Lightspring: Shaping the Future of 3D Lithography-based Optics
Photonics has quietly become crucial for AI infrastructure. Moving that much data between chips is starting to outpace what copper wiring can handle, so the industry is turning to light. The catch is nobody has found a way to manufacture that connection at scale.
Fabricating the photonic chip itself is now fairly mature work. Getting light in and out of it is not. Assembly and packaging eat up roughly half the cost of a finished chip, against about a fifth for the foundry step and the process still loses 2 to 3 dB of light at every connection, with throughput and yield to match.
Lightspring’s bet is that this step should not be manual in the first place. Rather than nudging fibres and lenses into position and (basically) gluing them piece by piece, the company prints the coupling structures themselves straight onto the chip. They use two-photon polymerisation, a laser-based 3D-printing process precise enough to work at that scale. What used to be a slow, bespoke assembly job becomes a single, repeatable manufacturing step. It’s the kind of standardized process the industry needs if optical packaging is going to scale with AI demand rather than remain its bottleneck.
Lightspring was founded in Besançon in France by Grégoire Bonnat (CEO), Adrià Grabulosa (CTO) and Daniel Brunner (Scientific Director). The company has just closed a €3.5 million Seed round led by OVNI Capital and Atlantic, alongside Concept Ventures, Ixcore, Plug and Play and strong angels from the industry.
Learn more about how 3D-lithography-based optics could reshape photonic packaging from our interview with Lightspring co-founder and CEO, Grégoire Bonnat:
What Inspired You to Start Lightspring?
I trained as an engineer and studied electronics but finally moved to entrepreneurship. My first company was in a completely different industry: I started building something like an Uber competitor for public transport, which pivoted into on-demand public transport software deployed in hundreds of cities and we eventually sold it to Siemens. That was the first chapter of my entrepreneurial life, ten years in total between the pivots, the scaling and the integration into Siemens.
Afterwards I was looking for a new project and started exploring the deep tech community. Daniel (now Scientific Director) was one of the first people I met and our first discussions were around optical computing, which has been his research for more than 15 years, building optical neural networks, an idea still considered largely experimental. That discussion pulled me in. He was solving a real pain point for the industry and the market looked like it was at a tipping point, the foundation for building a great venture.
We started from this optical neural network idea, which is still a bit speculative today. But that kind of system only makes sense at a certain scale and that’s why Daniel wanted to start research specifically on photonic integration: coupling lots of photonic components into a dense, high-performance system. That’s the basis of what Daniel started in 2020 with Adrià (now CTO) as a PhD student.
Editor’s note: Optical neural networks use light instead of electrical signals to perform the layered, weighted computations behind AI, with structures like lasers and waveguides carrying out the equivalent of a neural network’s matrix multiplications as light passes through them.
What Problem Does Lightspring Solve and Why Is the Timing Right?
AI is pushing optics closer and closer to compute. That shift is turning photonic packaging into one of the industry’s most critical manufacturing bottlenecks and an increasingly expensive one, since packaging already costs more than making the chip itself. When we looked at what Daniel and Adrià had built at FEMTO-ST around coupling many photonic components together at density, originally for optical computing, it was clear the timing had changed. AI infrastructure now needs optics sitting right next to the chip rather than running between racks, at a scale nobody anticipated even a few years ago.
Editor’s note: Photonic packaging is the step where a photonic chip gets connected to fibres, lasers or other chips so light can actually get in and out of it, as distinct from fabricating the chip itself.
Editor’s note: FEMTO-ST is the CNRS-affiliated research institute in Besançon, France where Daniel and Adrià ’s underlying technology originated, one of Europe’s leading centres for optics and micro-nanotechnology.
The problem underneath that shift is connecting components optically, especially across different materials. That has never been solved at scale. Today it still happens one connection at a time by hand and that is the industry’s real bottleneck, not chip fabrication, which is already mature. There is also no shared standard for how it should be done. That is the layer we are building toward: an equipment-independent way to couple light into and out of a chip that the rest of the industry can design against, the same way a manufacturing standard becomes something everyone builds on rather than something everyone reinvents.
What Does Lightspring Actually Build to Solve This, on the Hardware and Software Side?
