The archive · Hardware & Devices · Technical decision · 2023–2026
Oriole Networks bets light, not Ethernet, connects AI GPUs; $22M Series A, AMD deployment
UCL spinout replaces electrical switches with pure-photonic ones to speed AI training and cut data-center power; $35M raised, first deployment with AMD in 2026
Oriole Networks
What the business is
Oriole builds photonic networking hardware and software that connects AI accelerators with light instead of electrical Ethernet switches, for AI data centers and HPC
Starting capital：£10M seed (Mar 2024); £17.5M (~$22M) Series A (Sep 2024), led by Plural
How it started
Founded in 2023 as a University College London spinout by CEO James Regan, who previously built optical-systems company EFFECT Photonics, together with UCL researchers George Zervas, Alessandro Ottino and Joshua Benjamin; the IP was developed at the university. A £10M seed followed in March 2024 from XTX Ventures, Clean Growth Fund, Dorilton Ventures and the UCL Technology Fund.
What happened
Six months later, in September 2024, Oriole closed a £17.5M Series A led by Plural — partner Ian Hogarth joined the board — on claims of up to 100x faster interconnects at 2–3% of Ethernet's energy use. In May 2025 it launched PRISM, a full-stack pure-photonic network (NIC, integrated photonic switch, passive router, software). The company stayed pre-revenue while lining up its first deployment.
How it ended up
In June 2026 Oriole announced its first commercial deployment: a large-scale AI system with AMD pairing Instinct GPUs and EPYC CPUs to test PRISM for inference. CEO James Regan promised more deployments in 2027 and said the company was raising again.
Background
Oriole Networks is a London startup spun out of University College London in 2023 by serial optical-systems entrepreneur James Regan and UCL researchers George Zervas, Alessandro Ottino and Joshua Benjamin. Its premise: the reason large AI training runs hit a wall is not the chips, it is the network — the Ethernet switches connecting thousands of GPUs. Oriole's answer is photonics: move data between accelerators as light, replacing electrical switches in the network core with passive optical routers.
The company raised a £10M seed in March 2024 from XTX Ventures, Clean Growth Fund, Dorilton Ventures and the UCL Technology Fund, then a £17.5M Series A led by Plural six months later, with Ian Hogarth joining the board. The pitch to investors: interconnects up to 100 times faster, energy consumption at 2–3% of Ethernet's, in a market where networks already consume about 20% of an AI cluster's power.
In May 2025 Oriole launched PRISM, a full-stack pure-photonic network combining a network interface card, an integrated photonic switch and a passive, unpowered optical router, plus software that plugs into GPU collective-communication stacks. The company remained pre-revenue while it prepared its first deployment.
In June 2026 Oriole announced its first commercial deployment: a large-scale AI system built with AMD, pairing Instinct GPUs and EPYC CPUs with PRISM and aimed at AI inference. Regan says its switching latency is around 10 nanoseconds versus about 250ns for leading electrical switches, with broader deployments expected in 2027 and a new round under way. Dell'Oro analysts estimate AI back-end networking will be a close to $200B market by 2030.
What has to be true
- The bet rests on a physics claim — keeping signals optical end-to-end beats converting them at every switch — which only large-scale deployments can prove
- Two rounds in six months (£10M seed, £17.5M Series A) funded years of R&D before any revenue, typical of deep tech that must win a standards battle
- The market is enormous if the claim holds: Dell'Oro projects AI back-end networking near $200B by 2030, and power efficiency now matters as much as bandwidth
- Oriole competes against both Ethernet incumbents and photonics rivals such as Google's Apollo project and Ciena, so adoption timing, not just technology, decides the outcome
- Being pre-revenue into 2026 with a first deployment at AMD is the hinge: one failed deployment could stall the company before volume sales begin
What can be applied
Rather than improve the incumbent, remove it: Oriole bets pure-photonic switching replaces Ethernet in AI clusters. The upside is huge if physics holds at scale, but buyers only gamble once
Aftermath
As of September 2026 Oriole is pre-revenue but past its first commercial deployment: PRISM is being tested with AMD on a large-scale inference system in the UK, with broader deployments promised for 2027 and another raise under way. Analysts at Dell'Oro peg AI back-end networking at close to $200B by 2030, with power efficiency as important as bandwidth — the market Oriole is aiming at. The fundamental risk remains: a photonics-first fabric must prove reliability and economics at hyperscaler scale against Ethernet incumbents and rival optical-circuit-switching approaches.
Sources
- Plural backs Oriole Networks in £17.5m round to develop tech to train LLMs 100 times faster
- Photonics startup Oriole Networks raises $22m
- Oriole Networks pushes pure photonic networking for AI data centers
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card