As AI workloads drive unprecedented demand for bandwidth, lower latency and greater network density, Rahul Puri, CEO, STL, examines how multi-core and hollow-core fibre technologies are transforming the physical infrastructure underpinning the next generation of AI computing.
When generative AI video models first simulated hyper-realistic, photorealistic physical worlds in motion, they signalled a massive, structural transition in computing. Generating seconds of high-fidelity, multimodal video or running real-time, multi-step logical reasoning models demands an unprecedented scale of parallel computation.
Yet, while the world focuses on software capabilities and silicon chips, the physical highway that binds them together remains largely invisible.
Behind every breakthrough AI model lies a physical reality: high-density compute clusters that must continuously exchange petabytes of complex data.
In this new era of distributed intelligence, the limiting factor is no longer just the computational capacity of individual accelerators, but the latency and density of the fibre optic highways connecting them.
As artificial intelligence scales from isolated text queries to massive, persistent physical-world simulators, the physical layer must undergo its own evolutionary leap. Fibre is no longer a passive utility; it is the active, high-speed nervous system of the AI factory.
Inside the AI cluster
To understand why this infrastructure must evolve, we must look at how AI has transformed data traffic flows within modern network facilities.
Traditional cloud computing architectures were primarily designed around data flowing between external users and centralised servers. In contrast, large-scale AI training models rely on massive parallel processing across thousands of interconnected accelerators, completely inverting these dynamics.
Within these high-density AI clusters, traffic is overwhelmingly moving continuously and in parallel between switches, GPUs and memory banks.
Driven by the constant need for parameter synchronisation and gradient distribution, this internal data exchange can be 5x to 10x higher in volume than external traffic.
This shifting architectural demand is triggering a historic surge in fibre densification across the industry.
Market projections from McKinsey indicate that AI-ready workloads will represent ~70% of total data centre capacity demand by 2030, representing a remarkable 33% CAGR from 2026 to 2030.
Attempting to squeeze these massive, continuous traffic volumes through legacy cabling infrastructure leads to severe operational bottlenecks. Network operators must transition to next-generation physical-layer architectures.
The pathway congestion: The economics of ‘not digging’
To connect GPU clusters, operators require ultra-high-density cables containing 1,728 to 6,912 fibres.
However, physical pathways leading into and running between data halls, including conduits, trays and ducts, are fixed assets.
Standard 250-micron coated fibre cables are too thick and stiff to navigate congested pathways without restricting cooling airflow or risking signal loss from micro-bending.
Digging trenches is a logistical nightmare in dense urban areas. More critically, civil works represent up to 80% of total fibre deployment costs.
Under these constraints, the economics of ‘not digging’ become a strategic mandate. To multiply capacity within standard conduits, we must move beyond conventional glass geometries.
Multi-Core Fibre (MCF): The spatial multiplier ready for scale
Where traditional fibres squeeze data throughput by pushing the spectral efficiency of a single glass core, they eventually collide with the physical ‘Shannon Limit’, where packed wavelengths or high baud rates introduce severe optical noise and distortion.
Multi-Core Fibre (MCF) solves this challenge through Space Division Multiplexing (SDM). Instead of stressing a single channel, MCF embeds multiple independent cores within a single strand of glass, multiplying capacity within the same physical footprint.
Crucially, MCF has graduated from research labs to become a mature, commercially viable technology. MCF is a primary focus for modern capacity planning.
By packing multiple cores (typically four or seven) into a standard 125-micron cladding diameter, MCF maintains the exact physical geometry of conventional SMF while keeping coating size at 200 to 250 microns.
This keeps the cable’s bend radius, weight and form factor fully compatible with existing conduit systems. Crosstalk between these closely packed cores is validated to less than -40 dB, ensuring absolute signal isolation and integrity.
The commercial readiness of MCF is validated by successful deployments.
For example, we collaborated with Colt Technology Services to execute MCF trials across their London Metro optical network, achieving an 800 Gbps line rate over existing metro ducts.
Furthermore, MCF has entered commercial outside plant (OSP) and inside plant portfolios, marked by the world’s first deployment of MCF in both aerial and underground networks and the UK launch of an advanced Unitube Single Jacket Indoor Optical Cable featuring 4-core MCF.
For network planners, MCF represents a mature, immediately deployable blueprint to quadruple spatial capacity without civil works.
Hollow-Core Fibre (HCF): Light at the speed of air
During large-scale LLM training, thousands of GPUs must wait for parameter synchronisation before proceeding to the next calculation.
Standard fibre optic cables guide light through a solid silica glass core, which slows light down to roughly 124,000 miles per second.
Hollow-Core Fibre (HCF) addresses latency by guiding light through an air-filled structure rather than solid silica.
Because light travels ~46% faster in air than in glass, HCF reduces signal transit time (group index 1.0006 for HCF vs 1.4678 for SMF), shaving off ~2.5 microseconds of latency per mile.
Research published by Microsoft indicates that latency reductions of this scale can improve large-model training throughput by several percentage points, a massive saving on multimillion-dollar training runs.
Today, advanced HCF designs have achieved a compact 400-micron fibre diameter, preventing the cable from becoming too bulky.
Testing has demonstrated error-free 200G, 400G and 600G transmission across the entire optical C-band, proving HCF can support high-speed coherent transmission without compromising signal stability or DGD.
The ecosystem blueprint: Bridging the gap to field readiness
Transitioning these optical innovations into wide-scale field deployments requires solving four critical ecosystem pressure points:
- Standards and Interoperability: Establish unified ITU-T and TIA standards defining core geometries, crosstalk thresholds and optical parameters.
- Splicing and Test Workflows: Mass fusion splicers must be optimised to align and splice multi-core fibres in the field. Similarly, OTDR testing equipment must evolve to automate multi-core verification.
- Advanced Connectorisation: Pre-terminated assemblies are essential to bypass skilled labour shortages (facing a gap of 180,000 workers in the US alone). For HCF, Thermal Expansion Core (TEC) connectors ensure backward compatibility with standard network nodes.
- Hybrid Fibre Architectures: Deploy HCF on latency-critical synchronisation loops, MCF on high-density distribution backbones and standard SMF for legacy applications. STL has engineered this hybrid approach into a single ‘Tri-Fibre’ cable architecture, bundling HCF, G.654.E and standard G.657.A1 fibres within a single protective sheath.
As we look towards 2030, fibre is no longer a utility. The optical physical layer has officially evolved into a strategic asset that defines the latency, density and deployment speed of the entire AI compute stack.
By establishing unified standards, scaling installer training and designing hybrid architectures today, we ensure that the physical highway remains the enabler, and never the bottleneck, of the AI-driven future.
