AI Infrastructure • 6 June 2026 • By AI Conference London Editorial

Optical Networks and the AI Infrastructure Boom

Ciena's optical solutions are critical for managing the massive bandwidth demands driven by AI in data centers, ensuring high-speed, reliable connectivity.

The artificial intelligence boom, fuelled by ever-larger models and insatiable datasets, is placing unprecedented strain on the world's digital infrastructure. While much of the focus has been on the computational power of GPUs, the underlying network fabric responsible for moving these vast quantities of data is reaching a critical inflection point. The performance of AI is increasingly becoming a function of network capacity, making optical networking a cornerstone of the next wave of innovation.

The Unseen Backbone: AI's Dependence on Data Centre Interconnects

The nature of high-performance computing has fundamentally shifted with the rise of generative AI. Traditional enterprise workloads were often siloed, with predictable traffic patterns. In contrast, training a large language model (LLM) or a complex computer vision system involves a distributed process where massive model parameters and training datasets are constantly shuffled between thousands of processors. This creates an enormous amount of "east-west" traffic within and between data centres, a workload that legacy network architectures were not designed to handle efficiently. Source

This is where Data Centre Interconnect (DCI) technology becomes paramount. DCI refers to the high-speed network links that connect two or more data centres. For AI, these are not just passive data pipes; they are active components of a single, distributed supercomputer. If a training cluster is spread across multiple facilities, the DCI network must provide seamless, high-bandwidth, and low-latency connectivity to ensure all processing units remain synchronised. A bottleneck in the DCI can leave billions of pounds worth of GPU hardware sitting idle, negating any computational advantage. Source

The scale of data involved is staggering. A single training run for a frontier model can involve moving petabytes of data. This traffic flows not only between server racks but often across entire metropolitan areas or even countries to leverage available compute and energy resources. As models continue to grow in size and complexity, the demand for DCI bandwidth is forecast to grow exponentially, pushing the limits of current optical fibre capacity and creating a significant engineering challenge for network providers and data centre operators alike. Source

Ciena's Coherent Optics: Meeting the Bandwidth Challenge

At the forefront of addressing this bandwidth crisis are innovations in optical networking, led by companies such as Ciena. The key technology is coherent optics, which transforms how data is encoded onto light waves. Unlike simpler on-off keying, coherent detection uses properties of light like phase, amplitude, and polarisation to embed significantly more data onto a single wavelength. This allows for a dramatic increase in the information-carrying capacity of a single strand of optical fibre, moving from 10 gigabits per second (Gbps) to 800 Gbps and beyond on a single channel. This is a crucial topic that has major implications for future enterprise strategies. Source

A pivotal development in this space is the miniaturisation of this complex technology into "pluggable" form factors. Ciena's WaveLogic series of coherent pluggable optics, for example, enables data centre operators to upgrade their network capacity incrementally. These small modules can be directly inserted into routers and switches, allowing for a pay-as-you-grow model to 400G, 800G, and soon 1.6T (terabit) speeds. This approach avoids the prohibitively expensive and disruptive need to replace entire chassis-based transport systems, providing a pragmatic path to scaling infrastructure in line with AI-driven demand. The future of AI infrastructure is a core theme of the upcoming AI World Congress 2026.

This technological leap also directly addresses the critical issues of cost and sustainability in AI. By transmitting more data per wavelength, coherent optics reduce the cost-per-bit transported. More importantly, newer generations of these optics are significantly more power-efficient. Reducing the watts-per-gigabit is a primary goal for hyperscalers and data centre operators, who face immense pressure from both energy costs and environmental, social, and governance (ESG) mandates to build more sustainable AI infrastructure. Source

Beyond Speed: Latency and Jitter in AI Networks

While raw bandwidth is the most cited metric, it is far from the only one that matters for AI workloads. The performance of large-scale, synchronous parallel training is acutely sensitive to latency, the time it takes for a packet of data to travel from one point to another. In a distributed training job, the slowest communication link between any two GPUs can dictate the pace of the entire computation, as the system must wait for all parameter updates to be exchanged before proceeding to the next step. Source

Equally disruptive is jitter, which is the variation in latency over time. An unpredictable network with high jitter means that data packets arrive at inconsistent intervals. This forces receiving processors to buffer more data and can cause GPUs to stall periodically, waiting for the last packet of a data block to arrive. This "tail latency" problem significantly reduces overall cluster efficiency, turning a theoretical petaflop-scale machine into a far less effective one in practice. Eliminating jitter is therefore a primary design goal for AI network fabrics.

Modern optical transport networks are engineered to address precisely these challenges. By using technologies like wavelength-division multiplexing (WDM), optical networks provide dedicated, deterministic light paths between endpoints. This creates a highly stable, ultra-low latency connection that is immune to the congestion and queuing delays that can plague traditional packet-switched electronic networks. Understanding the intricate relationship between network performance and model training efficiency will be a key focus on the Day 1 and Day 2 agenda in London this June.

Architecting the AI Data Centre: Spine-Leaf vs. Optical Switching

For years, the "spine-leaf" architecture has been the gold standard for cloud data centre networking. This design uses multiple layers of electronic switches to provide high-bandwidth connectivity between any two servers. However, it faces challenges with the unique traffic patterns of large AI clusters. The "all-to-all" communication required for many distributed training algorithms can overwhelm the spine switches, creating bottlenecks and limiting the effective size of an AI cluster. The sheer density of cabling required to scale these architectures also presents physical and financial constraints.

