silent Performance Killer

AI performance is highly sensitive to latency. When scaling to 10,000+ GPUs, a three-tier switching architecture may introduce up to five switch hops — contributing far more latency than the fiber or copper links themselves.

To keep pace with AI scale, cabling infrastructure must evolve faster than ever. Flexible, forward-looking structured cabling ensures operators do not need to repeatedly overhaul concealed infrastructure as speeds rise.

Standards such as ISO/IEC 11801-5 and TIA-942 support scalable designs including ToR, Spine-Leaf, and Mesh topologies.

What to Look for in AI-Ready Data Centre Cabling

Building a future-proof AI backbone requires cabling designed for:

1. Alignment with Optics and Switch Roadmaps

• Native support for Base-8 and Base-16 parallel fiber

• Compatibility with today’s 400G/800G SR optics

• No stranded fiber when migrating to 1.6T

The IEEE P802.3dj project is already defining 200G/400G/800G/1.6T standards for the coming wave of AI hardware.

2. Simplified Installation and Reduced TCO

Pre-terminated, high-density cabling systems can reduce install time, minimise risks, and save up to 40% in pathway space.

3. High-Density Patch Systems

These systems:

• Reduce complexity

• Enhance maintenance efficiency

• Support VSFF connectors and emerging multi-core fiber technologies

4. Design for Density and Migration

Whether in small enterprise facilities or hyperscale clusters, the approach is consistent:

• Plan for LC/MPO density and future migration

• Use auditable patching workflows

• Right-size media by room, distance, and topology

Tailored Considerations by Facility Type

• Enterprise / Small DCs: Frequent MACs → mixed copper/multimode, UHD LC, MPO-LC.

• Colocation: Rapid tenant turn-ups → HD/UHD frames, MPO backbones, strong labeling, AIM workflows.

• Hyperscale / AI Rows: OS2 fibre, ultra-low-loss MPO, micro-bundles, pre-terminated trunks.

• Edge / Micro DCs & Outdoor POPs: Compact, low-staff → pre-terminated LC/MPO, remote visibility, rugged enclosures.

Conclusion

AI is redefining every layer of the data-centre ecosystem — from rack densities and cooling, to grid design, to the cabling and connectivity that hold everything together. As the industry pivots towards 800G and 1.6T, structured cabling and forward-thinking connectivity strategies become critical to ensure efficiency, scalability, and long-term performance.

The data centres that thrive in the AI era will be those built on dense, flexible, migration-ready physical infrastructure — engineered to evolve as fast as AI itself.

Share this