A UK startup claims it can reduce the power consumed by AI data centre networks by 81 per cent by replacing conventional electronic switching equipment with technology that routes data using light.
Why AI's Energy Problem Is Growing
The rapid growth of artificial intelligence is creating a major sustainability challenge. As AI models become larger and more widely used, the data centres that power them are consuming increasing amounts of electricity. Industry forecasts suggest global data centre energy demand could rise significantly over the coming decade, driven largely by AI training and inference workloads.
Much of the attention has focused on the energy consumed by powerful processors such as GPUs. However, another important source of energy consumption sits in the networks that connect those processors together.
Modern AI systems rely on thousands of chips constantly exchanging data. Every time information moves through conventional networking equipment, energy is consumed and heat is generated. As AI clusters grow larger, those networking systems are becoming increasingly expensive to power and cool.
That has prompted researchers and technology companies to look for ways of making AI infrastructure more efficient.
What Oriole Networks Has Developed
London-based startup Oriole Networks believes it has found one possible solution.
The company has developed a networking platform called PRISM that replaces traditional electronic switches in data centre networks with optical circuits that route information as photons rather than electrical signals.
For decades, data centre networks have depended on electrical switching technology. While highly effective, these systems consume significant amounts of energy and generate large quantities of heat.
Oriole argues that by allowing data to travel directly as light, much of that inefficiency can be removed.
According to the company, PRISM "removes the need for electronic switches entirely" within the network core and replaces them with "nanosecond-switched optical circuits".
The company claims this can reduce core network power consumption by 81 per cent. It also says GPU idle time can fall from around 60 per cent to less than 1 per cent because processors spend less time waiting for information to move through the network.
Why Energy Savings Matter
The sustainability implications extend beyond electricity consumption alone. For example, networking equipment generates heat, and removing that heat requires cooling systems. Cooling can account for a substantial proportion of overall data centre energy consumption and often involves significant water usage as well.
Reducing the amount of heat produced inside a facility can therefore create multiple environmental benefits simultaneously.
Oriole argues that its technology could help reduce cooling requirements while making better use of existing AI hardware. Rather than building more data centres or adding more processors to achieve higher performance, operators may be able to extract more useful work from the infrastructure they already have.
The company also believes its approach could reduce dependence on some of the complex supply chains associated with today's networking equipment.
Moving Into Real-World Testing
The technology is now moving beyond the laboratory. Oriole has announced that its system will be deployed as part of the UK's £50 million ARIA Scaling Inference Lab, a government-backed initiative designed to address performance and efficiency bottlenecks in large-scale AI infrastructure.
The deployment combines Oriole's networking technology with AMD Instinct GPUs and AMD EPYC processors.
Madhu Rangarajan, corporate vice president of Compute and Enterprise AI at AMD, described the technology as "a fundamentally different way to connect accelerators at scale" and said the collaboration is helping validate how photonic networking can provide the connectivity needed for AI inference workloads.
For Oriole, the deployment represents a significant milestone. Chief executive James Regan said: "A year ago, we were proving the physics; today, we're proving the business." He added that the project demonstrates how "photonic networking stops being a research curiosity and starts being the foundation of how serious AI infrastructure gets built."
The Important Caveat
The headline figures remain company claims rather than independently verified industry benchmarks.
The ARIA deployment will provide the first large-scale commercial test of whether the technology can deliver the same benefits under real-world conditions that it has demonstrated during development.
That distinction matters because many promising hardware technologies perform well in controlled environments but struggle when deployed at the enormous scale used by major cloud and AI providers.
The wider rollout planned for 2027 will provide a clearer indication of whether photonic networking can become a practical alternative to conventional data centre infrastructure.
What Does This Mean For Your Organisation?
For organisations concerned about the environmental impact of AI, the story highlights the increasingly important reality that making AI more sustainable is not simply about building better processors.
Attention is increasingly turning towards the wider infrastructure that supports AI, including networking, cooling, power delivery, and resource utilisation.
If technologies such as Oriole's can genuinely reduce network power consumption while improving hardware efficiency, they could help address some of the environmental pressures associated with AI's rapid growth. Lower electricity demand, reduced cooling requirements, and better utilisation of existing hardware would all contribute towards more sustainable AI infrastructure.
Whether Oriole's specific approach succeeds remains to be seen. However, the broader message is clear. As AI energy consumption continues to grow, innovations that reduce waste inside data centres may become just as important as advances in the AI models themselves.