Salience Labs, a startup in Oxford, England, that builds switches to carry data between AI chips as light, has closed an oversubscribed $125M Series B round. Bullhound Capital, a technology investor, led the round, and its principal Callum Stewart is joining the Salience board.
AMD Ventures and Intel Capital, the venture arms of the two chipmakers, also joined the round. So did SBI Group, a Japanese financial group, the trading firm Hudson River Trading and the state-owned British Business Bank. Existing backers Applied Ventures, Oxford Science Enterprises, Cambridge Innovation Capital and Silicon Catalyst Ventures also put in money.
Salience plans to spend it on rolling out its optical circuit switching (OCS) technology, scaling up volume manufacturing and growing its presence in Silicon Valley, including a second office on the West Coast. The money will also go towards future switches with 128 and 256 ports, larger than the 64-port switch it showed in September at the European Conference on Optical Communication (ECOC) in Málaga, Spain.
Vaysh Kewada, the company’s co-founder and chief executive, said Salience has shown “a fully integrated on-chip solution that scales well to high-volume manufacturing”. She said: “OCS will play a critical role in AI datacenter networking.”
What An All-Optical Switch Does
Inside an AI data centre, thousands of graphics processing units (GPUs), the chips that do most of the AI work, swap data all the time, and when links slow down or fail, a whole training job can stall. Much of that data already travels as light through optical fibre. The switches that direct it, though, are mostly electronic.
At each electronic switch, devices called transceivers turn the light into an electrical signal and, once it has been routed, back into light. Every conversion uses power and adds delay. Salience’s switch is built with silicon photonics, which places tiny light-guiding circuits on a silicon chip, and it steers the light from one fibre to another without converting it.
The company says this removes the need for transceivers inside the switch. Its website lists up to 8 times lower power per port, 10 times lower cost per port and 20 times lower network latency, the delay before data arrives, as benefits. These are Salience’s own figures.
Salience says its switches extend the “scale-up domain”, the tightly linked group of GPUs that work together like one large computer. Linking more of them that closely, with no electronic switching in between, is the problem Salience is trying to solve.
Building The Switch At Volume
Salience launched its first product, a 32-port all-optical switch, in March 2026 and said its range also includes 64- and 128-port designs. It said the switch is built to work with the transceivers and equipment already in data centres.
The month before, Salience agreed with Tower Semiconductor, an Israeli chip foundry, to make its switches using Tower’s high-volume silicon photonics processes, with the project entering pre-production. The Series B announcement names Tower again, alongside Keysight Technologies, the American test and measurement company, as firms Salience will keep working with to scale up manufacturing.
In March, Keysight said it was extending its KAI Data Center Builder, a set of tools for testing AI data centre networks, to cover optical circuit switches, working with Salience. The two companies planned a demonstration at the Optical Fiber Communication Conference (OFC) in Los Angeles. The Series B announcement mentions customer relationships but names no customers.
Optical Switching At Scale
Google has already used optical circuit switches at large scale. Its engineers described in a 2023 paper how such switches link the 4,096 tensor processing units (TPUs), Google’s own AI chips, in its TPU v4 supercomputer, so that connections between chips can be rewired as needed. They put the switches and other optics at under 5% of the system’s cost and under 3% of its power.
Patrick Rundell, a director at AMD Ventures, said: “AI infrastructure is increasingly being designed at rack scale, where networking has a direct impact on system performance, efficiency and economics.” Srini Ananth, a managing director at Intel Capital, said Salience stands out for its “integrated silicon photonics architecture and proprietary optical amplification”, its own way of boosting the light signal.
Salience also cites a Dell’Oro Group forecast that cumulative spending on switches for the back-end networks linking AI chips will approach $1T between 2026 and 2030. Bullhound’s Stewart said Kewada has built a team designing a switch architecture that “aligns with customers’ near-term roadmaps” and can keep growing with them.
From Photonic Chips To Switches
Kewada founded Salience in 2021 with Johannes Feldmann, its chief technology officer, building on more than a decade of research at the University of Oxford and the University of Münster in Germany. Feldmann earned a physics PhD focused on photonic computing at Münster and worked as a postdoctoral researcher at Oxford, and he designed and tested the first Salience prototype.
Kewada studied physics at Imperial College London and worked at McKinsey and at Oxford Sciences Innovation, the university investor now called Oxford Science Enterprises and one of Salience’s backers. Feldmann was lead author of a 2021 paper in Nature, written with researchers at Münster, Oxford and elsewhere, showing a photonic chip running the maths behind AI many times in parallel.
That work shaped the company’s first product idea. When Salience raised an $11.5M seed round in 2022, it was building a processor that paired a photonic chip with standard electronics to speed up AI calculations. By the time it raised $30M in a Series A round led by ICM HPQC Fund and Applied Ventures in February 2025, it had moved to optical switches for connecting large AI clusters.
Where Salience Labs Stands Now
Salience Labs now has fresh money from the venture arms of two major chipmakers, financial investors and its existing backers. It plans to bring its all-optical switches into volume production, build larger versions and expand its Silicon Valley offices.
The company’s case rests on keeping data as light while it moves between AI chips, rather than converting it at every switch. Its performance figures come from Salience itself, and it has not yet named customers. With Tower and Keysight, it now aims to take the switch from live demonstrations to high-volume production.
