In August 2025, I wrote about Finchetto, a British photonics startup working on an optical packet switch that stores data entirely in the optical domain without bouncing back and forth between light and electronics.
The company’s innovative technology could dramatically accelerate hyperscale networks once AI systems begin to overwhelm current infrastructure. The concept also aims to reduce power consumption while maintaining scalability and increasing communication speed.
What motivated Finchett to focus on photonic packet switching and how does it differ from conventional electronic switching?
With Finchetto, we looked at how networks work today and saw a lot of red tape going on.
The server or GPU often sends data as light, which is then converted to electrons inside the switch so the CPU knows where to go. The light then comes back on to exit the box. This results in energy costs and delays in the round trip.
Then we asked ourselves if we could do this without going back into the electronic domain. To do this, we developed a technology that uses light to control light, so everything changes in the optical domain.
Much of the photonic work seen elsewhere is still circuit switching, which establishes a path between two endpoints using things like MEMS mirrors or thermo-optic devices to transmit light.
The downside is relatively slow recovery and can’t keep up with burst-by-burst solutions at 1.6 or 3.2 Tbps. It’s through packet switching over optics that you get real flexibility and performance, and that’s the gap we’re going to fill.
What benefits do you see in terms of speed, efficiency, and scalability when you use this in large networks?
I’d say speed is the most obvious advantage, but efficiency is just as important. When you store the signal as light instead of converting it from light to electrons and back, you don’t burn as much power as you experience delay.
In terms of scalability, all optical packet switching allows very large and very flexible networks to be built. It can make routing decisions at the packet level to spread the workload more evenly across a large fabric.
Using standard concepts like backbone and leaves, but implemented with our photonic switches, you can access tens of thousands of nodes without the network itself being a bottleneck.
How does this translate into real-world implications for hyperscale data centers in terms of performance and capacity?
Power is now at the top of any hyperscaler’s agenda. Anything that reduces network power consumption without compromising performance will have a positive impact on profits and, by extension, on competition.
Our method eliminates many electro-optical conversions and many transceivers that often fail, resulting in a network that consumes less power and is more robust at the same time.
You can periodically add Finchetto switches to improve performance and energy efficiency over time as you use existing resources. It’s a lot easier than tear down and replace.
What does this mean specifically for new workloads like artificial intelligence and other advanced computing?
AI is a great example of how a network can silently kill its productivity. These training sets require large amounts of data to be transferred between GPUs in a very short time. If you can’t save the fabric, you end up with expensive dead silicone.
By switching packets over ultra-low-latency optics, we eliminate many bottlenecks at the hardware level. It also opens up options that were previously impractical. Some of the more exotic topologies (torus, dragonfly-style architectures, etc.) have historically been difficult to justify because the latency budgets simply don’t work with conventional switching.
When your switch is no longer a limiting factor, network architects can revisit these ideas and choose a topology that works with the hardware that really fits the workload.
How easily can data centers connect Finchetto to what they already have?
This has been one of our core design principles since day one. The reality is that hyperscale data centers are already operating at a level that the market accepts, and a lot of capital has been invested to get there.
No one will say: “Good idea, we will rebuild everything around.” We’ve spent a lot of time making our technology look and feel like a good citizen on today’s web.
It interoperates with existing transceivers, network cards, GPUs and cables and runs on a familiar architecture rather than requiring a complete redesign. This means you can start with specific implementations (a new AI module or a performance-critical piece of fabric) and evolve from there as you see the benefits.
Stepping back a bit, what trends in photonics and networking are you most excited about right now, and what are the main obstacles to mass adoption?
Photonics has evolved from an interesting research focus to a central part of the roadmaps of the most important players in the industry. You can see this in package optics and in the large acquisitions of early-stage photonic companies.
When leaders like Nvidia say, “We need optics along with computing,” the rest of the industry listens. The hard part is building a complete system that traders trust. It must be tightly integrated with the GPU, network card, motherboard, and hardware they already use; It must be reliable throughout its lifetime; And it should be easy to manage and update.
Our answer is to make the optical core as passive as possible and independent of line speed. If you go from 800GB to 1.6TB, you won’t have to replace the middle switch, which is different from replacing an entire layer of electronics every time you increase the speed.
If your switch is purely optical and has no internal buffers, how can you stop packet loss and hotspot collisions?
In a traditional electronic or hybrid switch, you rely on memory and buffering to smooth things over. In a purely optical system, this is not achieved, so you have to think differently.
All we have done is avoid collisions and return to the sender in the optical layer.
A switch can effectively determine whether a given route is free before sending traffic. If not, the packet doesn’t go through, so you can avoid most collisions in the future.
In the rare event that two packets collide, there is a mechanism to send the packet to the sender to retry.
All of this takes place in the optics, which are the complex part and mean that the benefits of an all-optical structure are preserved along with the complexity of the network’s packet-switching functionality.
Focusing specifically on the UK: As the country ramps up investment in AI and data centres, what needs to be done to ensure local photonics and networking technologies are used?
Much of the real innovation in the UK in this area comes from start-ups as there are no major switch suppliers in the country.
The risk is that we spend a lot of public money building an AI infrastructure that is essentially a showcase for foreign suppliers, while British companies that do thorough research and development will never be able to get a foothold.
What will really help is the right support during the scale-up and deployment phase: funded test beds, as seen at Quantum, where new technologies can be tested in realistic settings, and procurement structures that make the inclusion of UK-developed technologies the norm, not the exception.
If we’re serious about “sovereign” power in data centers and AI, we need to go beyond simply hosting other people’s hardware.
Where else do you see networks transforming photonics?
It’s easy to focus on the big data centers because that’s where AI and the cloud live today, but networks are much wider.
Think of communications between satellites in space, optical communication links in free space that provide communication to hard-to-reach areas, or secure, high-bandwidth connections between aircraft or autonomous vehicles in defense.
These are all basically network issues, and they are all places where photonics can have a big impact.
Finally, how do you think the Finchetto architecture will evolve to meet future needs such as quantum networks, optical computing, or photonic memory?
The way we structure our intellectual property is quite thoughtful. Essentially, what we have patented is a method and device for altering data using non-linear optics. In other words, you are not limited to a very limited implementation or use case.
This gives us a lot of room to maneuver. The same basic switching principle can be applied to different types of networks, be they classical high-speed packet networks, future quantum adjacency architectures, or systems in which the computation and memory are themselves optical.
We’re focused on solving the current challenges of AI and hyperscale networking, but we’re doing it with a technology foundation that can keep up with the industry, not get caught when the next wave hits.
Follow TechRadar on Google News I Add us as your preferred source Get news, reviews and opinions from our experts in your feed. Don’t forget to click the follow button!
And, of course, you can Siga TechRadar on TikTok Get our regular updates in the form of news, reviews, unboxing and videos. WhatsApp very
