AI Infrastructure • 16 May 2026 • By AI Conference London Editorial

Optical Networks and the AI Infrastructure Boom

Ciena's role in scaling optical networks for AI data centers amid surging bandwidth demands and projected growth by 2026.

Optical Networks and the AI Infrastructure Boom – AI World Congress 2026, London, 23-24 June 2026

The gargantuan models driving the artificial intelligence revolution are built on a foundation of data, but it is the silent, unseen network of light that allows this data to flow. As large language models and generative AI consume computational resources at an exponential rate, the underlying physical infrastructure, particularly optical networking, is facing an unprecedented demand for bandwidth. This surge is forcing a fundamental rethink of data centre architecture, a key theme for the upcoming AI World Congress 2026.

The Unseen Engine: AI's Voracious Appetite for Data

The scale of modern AI is difficult to comprehend. Models like OpenAI's GPT-4 or Google's Gemini involve training on petabytes of data, with trillions of parameters that need to be constantly refined and updated across vast clusters of GPUs. This process creates a deluge of data traffic. During the training phase, huge datasets are moved into the processing cluster, and immense volumes of intermediate data, known as gradients, are exchanged between thousands of interconnected processors. This is not a one-time event; inference, the process of using a trained model to generate a response, also requires high-speed data access, especially for real-time applications that serve millions of users. The sheer volume and velocity of this data render traditional networking solutions increasingly inadequate, creating bottlenecks that can throttle the performance of multi-million-pound AI investments. Source

Data Centre Architecture Under Strain

For years, data centre network traffic was dominated by a "north-south" pattern, where data flowed from users on the internet, into the data centre for processing, and back out again. Architectures were optimised for this flow. However, AI workloads have triggered an explosion in "east-west" traffic—communication that occurs entirely within the data centre, between servers and processing units. This server-to-server communication, essential for distributed training and data synchronisation across GPU clusters, now constitutes the majority of traffic in an AI-focused facility. The need to connect these sprawling clusters, which may span multiple buildings or even different campuses, has elevated the importance of high-capacity Data Centre Interconnect (DCI) solutions. Source

This architectural shift places immense pressure on the physical network layer. Traditional electrical interconnects begin to falter over longer distances and at higher speeds, suffering from signal degradation and high power consumption. As clusters grow from hundreds of GPUs to tens of thousands, the requirement is for a fabric that can provide low-latency, high-bandwidth connections across distances of several metres to multiple kilometres. This is the domain where optical networking excels, using light to transmit data over fibre optic cables at speeds and efficiencies that copper wiring cannot match. The challenge is no longer just about connecting a data centre to the outside world, but about creating a seamless, high-performance network that functions as a single, cohesive computational unit. Source

Ciena's Vision: The Coherent Optical Revolution

Leading optical networking firms like Ciena are at the forefront of addressing this bandwidth crunch. Their strategy centres on the advancement and application of coherent optical technology. Unlike earlier "direct detect" methods that only modulated the intensity of light, coherent optics manipulate multiple properties of the light wave—its amplitude, phase, and polarisation—to encode significantly more data onto a single wavelength. This is achieved through sophisticated Digital Signal Processors (DSPs) which precisely modulate the light at the transmitter and decode it at the receiver, correcting for impairments that occur along the fibre. It is analogous to the difference between a simple torch flashing on and off versus an advanced radio signal carrying complex information. Source

Ciena's WaveLogic series of coherent DSPs exemplifies this technological leap. Each generation has delivered substantial increases in data rate per wavelength, pushing from 100 gigabits per second (Gbps) to 400Gbps, 800Gbps, and beyond. For AI data centres, this translates into a massive increase in bandwidth density. A single fibre pair can now carry tens of terabits of data, enabling the construction of powerful east-west interconnects and scalable DCI links. This capability allows data centre operators to connect geographically distributed GPU clusters as if they were in the same room, ensuring that expensive processing units are not left idle waiting for data. These technical advancements are driven by insights often shared by confirmed AI World Congress 2026 speakers, who are witnessing the direct impact on model development. Source

East-West Traffic: The New Data Superhighway

To fully appreciate the role of optical networking, one must understand the unique demands of distributed AI training. A large model is broken down and distributed across thousands of GPUs, each working on a small piece of the puzzle. For the model to learn effectively, these GPUs must constantly exchange updates and synchronise their state in a process called "all-reduce". This creates a complex and intense pattern of east-west communication. The efficiency of this communication is paramount; any latency or bandwidth limitation in the network fabric directly increases the total training time, which can run into weeks or months and cost millions of pounds in compute time. Optical interconnects, both within the rack (co-packaged optics) and between rows and buildings (pluggable coherent optics), are therefore becoming the default choice for building these high-performance AI fabrics. Source

Power, Cooling, and the Sustainability Imperative

The AI infrastructure boom carries a significant environmental cost. The power consumption of data centres, driven largely by GPU clusters and their associated cooling systems, is a growing global concern. A single high-density AI rack can consume over 100 kilowatts of power, more than an entire row of traditional server racks. In this context, network efficiency is not just about performance, but also about sustainability. Here, optical solutions offer a compelling advantage. Transmitting data as photons over fibre is inherently more energy-efficient than sending electrons over copper, especially at high speeds and over distance. The metric of 'bits per watt' has become a critical benchmark for infrastructure components. Source

