Optimising Data Centres for AI: The Potential of Optical Interconnects

Introduction

Artificial Intelligence (AI) is poised to transform economies by delivering substantial productivity improvements. According to projections from the International Monetary Fund, successful and widespread adoption of AI in the UK could boost economic productivity by approximately 1.5% per year. Recognising this transformative potential, the UK government has identified AI as a key component of its agenda to boost economic growth and improve public services [1]. However, the widespread adoption of AI will necessitate unparalleled processing power and data storage capabilities.

Increasing demand for AI is driving massive investment in data centre infrastructure. For example, Amazon, a leading player in cloud computing and data storage, plans to invest nearly $150 billion in data centre expansion over the next 15 years, anticipating a significant surge in AI-related demand [2].

Nevertheless, some industry experts warn that the need for data centre capacity may still be vastly underestimated [3]. The growth of data centre development is also becoming increasingly energy intensive. According to Goldman Sachs, global data centre energy demand could rise 160% by 2030, potentially accounting for 3-4% of the world’s total power usage by the end of the decade, up from the current 1-2% [4].

This dual challenge of scaling AI processing infrastructure while minimising energy impact highlights the necessity for technological innovation. Although expanding the number of data centres is unavoidable, constraints related to space, materials, energy, and finances emphasise the importance of enhancing their efficiency. Companies that can innovate to boost the performance of AI data centres, while cutting costs and energy consumption will continue to be well-placed to attract significant attention and investment.

Optical Interconnects: A Promising Innovation

One potential way to address these challenges is the advancement of optical interconnects. Optical interconnects transmit data from one location to another while maintaining minimal transmission errors [5]. This technology leverages light-based signals, typically through fibreoptics or photonic devices, to facilitate data exchange between components, replacing traditional electrical signals carried by copper wiring [6].

Fibreoptic technology is already widely used for long-distance data transmission and often data centres use fibre optics for their external communications networks. However, the racks inside a data centre still predominantly rely on copper-based electrical wires [7].

Inexpensive and reliable, passive copper remains the predominant technology for connections of five metres or less. However, optical connections are favoured because as the connection speeds shift from 50 to 100 Gbps, copper cables cannot reliably send signals at even modest distances. In addition, AI processing chips exchange data at approximately 10 times higher speed than those of general-purpose processors, pushing the limits of copper wiring’s performance [5].

Thicker copper cables could be utilised, but they introduce bulk, cost, installation and management headaches [5]. Ultimately, optical interconnects could offer several advantages for AI, including higher bandwidth, faster data transfer speeds, lower latency, and reduced power consumption [6].

A Surge in Investment

The potential of optical interconnects has spurred significant investment in this space.  Companies such as Lightmatter and Ayar Labs are at the forefront of the development.

Lightmatter recently secured $400 million in a Series D funding round, bringing their valuation to $4.4 billion, an impressive fourfold increase despite only doubling their funding [8]. Similarly, Ayar Labs based in Silicon Valley, announced in December 2024 that it had raised an additional $155 million in new venture capital to accelerate its transition to large-scale production [9]. Ayar Labs received backing from major industry players such as AMD Ventures, Intel Capital, and NVIDIA, which reinforces the strong demand for optical interconnect technology and enhanced efficiency in AI data centres [8].

Conclusion

As the demand for AI-driven solutions continues to rise, so will the need for scalable, efficient, and sustainable processing infrastructure. Addressing the challenges of energy consumption and scalability will drive advancements in data centre design, and create significant opportunities for companies at the forefront of innovation. Technologies that enhance data centre efficiency while addressing these challenges are poised to play a critical role in shaping the future of AI infrastructure.



Written by Scott Kennedy

Edited by Connor Ovenstone

Previous
Previous

A Grounded Perspective on Heathrow’s Proposed Third Runway

Next
Next

Smarter, Faster, Better? How AI Startups are Transforming Industries