14JUN

Oriole moves data with light, not copper

3 min read

16:35UTC

Oriole Networks deployed what it calls the world's first large-scale photonic AI network at London Tech Week, claiming an 81% cut in the energy data centres spend shuffling data.

← Back to London startup raises Britain's own AI model Jump to analysis ↓

TechnologyDeveloping

Key takeaway

Oriole's photonic interconnect claims to cut data-centre transmission energy 81%, easing the grid limit on UK compute.

Oriole Networks, a London company, deployed what it calls the world's first large-scale photonic AI network at London Tech Week, moving data between chips with light rather than electrical signals ¹. It ran the deployment through ARIA (the Advanced Research and Invention Agency) and its Scaling Inference Lab, partnered with AMD, and claimed an 81% cut in the energy data centres spend shuffling data over conventional copper interconnects ².

The physics explains the figure. In a large AI cluster, most of the power goes not into computation but into pushing data between thousands of chips down copper wires that heat up and slow over distance. Swap the copper for optical links and that transmission cost collapses. The 81% claim is Oriole's own and has not been independently benchmarked at production scale.

That figure, if it holds, speaks directly to the saturated London data-centre grid Lowdown flagged in April , where power is the binding constraint on adding compute. An operator could fit more compute under the same fixed power envelope. Oriole is among the British firms named in the state hardware plan announced that week, with AMD here as a foreign partner to a UK deployment, nothing more.

Deep Analysis

In plain English

Inside AI data centres, thousands of computer chips need to exchange data constantly. Until now, that communication has happened through copper cables, which generate heat and use significant amounts of electricity. Oriole Networks has built a system that replaces those copper links with fibre-optic cables, moving data as pulses of light rather than electrical current. Light uses far less energy for the same job. At London Tech Week in June 2026, Oriole demonstrated this technology working at scale for the first time anywhere, in partnership with AMD (the chip company) and funded through ARIA, the UK's government blue-sky research agency. The company claims an 81% reduction in energy used for data transmission. That matters because London's data-centre grid is already full, and operators cannot get permission to use more electricity. A technology that cuts the energy cost of moving data inside a building means more AI computing power from the same power supply.

Deep Analysis

Root Causes

Two infrastructure-level constraints converged in 2026 to create a market pull for photonic interconnects that did not exist two years earlier, and they explain why Oriole chose London Tech Week as the deployment date rather than a private demonstration.

London's data-centre grid reached saturation in the Slough and West London corridor in April 2026, with 35 data centres already operational and National Grid refusing new large-load connections. For AMD and any GPU cluster operator, photonic interconnect now offers the only mechanism to increase compute density within a fixed power envelope: if you cannot get more electricity to the building, you must extract more compute from the electricity you have.

AMD's partnership with Oriole, rather than NVIDIA, reflects AMD's specific competitive position. NVIDIA's NVLink interconnect is copper-based and proprietary, locking customers into NVIDIA's own ecosystem. AMD's ROCm software stack is chip-agnostic by design, making AMD the natural partner for a photonic interconnect company that needs to work across multiple GPU architectures.

Source Landscape

This story draws on neutral-leaning sources

The 81% energy reduction in data transmission, if confirmed at production scale, changes the amortised compute economics for London-based AI inference operators in a specific way: the cost of inference per token is approximately 60-70% power cost and 30-40% hardware depreciation at current UK electricity prices of roughly £0.18-0.22 per kWh commercial.

An 81% reduction in transmission energy reduces total inference power consumption by approximately 25-30% (transmission is roughly one third of total cluster power at inference scale), which at London electricity prices is worth approximately £0.04-0.06 per GPU-hour in saved power costs.

At 10,000 GPU-hours per day, that saving is £400-600 per day per cluster, or roughly £150,000-220,000 per year. At that magnitude, the savings justify photonic interconnect in new cluster builds, where copper is not already installed, but fall short of the OpEx threshold needed to justify ripping out an existing copper infrastructure mid-lease.

