French chip startup Arago has raised $26 million in seed funding to develop its light-based AI accelerator, codenamed Jef. The company aims to reduce AI inference energy by 10x compared to today's GPUs by combining analog and digital electronics, silicon photonics, and free-space optics.
www.eetimes.com, Jul. 08, 2025 –
While most photonic computing startups are founded by optics or photonic researchers, Arago’s founders are physics and AI researchers, which Arago CEO Nicolas Muller argued gives the team a different perspective.
“We started from the top, meaning: what does it take to infer a model very efficiently?” Muller told EE Times. “Then we turned that into a data sheet for a product.”
Arago, which is less than a year old, was founded by three young entrepreneurs who met at a maker space – Muller, Eliott Sarrey, and Ambroise Müller. Sarrey and Müller met at ETH Zurich while working on silicon photonics and AI algorithm development, respectively. They will serve as co-CTOs.
Arago’s key challenges, as identified by Muller, are to get a photonic computing chip working reliably and to make it performant enough and easy enough to use so that it can be adopted into AI technology ecosystems.
“Historically, companies that were tackling this space were born around 5 to 10 years ago, where the technical and scientific landscape was extremely different,” Muller said. “You had pure optical technologies and different technologies that are based on free space optics. What we built is a technology that overcomes all of those challenges by only using optics at very specific places.”
Arago calls its inference chip, Jef, a “multiphysics processor” because it combines elements of analog and digital electronics with silicon photonics and free-space optics. Techniques like spatial light modulators (SLMs) from free space optics are limited by modulation frequency, he said, while silicon photonic modulators are limited by the maturity of the manufacturing processes.
“We combine some classical silicon processes with specific optical components that we integrate into a unified and coherent process and architecture to make it work,” Muller said.
Combining electronic and photonic elements enables a functional processor that can be manufactured using current processes and integrated into today’s deployment ecosystems, Muller said.
“The second aspect that made Arago possible is the timing of this project,” he said. “From a pure supply chain perspective, because we are a fabless company, we now have some components and some processes that only became mature and cost-effective enough around one to two years ago.”
So, versus competitors like Lightmatter, Lightelligence, Salience, and Lumai, Arago’s starting point was very different, but its technology and approach are different too, Muller said.
Half of Arago’s team of 20 is working on software.
“What we don’t want is that people need to learn a new language, change their habits, or just take ages to integrate our processor into their existing ecosystem,” Muller said. “So, the team at Arago is building a dedicated software stack that abstracts all these differences so that AI developers, researchers, or data scientists can access our processor by changing a single command line.”
The eventual product will be a data center inference accelerator, though the technology could be applied to edge applications like robotics further down the line, Muller said.
“We received lots of interest for different applications more directed toward the edge,” Muller said. “If you think about a data center, a processor like Argo that cuts the energy consumption by 10 to 30x is very interesting from a scalability and financial perspective. But for edge applications such as robotics, this kind of performance is the difference between a prototype and mass production.”
The technology could also be applied to training chips, Muller said, but system level challenges like building large clusters, plus strong competition from the market leader, make this market less attractive.
Arago has a Jef test chip up and running on a small card which is already running AI inference for multiple models, Muller said.
The team is currently at around 20 people in France, the UK, the U.S., and Israel. Funding will be used for product development and to grow teams at each of its technical hubs.