17 June 2026

 

 

For the past three years, the artificial intelligence industry has focused on a familiar challenge: building faster chips.

NVIDIA became one of the world's most valuable companies by supplying the graphics processors that power large language models. AMD, Intel and a growing number of specialist chipmakers have raced to develop increasingly powerful AI accelerators. Governments have launched billion-pound infrastructure programmes designed to secure access to computing capacity.

Yet a growing number of engineers believe the next major AI bottleneck may lie elsewhere.

As AI systems become larger and more distributed, moving data efficiently between processors is becoming almost as important as the processors themselves. The challenge is particularly acute for AI inference workloads, where thousands of accelerators must continuously exchange information while serving users in real time.

That is the problem London-based Oriole Networks wants to solve.

In June 2026, the company announced its first commercial deployment through a collaboration with AMD and the UK's Advanced Research and Invention Agency (ARIA) Scaling Inference Lab. The deployment combines Oriole's PRISM photonic networking platform with AMD Instinct GPUs and AMD EPYC CPUs in what the company describes as the world's first large-scale AI system powered by a pure photonic network.

For Oriole, the project marks a significant milestone. For the wider AI industry, it highlights a growing shift in attention towards the networking infrastructure that underpins modern AI systems.

 

From university research to commercial deployment

 

Oriole's story began long before generative AI became a global phenomenon.

The company was founded in 2023 as a spinout from University College London by Professor George Zervas, James Regan, Alessandro Ottino and Joshua Benjamin. Its core technology is rooted in nearly two decades of optical networking research led by Zervas.

This is an interesting heritage: many companies have repositioned existing technologies around the AI boom. Oriole, by contrast, was already working on a problem that AI has since brought into sharp focus: how to move ever-larger volumes of data through increasingly complex computing systems.

The company also combines academic expertise with deep industry experience.

Before joining Oriole, chief executive James Regan served as CEO of EFFECT Photonics, one of Europe's best-known integrated photonics companies. During his time there, he worked on commercialising photonic technologies for telecommunications and data communications markets, experience that is directly relevant to Oriole's current challenge.

Building a photonic technology is one thing. Convincing data centre operators to deploy it at scale is another.

That distinction was reflected in Regan's comments accompanying the announcement.

"A year ago, we were proving the physics; today, we're proving the business," he said.

The statement captures a challenge that has faced the photonics sector for years. The science is often compelling. Commercial deployment is far harder.

The AMD collaboration represents Oriole's first opportunity to demonstrate that its technology can operate in a real-world AI infrastructure environment rather than a laboratory setting.

 

Why networking is becoming AI's next bottleneck

 

The timing may be advantageous.

Training and running modern AI models requires vast quantities of data to move between processors, memory and storage systems. As AI clusters expand, the networking infrastructure connecting those components becomes increasingly critical.

Today's AI data centres rely heavily on electronic switches. Every time information passes through those switches it incurs latency, consumes power and generates heat.

As workloads scale, those inefficiencies accumulate.

Industry-wide estimates vary, but infrastructure specialists increasingly recognise networking as one of the major constraints on AI system performance. Simply adding more accelerators does not automatically increase efficiency if those accelerators spend significant amounts of time waiting for data.

According to Oriole, GPU idle time in conventional AI environments can approach 60%. The company claims its photonic networking architecture can reduce that figure to less than 1%.

Those figures are company-reported metrics and have not yet been independently validated at large commercial scale. However, they illustrate the magnitude of the problem Oriole is attempting to address.

For AI infrastructure operators, improving utilisation may become just as important as increasing compute capacity.

 

Reimagining the AI network

 

At the centre of Oriole's strategy is PRISM, an AI networking platform designed to route data as photons rather than electrical signals.

Instead of relying on conventional electronic switching infrastructure, PRISM uses nanosecond-scale optical circuit switching.

The goal is straightforward: reduce latency, improve throughput and lower energy consumption across the entire system.

According to the company, PRISM can reduce network core power consumption by 81% by eliminating electronic switches from the network core. Oriole also claims the architecture can increase AI inference throughput and interactivity by an order of magnitude, allowing more users to be served simultaneously from the same hardware.

If those gains prove achievable in production environments, the implications could extend beyond performance alone.

AI infrastructure operators are increasingly constrained by energy availability, cooling requirements and operating costs. Reducing the amount of energy required to move data could improve the economics of AI deployment while easing pressure on data centre infrastructure.

