Oriole launches its inaugural fully photonic AI network, claiming an 81% reduction.

      TL;DR: UK startup Oriole Networks is launching the world's first large-scale pure photonic AI network, claiming an 81% reduction in core power usage and under 1% GPU idle time. This system integrates with AMD hardware within the UK’s £50m ARIA Scaling Inference Lab.

      Traditionally, data center networks rely on electrical switches, which are energy-intensive and generate significant heat, causing delays that hinder AI systems from quickly processing and sharing data. Oriole Networks proposes a solution by substituting all electrical switches in the network’s core with optical circuits that transmit data using photons rather than electrons.

      On Monday, Oriole announced it will implement what it calls the world’s first large-scale AI system powered by a pure photonic network, part of the UK’s ARIA Scaling Inference Lab. This setup combines Oriole’s PRISM networking platform with AMD Instinct GPUs and AMD EPYC CPUs, representing the company's first commercial deployment, with plans for broader industry adoption in 2027.

      What PRISM does:

      PRISM entirely removes electronic packet switches from the network core. In standard data centers, electrical switches connect GPUs, causing latency, consuming energy, and generating heat. Oriole replaces them with optical circuit switching at nanosecond speeds, enabling direct transmission of photons between chips.

      The company asserts this leads to an 81% reduction in core network power consumption and decreases GPU idle time from about 60% in current systems to less than 1%, as the network is no longer a limiting factor. Oriole claims this results in a substantial increase in inference throughput, allowing for more tokens processed per second and more users served simultaneously by the same hardware.

      These claims are significant; however, the 81% power cut and sub-1% GPU idle time have not been independently verified at production scale. The ARIA deployment will serve as the initial real-world test to see if lab results can be replicated under commercial conditions.

      The ARIA Scaling Inference Lab:

      The deployment takes place in the ARIA Scaling Inference Lab, a £50 million ($68 million) project funded by the UK government through the Advanced Research and Invention Agency to alleviate bottlenecks in large-scale AI inference. Established by Act of Parliament, ARIA is supported by the Department for Science, Innovation, and Technology. The lab is managed by CommonAI and aims to evaluate and enhance AI systems under practical conditions.

      Inference, the phase in which trained models generate predictions and outputs, comprises the bulk of AI compute costs and energy consumption. It is also where global AI infrastructure development faces the most limitations due to network performance.

      “AMD is thrilled to partner with Oriole on the ARIA Scaling Inference Lab cluster,” remarked Madhu Rangarajan, corporate vice president of compute and enterprise AI at AMD. “Oriole’s AI backend networking utilizing nanosecond optical circuit switching offers a fundamentally new approach to large-scale accelerator connectivity.”

      From R&D to deployment in three years:

      Founded in the UK, Oriole has raised around $35 million from investors such as Plural, UCL Technology Fund, Clean Growth Fund, XTX Ventures, and Dorilton Ventures. The company transitioned from research to commercial application in just three years, a notably rapid timeline for photonic hardware.

      CEO James Regan framed the announcement as a shift from proving theoretical physics to demonstrating commercial viability. “A year ago, we were validating the physics; today, we’re validating the business,” he stated. “This illustrates how photonic networking evolves from a research novelty to a foundational element of serious AI infrastructure.”

      Importantly, PRISM is designed to be chip-agnostic. It can function with any accelerator platform, not just AMD, offering data center operators a way to enhance network performance without locking into a single proprietary stack. The planned broader rollout in 2027 will test whether this agnosticism remains effective across various hardware setups.

      Why it matters:

      AI data center energy use is projected to double by 2030, with cooling alone responsible for roughly 40% of a data center's energy consumption. Networks introduce an additional layer of waste; every electrical switch that connects GPUs consumes energy converting photons to electrons and back, generating heat in the process.

      If PRISM achieves its claims, the implications extend beyond mere power savings. Improved chip-to-chip communication allows for a more efficient use of costly GPU resources, thereby reducing the inference cost per token. Given that enterprises are already facing soaring AI expenses, a network that enables existing hardware to deliver greater output without additional investment in hardware has a clear commercial advantage.

      However, the transition from a government-funded test facility to a commercial hyperscale data center presents challenges. While Oriole’s ARIA deployment is authentic, it still lacks the scale of operations seen in Meta or Google clusters. The 2027 deployment will reveal whether PRISM can successfully scale from a lab backed by £50 million in public funding to the production environments of companies investing vast sums in AI infrastructure