Arago bets light-powered chip JEF cuts AI energy 10x, stays GPU-compatible

Arago is a fabless deep-tech startup building 'JEF', a multiphysics AI inference processor that combines analog and digital electronics, silicon photonics and free-space optics; it claims GPU-level performance at up to 10x lower energy while running standard AI frameworks.

Light can do AI inference with ~10x less energy than GPUs — provided the chip drops into today's software, fabs and data centers with no ecosystem rewrite.

$26M (€22.1M) oversubscribed seed, co-led by Earlybird, Protagonist and Visionaries Tomorrow with Generative IQ, C4 Ventures and angels including Bertrand Serlet, Olivier Pomel (Datadog), Thomas Wolf (Hugging Face) and Jack Abraham; announced 2025-07-07.

Arago was founded in Paris less than a year before its July 2025 announcement by Nicolas Muller (CEO) with Eliott Sarrey and Ambroise Müller — physics and AI researchers who met at a maker space, with Sarrey and Müller having worked on silicon photonics and AI algorithm development at ETH Zürich. They started from the top-down question of what it takes to infer a model very efficiently, then wrote the datasheet for a product; Muller argues the timing only became possible because key components and processes matured about one to two years earlier.

Arago built a ~20-person team across France, the UK, the US and Israel, half of it on software, and got a JEF test chip running AI inference for multiple models on a small card. In July 2025 it closed an oversubscribed $26M seed co-led by Earlybird, Protagonist and Visionaries Tomorrow, with Generative IQ, C4 Ventures and angels including former Apple VP Bertrand Serlet, Datadog co-founder Olivier Pomel, Hugging Face co-founder Thomas Wolf and Exowatt co-founder Jack Abraham; Earlybird called the architecture a potential 'DeepSeek moment' for AI chips. The round funds development of a data-center inference accelerator, with edge and robotics applications later, and the team is fabless, relying on mature processes rather than a new manufacturing line.

Inference energy is becoming the binding constraint on AI scale, and photons generate far less heat than electrons — a physics-level advantage over GPU roadmaps.

Past photonic attempts failed on integration, not physics; Arago limits optics to specific places and reuses mature silicon processes to stay manufacturable.

Drop-in compatibility is the adoption strategy: no new language, one-line software change, standard frameworks — the ecosystem does not have to adapt.

The timing claim is testable: components became mature and cost-effective only one to two years before founding, and the team got a test chip running inference within its first year.

In deep tech, compatibility is the wedge: the hardest part is not the new physics but today's software, manufacturing and deployment — design for the ecosystem that exists, not the one you hope for.

As of 2026-09-02 Arago is in development with no commercial product or announced customers, building the JEF processor and software stack across hubs in France, the UK, the US and Israel. Claimed early results (GPU-level performance at up to 10x lower energy on standard frameworks) come from the company and investors and are not yet independently benchmarked. The open questions are whether a hybrid photonic chip can reach datasheet performance, cost and reliability at data-center scale, and how it fares against incumbents and photonic rivals like Lightmatter.