Before the rise of generative artificial intelligence, few investors were interested in optical components in data centers. The spotlight was on processors, memories or software. Two years later, the situation is changing, and as AI infrastructures grow in size, another technology is coming to the forefront: silicon photonics. The raising of 4.5 million euros from the Swiss startup Aylight is an illustration of this. If it does not mark the appearance of a new technology, it reveals a more profound change where AI could offer silicon photonics a market capable of changing scale.
An old technology that’s finally finding its moment
The first work on silicon photonics dates back to the 1980s, while major industrial programs took shape in the early 2000s. The ambition is to use manufacturing processes from microelectronics to produce components capable of manipulating light directly on silicon chips.
The idea is to combine two worlds. Electronics remain the most efficient for carrying out calculations. Light transports information with very low attenuation and a much greater capacity when it comes to moving large volumes of data.
Until now, this technology has mainly found applications in telecommunications networks and optical links connecting servers in large data centers. Manufacturers like Intel have been investing in this area for more than a decade, while specialists like Coherent or Lumentum already supply numerous optical components used by operators and hyperscalers.
If artificial intelligence does not invent silicon photonics, it profoundly modifies the economic conditions of its adoption.
AI changes the nature of the problem
A GPU is only valuable if it can quickly exchange data with the other processors that make up a cluster. The most advanced models are trained on tens of thousands of GPUs running simultaneously. Future generations are expected to mobilize several hundred thousand accelerators spread across several buildings, or even several campuses.
In these architectures, data circulates constantly: synchronization of models, exchanges between memories, communications between processors, power supply to storage systems. Internal traffic is growing faster than computing power itself. One of the main challenges is no longer just to calculate faster, but to transport more information without degrading performance or exploding energy consumption. This is precisely where silicon photonics regains its relevance.
Replace some electrons with photons
In a classic electronic chip, information circulates in the form of electrical signals. Photonics uses, for certain transmission functions, light pulses which propagate in waveguides etched directly on a silicon chip. The calculation remains carried out by electronic components.
This distinction is essential. Photonics does not replace electronics, but complements it when exchanges become too important to be effectively ensured by electrical interconnections.
This architecture is based on several families of components: waveguides which direct the light, modulators which convert electrical signals into optical signals, photodetectors which carry out the opposite operation, but also multiplexing devices capable of circulating several wavelengths simultaneously on the same fiber.
At the heart of this whole is an often overlooked component: the laser.
Why lasers are becoming a strategic issue
Every optical link begins with a light source, and without a laser, there is no optical transmission. However, these components must meet several contradictory requirements: produce extremely stable light, consume little energy, be compact enough to integrate into network equipment and remain compatible with large-scale industrial manufacturing.
It is precisely on this link that Aylight intervenes, resulting from work carried out at ETH Zurich, the startup develops multi-wavelength lasers capable of generating several optical channels from a single chip. This approach aims to replace multiple separate lasers with an integrated architecture, thereby reducing the number of components, power consumption and assembly complexity.
The interest is not to transform the architecture of data centers, but to improve a component whose performance directly conditions that of optical networks.
The fact that this technology is designed to be manufactured in existing photonic foundries constitutes a decisive element. In the semiconductor industry, the most promising innovations are rarely those which impose new production chains, but those which fit into existing industrial tools.
A new layer of AI infrastructure
The rise of photonics reflects a broader evolution of the semiconductor ecosystem. Today, every layer of infrastructure becomes strategic.
Memory is progressing with HBM chips from SK hynix, Samsung Electronics or Micron Technology. The networks are dominated by Broadcom, Cisco or Marvell Technology. Advanced packaging mobilizes considerable investments at TSMC, while cooling systems are also becoming differentiating elements.
Photonics is part of this growing fragmentation of the value chain. This development also favors the emergence of a new generation of specialized startups. Some develop photonic engines, others optical interconnections, design software or packaging technologies. Scintil Photonics, EFFECT Photonics, Ayar Labs and Celestial AI illustrate this growing specialization.
Industrialization still far from complete
The current interest in silicon photonics should not, however, mask the challenges that remain to be met. The first concerns the manufacturing itself. Silicon is an excellent material for guiding light, but it does not naturally make a good laser source. Many players still have to assemble different materials or integrate separately manufactured components, which increases industrial complexity.
The second challenge is economic, electrical interconnections remain extremely competitive over many distances and for many uses. Photonics will not replace them everywhere. Its adoption will progress where flow rates, distances or energy consumption justify an additional initial cost.
Finally, the rise of this industry will also depend on production capacities. As with traditional semiconductors, value will not only be in the design of the components, but also in their large-scale manufacturing. Foundries capable of producing photonic chips in volume could become as strategic players as the manufacturers themselves.
Does Europe have a card to play?
Unlike GPUs, which are largely dominated by the United States and Asia, silicon photonics still offers a relatively open landscape. Europe has recognized academic research, several centers of competence in integrated optics and a particularly active network of startups in Switzerland, France, the Netherlands and Belgium.
However, this scientific advance does not guarantee any industrial leadership. The question is no longer just about developing the best components, but about having the production capacities, industrial partners and commercial outlets necessary to scale up.
Founded in 2025 by Bahareh Marzban and Dmitry Kazakov, following their research work at ETH Zurich, Aylight raised €4.5 million in a pre-seed round co-led by Elaia and Swisscom Ventures, with participation from Verve Ventures and Plug and Play. The capital will finance the first prototypes manufactured in the foundry, the strengthening of R&D teams and the acceleration of the industrialization of its on-chip multi-wavelength laser technology.
