Nicolas
Muller

Somewhere in Paris, at a maker space, three people who understood both the mathematics of machine learning and the physics of light decided to flip the AI chip problem upside down. Nicolas Muller, Eliott Sarrey, and Ambroise Müller didn't start with "how do we make photonics work?" They started with a much sharper question: what does it actually take to run a model efficiently? Then they built a chip to match that spec.

That distinction - top-down versus bottom-up - separates Arago from almost every other company working on optical computing. The photonic research field has been promising faster AI for years. Most of its startups are founded by optics specialists who later hunt for applications. Muller's view is that this is the wrong order of operations.

Muller brought to the table an ML degree from MIT alongside a background in physics and deep-tech company building. The intersection mattered. Understanding both how neural networks actually compute - specifically the relentless demand for matrix multiplication at scale - and how photons move through engineered optical systems gave Arago's founders a lens most chip designers don't have.

The result is JEF, Arago's multi-physics processor. It's not a purely optical chip - it's something more pragmatic. JEF merges analog and digital electronics with silicon photonics and free-space optics, using standard manufacturing processes that already exist in mature foundries. That "standard fab" requirement wasn't an afterthought; it was baked into the design brief from the beginning.

One laser. One number. The simplicity of the statement hides the engineering depth behind it. Photons don't generate heat the way electrons do. They travel at the speed of light through silicon waveguides without resistive loss. For the specific workload of matrix multiplication - the computation that dominates every transformer, every image model, every LLM inference pass - light does the math with far less energy than any transistor-based architecture can achieve. Arago claims 10x lower power consumption than top GPUs at equivalent performance and cost.

But performance claims are common in the chip industry. What makes Arago's position more credible than most is the execution timeline: they had a working prototype within 12 months of founding. In semiconductor development, where tapeout cycles alone often stretch past a year, that's a pace that gets attention. The prototype got investor attention, too.

Ecosystem-first by design

The chip was always going to be measured against a brutal real-world test: does it actually plug into the existing stack, or does it ask AI teams to relearn their tools? Muller made the answer non-negotiable early.

JEF is designed to integrate with PyTorch and TensorFlow without friction, targeting AI inference workflows as they exist today - not as some hypothetical future state. This compatibility constraint shaped the hardware architecture. It also shaped the commercial story: data center operators can slot Arago's chip in without a facility retrofit or a developer retraining budget.

There's also a supply-chain timing argument. Arago is a fabless company, meaning it designs chips but outsources manufacturing. Muller has been candid that the company's founding moment was partly determined by when specific components became mature and cost-effective enough to build around.

The $26M round and who signed the check

In July 2025, Arago closed an oversubscribed $26 million seed round. The lead investors - Earlybird, Protagonist, and Visionaries Tomorrow - were joined by C4 Ventures and Generative IQ. The angel list was its own signal: Bertrand Serlet, who ran software engineering at Apple before co-founding Fungible; Christophe Frey from Arm; Olivier Pomel, who built Datadog; Thomas Wolf from Hugging Face; and Jack Abraham of Exowatt.

That's a group that spans chip design, enterprise infrastructure, AI software, and energy - almost exactly the stack Arago needs to navigate to commercialize a new compute principle. The round was oversubscribed, which in the current deep-tech climate suggests the prototype conversations went well.