Consensus view: Research from the European Photonics Industry Consortium (EPIC) and the IEEE Photonics Society's 2025 interconnects roadmap both place silicon photonics-based chip-to-chip links at the 2026-2028 deployment window for AI inference clusters, precisely the window Oriole is targeting.

The claimed 81% transmission-energy reduction is physically consistent with published figures for WDM (wavelength-division multiplexing) optical interconnect against NRZ (non-return-to-zero) electrical signalling over 3-metre reach, the typical rack-to-rack distance in GPU clusters.

Counter-view: The University of Southampton's Optoelectronics Research Centre, one of the UK's leading groups in this field, has published work noting that large-scale photonic integration faces yield problems at volume: a single defect in an optical waveguide degrades signal quality across the entire chip-to-chip path in ways that electrical interconnects can tolerate.

The ARIA Scaling Inference Lab deployment is a controlled environment, not a production data centre under operational load variability.

Key tension: Whether Oriole's photonic architecture holds its energy advantage under the thermal and vibration conditions of production data centres, where rack temperatures fluctuate by 15-20 degrees Celsius between day and night cycles.

Sources:The Next Web

Mentions:London Tech Week →ESB Networks →Advanced Research and Invention Agency →AMD →London →Oriole Networks →Scaling Inference Lab →

First Reported In

Update #8 · London startup raises Britain's own AI model

The Next Web· 14 Jun 2026

Read original →

Causes and effects

Caused by

Slough saturates, AI datacentres head north

Oriole's 81% energy cut directly addresses the London data-centre grid saturation Lowdown flagged in April, where power is the binding constraint on adding compute.

Occurred 20 Apr 2026

Read story →

This Event

Oriole moves data with light, not copper

If the 81% figure holds at scale, operators could fit more compute under a fixed power envelope, the binding constraint on every new UK data centre.

Different Perspectives

European VC (Atomico, Plural, Highland Europe as PhysicsX / Lumen adjacents)

European growth funds have backed three of the week's largest UK rounds via follow-on positions and co-investments; the PhysicsX cap table includes Atomico (European-domiciled, Skype-founded) and Siemens (German industrial), both returning investors who view UK physical-AI as a supply-chain multiplier across Continental manufacturing. European LP capital is filling the growth tier UK state vehicles have not yet reached.

UK regulated-industry coalition (Lloyds, BAE Systems, LSEG via Lumen Sovereign)

Thirteen of Britain's most heavily regulated companies backed Cosine not as a philanthropic gesture but to acquire a data-compliant AI tool that replaces costly US API alternatives; each partner provides proprietary data in exchange for early access. Their participation signals that regulated incumbents, not venture funds, may be the structural customer base that sustains the UK's sovereign model tier.

US growth investors (General Catalyst, Intrepid Growth Partners)

US and allied growth investors followed Temasek into PhysicsX's Series C; General Catalyst also returned in the round after backing Geordie the previous week. The absence of any US-led domestic-capital equivalent is a structural reading: American funds enter at growth stage where returns are clearest, ceding seed and Series A economics to UK vehicles that are themselves contracting.

Temasek (Singapore sovereign fund)

Temasek led PhysicsX's $300m Series C, its second major UK deep-tech cheque in six weeks after co-investing in Isomorphic's Series B with the SAIU; its thesis runs through Southeast Asian advanced-manufacturing adjacencies, not bilateral UK policy. Singapore's sovereign capital is now the default lead for British scale-ups above £200m that fall outside the BBB's priority sectors.

UK Government (DSIT / Liz Kendall)

DSIT published its first sector scorecard on 10 June setting a £8.3bn 2025 baseline, and the Sovereign AI Unit's compute allocation enabled Cosine's Lumen Sovereign launch. The scorecard's own barbell figure, more capital in fewer rounds, exposes the policy gap DSIT has not yet addressed: no instrument currently leads venture rounds in industrial AI simulation sectors.

Spanish state finance (COFIDES, CDTI)

Spain's COFIDES and CDTI have co-invested alongside UK deep-tech rounds in prior cycles and track the British Business Bank's direct-investment activity as a benchmark for state-capital deployment in innovation. BBB's two direct co-investments in one week set a pace reference for Iberian equivalents.