A Paris and Palo Alto deeptech betting that the next AI chip will run on light instead of electrons - and that nobody will have to rewrite their PyTorch code to use it.
Arago, Inc. - Photographed on its first lap, somewhere between a foundry and a Series A. 40 people. Three continents. One beam.
Profile - June 2026
Picture a clean room in a leased corner of a Paris semiconductor lab on a Tuesday afternoon. A wafer the size of a salad plate is being aligned under a microscope. On it: tiny photonic waveguides routed like a city's subway map, designed to push light through matrix multiplications instead of pushing electrons through transistors. A small group of physicists is talking with software engineers about a PyTorch test that ran the night before. The test ran fine. That, more than anything else, is why Arago exists.
Arago is a 40-person company that did not exist 24 months ago. It now has $26 million in seed funding, a chip codenamed JEF, a software stack named CARLOTA, offices in Paris and Palo Alto, and a quietly audacious thesis: AI's real bottleneck isn't model size, it's the power bill. Solve the power bill with photons and the rest of the industry can keep using the tools it already knows.
That last clause is the trick. Lots of people have tried optical computing. The graveyard is famous. Arago's bet is that the way you win is not by reinventing the developer experience but by hiding the physics underneath it.
What it is
Handles the high-volume matrix operations - the n³ part of inference - using light moving through silicon waveguides. The unit aims for multi-PetaOp/s throughput at a fraction of the wall power a GPU would draw to do the same work.
A conventional electronic core handles the pointwise n² operations and orchestration. Photons are great at multiplying, less great at branching - so Arago lets each substrate do what it's best at.
The software stack that hides the photonic strangeness from developers. Drop in a model, get inference out. No re-architecting required, which is the whole pitch in three words.
By the numbers
AI inference is becoming the single largest line item in data-center power planning. Arago's stated goal is to deliver the same performance as today's leading GPUs at roughly a tenth of the energy. The bar chart below is the company's headline claim, normalized to GPU = 100.
Closed July 2025, oversubscribed.
Lower than leading GPUs at equivalent compute.
Across France, North America and Israel.
Photonics, ML and software, in one room.
It's a short sentence. It also reads like a mission statement, a pitch deck slide, and a recruiting page in one. The rest of the company is essentially an attempt to un-box it.
The trio
The team is structured the way the chip is. Photonics, electronics, and software each have a champion in the founding circle. That isn't a coincidence; it's the org chart.
Runs the business and the transatlantic axis. Carries the company between Paris and Palo Alto.
Owns the chip - the silicon photonics, the multi-die packaging, the path from idea to engineering sample.
Drives the science: the math that turns matrix multiplications into beams of light and back again.
The cap table
The $26M seed was co-led by Earlybird, Protagonist and Visionaries Tomorrow, with participation from Generative IQ and C4 Ventures. The angel list reads like a who's-who of operators who've built infrastructure at scale.
Recent past
Three co-founders meet around a thesis: AI's power problem is a hardware problem, and photonics deserves another look now that silicon photonics manufacturing is real.
Earlybird, Protagonist and Visionaries Tomorrow co-lead. The angel list signals that AI-software operators are paying attention to AI-hardware bets.
The chip is named and pitched publicly. Coverage in EE Times, Photonics Spectra, Data Center Dynamics. Early benchmarks claim 10x energy advantage at parity compute.
Team grows past 40 across three continents. Focus shifts to silicon-photonic integration and first deployment partners in AI inference, edge compute, and low-power data centers.
Inside the building
Arago publicly recruits engineers and physicists who, in the company's own phrasing, want "great science and fast achievements." Read: people who can publish a paper and ship a milestone in the same quarter.
Trust, respect, camaraderie - written down, on the website. Plus a stated allergy to unnecessary process. The org runs across France, Israel and North America, so a flat default is structural, not aspirational.
The chip wars
The photonic computing field has a few familiar names. Lightmatter, Lightelligence, Celestial AI and Q.ANT each carry a version of the optical thesis. Arago's distinguishing move is the multi-physics architecture: not photonics alone, not electronics alone, but both on the same die with a deterministic control plane sitting between them.
Photons multiply matrices; electrons handle branching, control flow, and the parts of AI that aren't linear algebra. The split is the architecture.
Built around existing silicon photonics manufacturing and existing developer tooling. Customers don't have to adopt a new fab or a new framework to adopt Arago.
Index
Coda
Return to the clean room. The wafer is still there. The PyTorch test from the night before is still saved on someone's laptop. What has changed in the last few minutes of reading is the framing of what's on the bench.
It isn't a science project anymore. It's $26 million of investor conviction, three co-founders with overlapping but non-redundant skills, and a forty-person crew who treat a foundry tape-out like a sprint demo. The wafer is just the physical form of a bet that the constraint on AI is no longer "make the model smarter" but "make the joule cheaper." If Arago is right, the room in Paris is where that gets answered. If Arago is wrong, the room is still a pretty good place to be wrong from. Either way, somebody had to try doing this with light. Arago put up its hand first - and then put up the funding to back the hand.
The salad-plate-sized wafer goes back into its case. Somewhere across the Atlantic, a customer slide deck is being edited. Outside, Paris has moved on to dinner. Inside, the test is still running.