The established way of doing things is high precision pick and place. Any misalignment loses a lot of signal between components so it requires very high precision. We break that: you can assemble your components with lower precision and in one process step we 3D-fabricate all the optical connections. Imagine two components separated by one tenth of a millimetre, about the width of a hair and in that trench we build a kind of miniature optical fibre. This works because our integrated, software-driven approach compensates for the misalignment between components and finely tunes the waveguides and coupling structures to the mode, meaning the characteristics of the light signal coming out of each component. It’s a manufacturing step that has been blocked and we turn it into something programmable, captured in a library of coupling designs known as a PDK.
Editor’s note: An optical fibre is a thin strand of glass or plastic that carries data as pulses of light, the standard way light is transported between components before it needs to enter a chip.
Editor’s note: Waveguides are microscopic structures that confine and guide light along a controlled path, the on-chip equivalent of what a fibre does over longer distances.
Editor’s note: Coupling structures are the physical elements, lenses, waveguides or similar shapes, engineered to get light to transfer efficiently from one component to another without losing signal.
Editor’s note: A PDK (process design kit) is a certified library that tells chip designers which shapes and materials reliably deliver a given performance, letting them design without needing to understand the fabrication process itself.
What Does the PDK Contain and How Does It Change the Way Designers Work?
We’re using a process that’s getting more popular but still isn’t widely adopted by the industry so there’s no standard for it yet. What we offer is a platform that lets designers build what they want without rebuilding the whole manufacturing chain around their own architecture. They come with their design and their core components and we build a high performance system around it, drawing on a library of coupling solutions that covers the range of use cases needed, paired with fabrication recipes that have already been qualified rather than reinvented per project.
Editor’s note: Fabrication recipes are the specific, tested sequences of process parameters, materials, exposure settings and so on, that reliably produce a given coupling structure, the manufacturing equivalent of a proven recipe rather than trial and error each time.
As far as we know it would be the first PDK built specifically for 3D optical coupling. In the end it mixes design expertise with process control, aiming for solutions that are as close to off-the-shelf and easy to use as possible.
Can You Walk Us Through How 3D Lithography Replaces Active Alignment?
Active alignment is the idea that to get most of your signal from component A to B, you slowly move the components and measure the output. When the coupling is at its maximum, you glue it in place. We replace that with lower accuracy passive alignment. It’s a much faster pick and place: you put the whole wafer or substrate into our machine, we detect the actual position of every component and we tune the coupling structures in real time to fabricate them exactly where they need to be.
Editor’s note: A wafer is a thin disc of semiconductor material on which multiple chips are fabricated at once, later cut apart into individual dies.
Editor’s note: A substrate is the base material a chip or its components are mounted on and connected through, similar in role to a wafer but used more broadly across packaging.
The deep expertise in designing those structures, especially in 3D, isn’t something the industry really has today. The whole semiconductor industry is used to 2D processing, that’s the basis of lithography and the existing software stack (the EDAs). Even when you stack a lot of 2D patterns, people are still managing 2D. Our case is pure 3D design and 3D fabrication control, which is genuinely new for the industry. The target is coupling losses below 0.5 dB per interface, with design cycles under two weeks against several months for bespoke optical packaging.
Editor’s note: EDAs (electronic design automation tools) are the software used to design and simulate chips, standard for 2D electronic circuits but not built for 3D optical structures.
How Does Lightspring Fit Into the Foundry and OSAT Ecosystem?
Our goal is to offer a platform we can fabricate ourselves at small to mid volumes. From there we want to reach a maturity stage where we can outsource fabrication to partners (foundries, OSATs). Our role is to bring the technology to a maturity level where we can guarantee yield and performance and have a clear playbook for how the solution is implemented. Then we go into the fab with our partner and implement it at the right scale. On the design side, we work as a partner to chip designers and system integrators, helping them build their systems on top of the coupling and integration platform we provide.
Editor’s note: Yield is the percentage of manufactured units that meet quality standards and can actually be used or sold, a key measure of how reliable and scalable a manufacturing process is.
What Advice Would You Give to Fellow Deep Tech Founders?
Something I’ve heard a lot from institutional stakeholders in the deep tech community is that deep tech should be done a certain way, because of tech transfer, IP and so on. You start adopting a certain mindset about how it should all play out. To some extent I think that’s wrong. I think you should apply the same playbook as a classical startup, because the whole startup world is moving towards harder problems anyway. Investors and markets expect you to work at the same velocity people did in software a few years back. So build quickly, iterate quickly, customer first, the same playbook. Don’t let yourself be influenced by people who say this will take two or three years for tech transfer and then another three for a prototype. You should work as fast as possible, the way people do in Silicon Valley and in China.