Arago, founded just over a year ago, believes it has cracked one of artificial intelligence's most pressing problems: its voracious appetite for electricity.
Last week, the company announced it had raised $26 million in seed funding to commercialize what it calls a "multiphysics processor" – a chip that uses photons instead of electrons to run AI models while consuming a fraction of the energy of today's graphics processing units (GPUs).
The timing couldn't be more critical. As AI systems proliferate across industries, their energy demands are becoming unsustainable. Data centers already consume about 1% of global electricity, and that figure is projected to reach 3-4% by 2030 as AI workloads explode. Each advanced GPU can consume as much power as a household oven – around two kilowatts – while running continuously in server farms.
"You can think of a processor, the GPUs, as basically a board with billions of transistors," said Nicolas Muller, Arago's CEO and co-founder. "Transistors are basically switches like you have in your house and that turn on and off. And every time they turn on and off, they dissipate a lot of heat. You can imagine that in a current GPU, you have 80 billion of them, and they switch around two to 3 billion times per second."
The result, he notes, is a processor "more dense than the heart of a nuclear power plant."
From Serendipity to Silicon Valley Backing
Arago's origin story began with a chance encounter at a maker space. Muller, who studied finance, economics, and machine learning at MIT, was working at a semiconductor startup in Silicon Valley when he met his two co-founders, Eliott Sarrey and Ambroise Müller, both physicists who had studied together at prestigious French universities.
"It was kind of serendipity," Muller recalled. "They knew each other for a long time. They studied at Polytechnique together, theoretical math and physics, and then they switched to ETH in Zurich, where one focused on optical research... and the other was developing quantum and AI algorithms."
The trio shared a common frustration: AI was being held back by hardware limitations.
"We saw from ourselves that AI is basically boxed in today's hardware," Muller said. "And so we wanted to create a new type of hardware to see what's next, what it feels like to have an accelerator, a processor that is purpose-built for AI, that unlocks new applications and new AI economics."
What sets Arago apart from other photonic computing attempts is its hybrid approach. Instead of building a pure optical processor, which faces significant technical constraints, the company developed what it calls a "multiphysics design" that combines digital, analog, and optical components.
"Instead of having thousands, if not hundreds of thousands, of these switches to represent one number, we have only one laser that encodes this value," Muller said. This approach allows the company to "leverage the benefits of light, which are reduced energy consumption and higher throughput, but also make sure that we are compatible with the existing ecosystem and manufacturing processes."
Beyond the Lab: Real Hardware, Real Applications
Unlike many deep-tech startups that exist primarily in simulation, Arago already has working hardware. "Today we have a processor, not simulation, not an academic experiment, a processor that fits into my hand," Muller said.
The company is targeting two main markets. First, data centers, where cloud providers are desperate to reduce energy costs. Second, edge computing applications like robotics, where power constraints severely limit capabilities.
"If you look at today's warehouses where robots operate, they operate for one hour and produce some results. Or, they just stand still in a corner because you have two huge GPUs in the back that are draining all the power out of it," Muller said. "And so you have, with GPUs, you have one hour of operations. But with Arago's processor, you have enough for one day of uninterrupted operations."
In terms of adoption, Arago's chip is compatible with all existing systems and platforms, making it plug-and-play. So companies don't need to change their code or higher an army of developers to re-architect their systems.
That compatibility will be crucial down the road. The company plans to extend this technology to personal computers as well, aiming to democratize AI by enabling powerful models to run on consumer hardware.
Currently, most people can only run relatively simple AI models locally. "More or less everybody can run probably 7 billion parameters on their computer," Muller noted. "But basically, a 7 billion parameter model struggles to convert a text from French to English, right? So in reality, it's impractical, and it does not deliver any results."
Manufacturing Without Asia
Arago has made a strategic decision that differentiates it from most semiconductor companies: manufacturing entirely outside of Asia. As a fabless company, Arago designs chips but outsources production to third-party manufacturers in Europe and the United States.
"We manufacture these processors in Europe and in the US. There is no part being manufactured in Asia, because we don't need to be where there are the most advanced processes," Muller said. This approach not only addresses supply chain sovereignty concerns, but also reduces costs by using more mature manufacturing processes.
"We can go for basically more mature and lower-cost processes," he said, noting that the company doesn't need the extreme precision required for cutting-edge transistors.
Heavyweight Backing and Global Talent
The company's oversubscribed funding round was co-led by European VCs Earlybird, Protagonist, and Visionaries Tomorrow, with participation from notable angel investors including former Apple VP Bertrand Serlet, Nvidia veterans, and Hugging Face co-founder Thomas Wolf.
Arago has assembled an international team of 20 people representing nine nationalities, drawn from leading tech companies and research institutions. "We have people that are more junior and people that have 25-30 years of experience," Muller said. "And so we have this very interesting blend where people are extremely gifted in terms of engineering and science, and they really love to ship product extremely fast."
The team's diversity extends beyond nationality. "Everybody within Arago has an interesting story. People left their country because of war, and they did not speak a single word of French," he said. "For example, they were studying another field, but turns out they can do mathematics, because this was the only language that they could understand and graduate from the best universities in the world."
The NVIDIA Question
Given the dominance of NVIDIA in AI chips, the obvious question is whether the GPU giant might simply copy Arago's approach or acquire the company outright. Muller sees several factors protecting his startup's position.
"NVIDIA and Intel spent the last 30-50 years optimizing transistors, and they are now in a phase where they need to ship hundreds of thousands of these processors per month. So they cannot afford to do some breakthrough R&D at the moment," he argued.
Additionally, the market dynamics favor incremental improvements over revolutionary changes. "As for now, it's still less expensive and less risky to add more silicon, add more transistors," Muller said. "And customers are very risk-averse."
But he acknowledges NVIDIA's financial power: "They have so much capital now and they are approaching $4 trillion of valuation, so they can just afford to wait and see who's going to be the winner and just buy them right."
When asked if being acquired by NVIDIA is the plan, Muller is emphatic: "The answer is no, we wanted to really build a champion and a challenger for this industry."
Racing Toward Commercialization
With funding secured, Arago plans to dramatically scale its operations over the next 12 months. The company will double its headcount across its offices in Paris, the US, and Israel while accelerating product development toward mass production.
"We will use our funds to continue what we started. So, invest heavily in the R&D of our project. It's a race, so we need to go extremely quickly," Muller said.
While he can't disclose specific customers or timelines due to competitive sensitivities, Muller hinted that commercial deployments are imminent: "You can expect to see, in the near future, more and more Arago processors in the next few months."
As AI continues its rapid expansion across industries, the race to solve its energy problem has become critical. Whether Arago's photonic approach can deliver on its ambitious promises – and scale quickly enough to matter – may determine not just the startup's fate, but the viability of AI's continued growth in an energy-constrained world.
"What drives us at Arago is the opportunity to do something that is non-incremental," Muller said. "Trying to reinvent part of the computer, not all the computer, but a part of it. That's what we're doing."
