A UK startup says it can cut data centre network power by 81% by replacing every electrical switch with light

For decades, the networks inside data centres have run on electrical switches. They are power-hungry, generate enormous heat, and are increasingly the bottleneck that limits how fast AI systems can process and exchange data. Oriole Networks, a UK startup, says it has a fix: replace every electrical switch in the core network with nanosecond-scale optical circuits that route data as photons instead of electrons.

On Monday, Oriole announced that it will deploy what it describes as the world’s first large-scale AI system powered by a pure photonic network, as part of the UK’s ARIA Scaling Inference Lab. The system pairs Oriole’s PRISM networking platform with AMD Instinct GPUs and AMD EPYC CPUs. It marks the company’s first commercial deployment, with wider industry rollout planned for 2027.

What PRISM does

PRISM eliminates electronic packet switches entirely from the network core. In a conventional data centre, electrical switches sit between GPUs and introduce latency, consume power, and generate heat. Oriole replaces them with optical circuit switching at nanosecond speeds, allowing photons to travel directly from chip to chip.

The company claims this cuts core network power consumption by 81%. It also says GPU idle time drops from roughly 60% in current systems to less than 1%, because the network is no longer the constraint. The result, according to Oriole, is an order-of-magnitude increase in inference throughput, meaning more tokens per second and more users served simultaneously from the same hardware.

Those are significant claims. The 81% power reduction and the sub-1% GPU idle time have not been independently benchmarked at production scale. The ARIA deployment will be the first real test of whether lab performance translates to commercial workloads.

The ARIA Scaling Inference Lab

The deployment sits within the ARIA Scaling Inference Lab, a £50 million ($68 million) testbed funded by the UK government through the Advanced Research and Invention Agency to address bottlenecks in large-scale AI inference. ARIA was created by Act of Parliament and is sponsored by the Department for Science, Innovation, and Technology. The lab is hosted by CommonAI and designed to test and optimise AI systems under real-world conditions.

Inference, the operational phase where trained models serve predictions and generate outputs, accounts for the majority of AI compute cost and energy use. It is the phase where the global AI infrastructure buildout is most constrained by network performance.

“AMD is excited to collaborate with Oriole on the ARIA Scaling Inference Lab cluster,” said Madhu Rangarajan, corporate vice president of compute and enterprise AI at AMD. “Oriole’s AI backend networking with nanosecond optical circuit switching represents a fundamentally different way to connect accelerators at scale.”

From R&D to deployment in three years

Oriole was founded in the UK and has raised approximately $35 million to date from investors including Plural, UCL Technology Fund, Clean Growth Fund, XTX Ventures, and Dorilton Ventures. The company went from research to commercial deployment in three years, an unusually fast timeline for photonic hardware.

CEO James Regan framed the announcement as a transition from physics proof to commercial proof. “A year ago, we were proving the physics; today, we’re proving the business,” he said. “This is what it looks like when photonic networking stops being a research curiosity and starts being the foundation of how serious AI infrastructure gets built.”

Crucially, PRISM is designed to be chip-agnostic. It works across any accelerator platform, not just AMD, giving data centre operators a path to improved network performance without committing to a proprietary stack. The wider industry rollout in 2027 will test whether that agnosticism holds at scale across different hardware configurations.

Why it matters

AI data centre energy consumption is projected to double by 2030. Cooling alone accounts for roughly 40% of a data centre’s power use. Networks add another layer of waste: every electrical switch that sits between GPUs burns energy converting photons to electrons and back again, heating the room in the process.

If PRISM delivers on its claims, the implications extend beyond power savings. Faster chip-to-chip communication means more efficient use of expensive GPU capacity, which means lower inference cost per token. In a market where enterprises are already struggling with runaway AI bills, a network that makes existing hardware produce more output without buying more hardware has an obvious commercial case.

The caveat is the distance between a government-funded testbed and a commercial data centre at hyperscale. Oriole’s ARIA deployment is real, but it is not yet operating at the scale of a Meta or Google cluster. The 2027 rollout will determine whether PRISM can survive the jump from a lab backed by £50 million of public money to the production floors of companies spending hundreds of billions on AI infrastructure. That is the gap where most hardware startups fail.