THOMAS RITZ THE FIBRE BOTTLENECK
THE AI BOOM IS CLEARLY VISIBLE BUT THE FIBRE BOTTLENECK IS HIDING AI hype, U.S. vs. China competition, GPU shortages, new hyperscale campuses, power grid pressure, and liquid cooling… for all of these, the optical ‘nervous system’ is the unseen enabler. AI does not happen inside a single chip but across thousands of processors that need to exchange data constantly. That makes data movement as important as data processing, writes R&M Market Manager Public Networks Thomas Ritz, MBA .
AI-capable data centres require a different approach because training AI applications means moving unprecedented volumes of data in parallel, in real time, and over short distances between processors. Traditional data centre planning will not suffice for AI clusters. The AI data centre is no longer a single building but a connected fabric. We are seeing a topology shift from individual server rooms, isolated racks, and single data halls, data centres, and campuses to AI clusters, high-density rack rows, interconnected rooms, multi-building campuses and data centre interconnects. Meet-me rooms have become key ‘exchange points’ where carriers, cloud providers, colocation tenants, and internal networks meet. Building entrance facilities and optical distribution frames become the physical control points for AI-era capacity. Solution requirements need to encompass building entrance facilities, meet-me rooms, backbone cabling, fibre
horizontal cabling, computing solutions, operation, and maintenance.
Mass-fusion splicing is now a productivity and reliability requirement, not a niche technique. Single-fibre splicing becomes impractical at massive fibre counts and 12-, 16-, or potentially higher-count ribbon splicing speeds up project timelines. Routing large numbers of ribbons into one closure is not just a capacity problem but a handling, strain- relief, bend-radius, and identification problem. COOLING AND CABLE OVERLAP High-density racks exceeding 30 kW are now common in many data centres, especially for LLM and AI workloads, and liquid cooling is already state-of-the- art for high-performance computing or rack capacities of 30 kW and above. But cable diameter, cable-bundle density, and routing influence airflow; chaotic cabling in front of active equipment can block cooling openings. AI racks are not cooled in isolation. Physical discipline of the rack is also vital: cable paths, airflow, power distribution, sensor placement, and service access. Racks, liquid cooling, rack intelligence, PDUs, and DCIM should be covered by integrated infrastructure thinking rather than isolated products. FIBRE VERSUS COPPER: A DESIGN DECISION Copper still has a role in very short, dense scale-up links, especially within or near racks. Fibre dominates when distance, density, bandwidth, electromagnetic immunity, and energy efficiency become more important. IEEE Spectrum describes the distinction: scaling out mostly relies on photonic chips and optical fibre over hundreds or thousands of metres, while scaling up often relies on short copper links of around a metre or two. However, copper approaches physical limits at terabit-per-second bandwidths because it must become shorter and thicker.
THE HIDDEN COMPLEXITY BEHIND AI-SCALE FIBRE BUILDS One important, but often overlooked problem: thousands of fibres have to go somewhere. For large AI builds, the problem is not simply ‘install more fibre.’ The real challenge is where the fibre enters, where it is split, where it is spliced, where it is stored, how it is identified, how it is routed, and how technicians can touch it without damaging it. • Building entry: high-count cables arrive at the facility and must be routed into building entrance facilities or splice cabinets. Ribbon type cables have to be handled and routed to or inside the optical distribution frames (ODF) and modules where ribbon or single splicing takes place – translation/transition. • ODF upgrades: optical distribution frames need higher density, better cable guidance, better labelling, better access, and cleaner entry paths into cassettes. • Closures and connection boxes: outside the largest hyperscale data centres, the network still depends on outdoor closures, splice points, and access nodes. For 6,000-plus- fibercables, closures become large, complex engineering objects rather than simple protective boxes. • Ribbon fibre orientation: high-density ribbon fibres may need different vertical routing, cassette placement, bend-radius control, and cable-group handling. • Human factors: the more fibres in one closure, the more important it becomes to make the closure technician-friendly. A poor design raises the risk of fibre breaks, routing errors, untraceable fibres, and slow restoration.
20
www.opticalconnectionsnews.com
| ISSUE 44 | Q3 2026
Made with FlippingBook Annual report maker