An emerging approach that complements traditional designs is the use of an optical circuit switch (OCS) to form an underlying optical foundation. An OCS, also known as a reconfigurable optical add-drop multiplexer (ROADM) in the wide-area network context, can create direct, transparent, all-optical connections between racks of GPUs. This allows for massive, multi-terabit datasets to be moved directly between compute clusters, bypassing the electrical spine layer entirely. This is a key area ripe for exhibition and sponsorship from leading infrastructure firms.

Industry leaders like Ciena are innovating with hybrid electro-optical architectures that combine the best of both worlds. In such a model, the traditional electronic network handles the small, latency-sensitive control messages and general-purpose traffic, while the optical layer is dynamically provisioned to handle theelephant flows of bulk data transfers inherent in AI training. This creates a far more scalable, power-efficient, and cost-effective design tailored specifically for the demands of artificial intelligence, a vision that moves beyond simply adding more switches and fibres. Source

Software and Management: The Intelligence Layer

The advanced optical hardware powering AI infrastructure is only as effective as the software that manages it. These are not static, set-and-forget networks. To support dynamic AI workloads, the underlying optical layer must be programmable, observable, and adaptable. This requires a sophisticated software control plane that can intelligently orchestrate network resources based on the real-time demands of the applications running on top. Source

A critical component of this management layer is telemetry. Modern coherent optics are packed with sensors that provide a continuous stream of data on the performance of the optical link, including metrics like optical signal-to-noise ratio (OSNR), chromatic dispersion, and error rates. This telemetry data can be fed into AI-powered analytics platforms to predict potential link failures before they occur, automatically re-route traffic to avoid degradation, and optimise the network for performance and energy efficiency. Many AI World Congress 2026 speakers are at the cutting edge of applying AI to network management.

The Road to Terabit Ethernet and the Future of AI Infrastructure

The relentless pace of AI development guarantees that the demand for network bandwidth will not slow down. The industry is already on a clear path towards 1.6 Tbps and 3.2 Tbps Ethernet speeds, which will require corresponding advances in coherent optical technology to transport these signals over meaningful distances. This involves developing new materials, more sophisticated digital signal processing (DSP) chips, and advanced integration techniques to continue driving up performance while managing power consumption and physical footprint. Source

Furthermore, the physical architecture of AI systems continues to evolve. The concept of disaggregation, where resources like compute, memory, and storage are pooled and connected via a high-speed optical fabric, is gaining traction. This would allow for the dynamic composition of servers tailored to specific workloads, but it would place even greater demands on the network to act as a seamless, server-scale backplane. These future architectures are entirely dependent on continued innovation in optical interconnects.

Ultimately, the relationship between AI and optical networking is symbiotic. Advances in AI are creating an insatiable demand for more capable networks, while advances in optical networking are enabling the next generation of larger, more powerful AI models. This co-evolution will be a defining feature of the tech landscape for the next decade, with significant investment flowing into the foundational infrastructure that underpins the AI economy. Industry leaders and policymakers must collaborate to ensure this growth is sustainable and robust, and those looking to keep pace can find more AI news on our portal. Source

Frequently Asked Questions

What are optical networks?

An optical network is a type of data communication network built with optical fibre technology. It uses light signals, carried in hair-thin glass fibres, to transmit data over long distances at extremely high speeds. They form the backbone of the internet and modern data centre interconnects.

Why are optical networks critical for AI?

AI, particularly the training of large models, requires moving massive datasets between thousands of processors. Optical networks provide the ultra-high bandwidth (hundreds of gigabits or even terabits per second) and low latency needed to keep these processors synchronised and fed with data, which is essential for efficient AI computation.

What is a coherent pluggable optic?

A coherent pluggable optic is a small, hot-swappable module that can be inserted into network equipment like routers and switches. It contains advanced technology to encode and decode vast amounts of data onto light waves, enabling significant bandwidth upgrades without replacing the entire networking system.

How do optical networks help with sustainability in AI?

Newer generations of coherent optical technology are designed for greater power efficiency, reducing the amount of electricity (watts) required to transmit each gigabit of data. By moving more data with less power, optical networks help mitigate the large energy footprint associated with large-scale AI data centres.

Is upgrading to optical networks expensive?

While the initial technology is advanced, modern approaches like pluggable optics allow for a more cost-effective, incremental upgrade path. This avoids the massive capital expenditure of a full "rip-and-replace" upgrade. The increased efficiency and capacity often result in a lower total cost of ownership over time, especially for bandwidth-intensive AI applications.

Bibliography

1. McKinsey & Company. "QuantumBlack, AI by McKinsey". https://www.mckinsey.com/capabilities/quantumblack
2. MIT Technology Review. "Artificial Intelligence". https://www.technologyreview.com/topic/artificial-intelligence/
3. Stanford University Human-Centered Artificial Intelligence. "Research". https://hai.stanford.edu/research
4. Boston Consulting Group. "Artificial Intelligence & Generative AI Consulting". https://www.bcg.com/capabilities/artificial-intelligence
5. World Economic Forum. "Artificial Intelligence". https://www.weforum.org/agenda/archive/artificial-intelligence/
6. Nature. "Machine Learning". https://www.nature.com/subjects/machine-learning
7. Gartner. "Articles". https://www.gartner.com/en/articles
8. IBM. "Think Research and Insights". https://www.ibm.com/think/insights
9. Financial Times. "Artificial Intelligence". https://www.ft.com/artificial-intelligence
10. GOV.UK. "AI regulation: a pro-innovation approach". https://www.gov.uk/government/publications/ai-regulation-a-pro-innovation-approach

The intricate dance between network engineering and artificial intelligence will continue to define the next generation of technology. To connect with the experts shaping this future and explore these topics in greater depth, be sure to register for the AI conference London, taking place this June.