Arago

Arago is a Paris-based deeptech semiconductor company developing a photonic AI accelerator that uses light rather than electricity to perform AI computation. The company positions its product against transistor-based silicon accelerators, which it argues are constrained by compute, memory and power ceilings as AI models grow. The chip, codenamed JEF, is described as architected to deliver 10x-30x reductions in energy consumption relative to conventional GPUs while maintaining compatibility with industry-standard software frameworks and existing semiconductor manufacturing processes.

The architecture combines an "AI unit" built on a proprietary multi-physics computing core that accelerates high-volume tasks with n^3 complexity at low power, and a "control unit" with a deterministic architecture aimed at memory-transfer efficiency and pointwise (n^2 complexity) operations. Arago claims that a byte of data enables thousands of operations on its system compared with a few operations on a GPU, yielding multi-PetaOp/s performance, with the largest advantage on tensor sizes in the hundreds and thousands. A software stack branded CARLOTA interfaces with PyTorch and abstracts low-level hardware details at the instruction level so that models can be deployed without code changes.

The company was co-founded in 2024 by Nicolas Muller (CEO), Eliott Sarrey (CTO) and Ambroise Müller (CSO). Per investor materials, Muller has an ML academic background from MIT and prior early-team experience at EBIT-positive deeptech startups; Sarrey is a condensed matter physicist from ETH with a bachelor's in mathematics and physics from École Polytechnique and was the first French winner of the Google Science Fair; Müller studied at École Polytechnique and worked as an AI algorithm engineer at Eviden. One aggregator profile reports a team of roughly 20 people.

Founding story

Arago was co-founded in 2024 by Nicolas Muller, Eliott Sarrey and Ambroise Müller around the premise that transistor- and electricity-based chips are reaching limits in speed, energy consumption and cost as AI models scale, and that light can move faster with less energy than electricity. The founders combine backgrounds in machine learning, condensed matter physics and high-performance computational physics.

Business model

Arago designs and intends to commercialize AI accelerator hardware, paired with its own CARLOTA software stack, for AI compute workloads. Sources describe a combined hardware and software stack as a source of defensibility, but do not specify pricing, sales contracts or delivery models.

Traction

Arago reports that its results to date show its technology can run AI models using industry-standard software frameworks while remaining compatible with existing compute infrastructure and semiconductor manufacturing processes. It closed a $26M seed round in July 2025, described as oversubscribed, to commercialize an initial prototype; investor material references an Arago semiconductor device and a Paris lab. No customer, revenue or shipment figures are disclosed in the sources.

Latest developments

The July 2025 seed round of $26M is the most recent disclosed development. Proceeds are earmarked for accelerating R&D, scaling the team across France, Israel and North America, deepening business and industrial partnerships, and commercializing the initial prototype of the chip codenamed JEF.