Advancements in coherent optics contribute directly to improving this metric. By packing more data onto each wavelength, technologies like Ciena's WaveLogic reduce the number of transceivers, fibres, and supporting hardware needed to achieve a given bandwidth, leading to a smaller physical footprint and lower overall power consumption. This synergy between performance and power efficiency is vital for the sustainable scaling of AI. The economic and environmental implications of data centre energy use will be scrutinised in detail, with this topic being central to the infrastructure track on the Day 1 and Day 2 agenda. Source

The Road to 800G, 1.6T, and Beyond

The relentless growth in AI model complexity means the demand for bandwidth is not static. While 400Gbps (400G) connections are becoming common in modern AI clusters, the industry is already transitioning to 800G. This move is critical for feeding data to next-generation GPUs and preventing network bottlenecks. Major cloud providers and hyperscalers are aggressively rolling out 800G optical interconnects to support their expanding AI services. Looking further ahead, the industry roadmap includes 1.6 terabit per second (1.6T) and even 3.2T pluggable optical modules. These future standards will be essential for training the foundation models of tomorrow, which are expected to be orders of magnitude larger than today's state-of-the-art. Achieving these speeds requires continuous innovation in DSPs, silicon photonics, and packaging, offering a chance to explore these solutions firsthand through the event's exhibition and sponsorship programme. Source

Strategic Implications for Enterprise and National AI Policy

The development of a robust optical networking infrastructure is no longer a niche technical concern; it is a matter of strategic national importance. The ability to build and operate large-scale AI factories is becoming a key determinant of economic competitiveness and technological sovereignty. Nations and corporations that lack access to this advanced infrastructure risk falling behind in the global AI race. This has prompted governments to consider AI infrastructure as a critical asset, with regulations like the UK's pro-innovation approach to AI and the EU AI Act directly shaping the environment in which these technologies are deployed. For enterprises, the choice of data centre and network provider is now a strategic decision that directly impacts their ability to leverage AI for competitive advantage. Source

Frequently Asked Questions

What is an optical network?

An optical network is a communication system that uses signals encoded onto light to transmit information. Data is transported through thin, flexible strands of glass or plastic called optical fibres. It offers significantly higher bandwidth, lower latency, and greater distance capabilities compared to traditional networks based on electrical signals over copper wires.

Why is 'east-west' traffic important for AI?

East-west traffic refers to data communication between servers within a data centre, as opposed to 'north-south' traffic which flows in and out of the data centre. It is critical for AI because modern AI models are trained on distributed clusters of thousands of GPUs which must constantly synchronise with each other. This high-volume, inter-processor communication constitutes the majority of network traffic in an AI cluster.

What is coherent optics?

Coherent optics is an advanced method for transmitting data over optical fibres. It works by modulating multiple properties of a light wave—its amplitude, phase, and polarisation—to encode more bits of information per symbol. Using a sophisticated Digital Signal Processor (DSP), it dramatically increases the data-carrying capacity and reach of a single optical fibre compared to simpler, older technologies.

How does improved optical networking affect AI sustainability?

Improved optical networks enhance AI sustainability by increasing energy efficiency. Transmitting data as light (photons) consumes less power than as electricity (electrons), especially at high speeds and over long distances. Coherent optics increase the data rate per fibre, reducing the amount of hardware, rack space, and overall power needed for the network, thus lowering the carbon footprint of the AI data centre.

What is the significance of 800G networking for AI?

800G (800 gigabits per second) networking represents the next-generation speed for data centre interconnects. For AI, this massive bandwidth is crucial for preventing data bottlenecks and ensuring that powerful GPU clusters are not left idle waiting for data. It enables faster model training times and supports the ever-increasing scale of foundation models, directly impacting the performance and economic viability of AI operations.

Bibliography

  1. McKinsey & Company. "QuantumBlack, AI by McKinsey." https://www.mckinsey.com/capabilities/quantumblack
  2. Gartner, Inc. "Gartner for Technology Leaders." https://www.gartner.com/en/articles
  3. World Economic Forum. "Artificial Intelligence and Machine Learning." https://www.weforum.org/agenda/archive/artificial-intelligence/
  4. OECD. "OECD.AI Policy Observatory." https://www.oecd.org/digital/artificial-intelligence/
  5. Stanford University. "Stanford Institute for Human-Centered Artificial Intelligence (HAI)." https://hai.stanford.edu/research
  6. Nature. "Machine Learning Subject." https://www.nature.com/subjects/machine-learning
  7. MIT Technology Review. "Artificial Intelligence." https://www.technologyreview.com/topic/artificial-intelligence/
  8. Boston Consulting Group. "Artificial Intelligence Consulting." https://www.bcg.com/capabilities/artificial-intelligence
  9. Deloitte. "The state of generative AI in the enterprise: Now decides next." https://www.deloitte.com/global/en/issues/trust/state-of-generative-ai-in-the-enterprise.html
  10. UK Government. "A pro-innovation approach to AI regulation." https://www.gov.uk/government/publications/ai-regulation-a-pro-innovation-approach

The deep symbiosis between AI advancement and network infrastructure will continue to define the technological landscape. To explore these critical topics with industry pioneers and technical experts, secure your place and register for the AI conference London, taking place this June.