The company also argues that a photonic approach could reduce dependence on some of the complex hardware supply chains underpinning today's networking equipment.

Whether those benefits translate into large-scale deployment remains to be seen. That is precisely what the ARIA testbed is intended to evaluate.

 

Competing in a growing photonics race

 

Oriole is not alone in believing that light will play a larger role in future AI infrastructure.

Over the past several years, photonics has become one of the most closely watched areas of deeptech investment.

US-based Lightmatter has developed photonic computing and interconnect technologies designed to accelerate AI workloads. Ayar Labs focuses on optical input-output systems that replace traditional electrical connections between chips. Celestial AI has attracted significant funding for its Photonic Fabric platform, which aims to improve communication between AI compute and memory systems.

Collectively, these companies reflect a broader industry belief that future AI scaling will require innovations beyond compute alone.

Yet Oriole occupies a somewhat different position within the emerging ecosystem.

Where several photonics companies focus on chip-to-chip communication or memory bandwidth, Oriole is targeting the network layer itself. Its objective is not simply to improve individual connections but to redesign how large AI clusters exchange information.

That focus could become increasingly important as AI infrastructure continues to expand.

 

Why AMD matters

 

For a startup operating deep within the infrastructure stack, partnerships matter.

The involvement of AMD provides Oriole with validation from one of the world's largest semiconductor companies.

"AMD is excited to collaborate with Oriole on the ARIA Scaling Inference Lab cluster," said Madhu Rangarajan, Corporate Vice President, Compute and Enterprise AI Business at AMD.

"Oriole's AI backend networking with nanosecond optical circuit switching represents a fundamentally different way to connect accelerators at scale. We are helping to validate how photonic fabrics can work alongside AMD compute to deliver the low-latency, high-bandwidth connectivity that AI inference workloads demand."

AMD's participation is significant for another reason.

The company has emerged as one of the leading challengers to NVIDIA in AI infrastructure. As competition intensifies, improving system-level efficiency becomes increasingly important. The networking layer represents one potential source of future performance gains.

For Oriole, access to AMD hardware and engineering expertise offers an opportunity to test its architecture under realistic AI inference workloads.

 

The role of ARIA

 

The deployment is taking place within the UK's ARIA Scaling Inference Lab, a £50 million programme designed to investigate new approaches to AI infrastructure.

ARIA was established through an Act of Parliament and is sponsored by the Department for Science, Innovation and Technology. The agency was created to fund high-risk, high-reward research capable of generating long-term economic and technological benefits.

The Scaling Inference Lab focuses on a growing challenge within artificial intelligence: how to make AI systems more capable and efficient during inference, rather than relying solely on larger models and additional compute.

"Meeting the demands for modern AI requires rapidly identifying ways to improve the performance and cost-efficiency of large-scale AI clusters," said Suraj Bramhavar, Programme Director at ARIA.

For Oriole, the programme provides a pathway from research into deployment. For ARIA, it represents exactly the type of collaboration between emerging deeptech companies and established industry players that the organisation was designed to support.

 

A European strength hiding in plain sight

 

Oriole's emergence also highlights an area where Europe has developed significant expertise.

While public attention often focuses on American foundation models and semiconductor giants, Europe has built a strong position in photonics research and commercialisation. Universities, research institutes and startups across the UK, Germany, France and the Netherlands have spent decades advancing optical communications technologies.

AI may now provide a new commercial opportunity for that expertise.

Oriole is a product of that ecosystem: a university spinout commercialising long-term photonics research, led by executives with experience bringing optical technologies to market, and supported through a government-backed innovation programme.

It is a reminder that not all important AI innovation happens at the model layer.

Some of it happens much deeper in the technology stack.

 

The next phase of AI infrastructure

 

Oriole still needs to prove that its performance gains can be replicated beyond a government-backed testbed and across commercial AI infrastructure. It also faces competition from well-funded rivals pursuing alternative photonics strategies.

Yet the significance of this deployment lies in what it reveals about the direction of the industry.

For years, AI competition centred on building larger models and more powerful chips. Increasingly, attention is shifting towards the infrastructure that connects those systems together.

If that trend continues, companies such as Oriole may find themselves addressing one of the most important questions in artificial intelligence: not how quickly processors can compute, but how efficiently they can communicate.

The future of AI may not be determined solely by who builds the most powerful accelerator.

It may also depend on who can move information through those systems at the speed of light.