Swedish photonics company Solinide Photonics has raised €4m to commercialise optical technology intended to reduce the power, space, and hardware needed to move data between processors inside increasingly dense AI computing infrastructure.

The Chalmers University of Technology spin-out develops silicon nitride photonic integrated circuits and microcomb technology for optical communications. Its seed round brings together existing investors Chalmers Ventures and Navigare Ventures with PSV Hafnium, Turbine Capital, Norrsken Evolve, and Almi Greentech Fund.

Rather than funding another attempt to design an AI accelerator, the investment targets a part of the computing system that has become harder to ignore as clusters expand. Training and running large models requires thousands of processors to exchange enormous volumes of information, and installing faster chips achieves little when the interconnects between them become a bottleneck.

Solinide’s approach uses a photonic chip to generate many precise wavelengths of light simultaneously. Conventional optical systems can require separate laser sources for individual wavelengths, whereas a microcomb can consolidate part of that hardware while carrying multiple data channels through the same fibre.

AI’s constraint moves between the chips

The infrastructure surrounding AI processors accounts for a growing share of system complexity and electricity use. Accelerators attract most of the capital and attention, but large clusters also depend on switches, transceivers, lasers, memory, cooling, and high-speed networks capable of keeping those processors supplied with data.

As model sizes and workloads grow, processors have to communicate quickly enough that expensive computing capacity is not left waiting. That places additional pressure on electrical interconnects and optical networking, particularly when information has to travel beyond one server or rack.

Solinide says its silicon nitride microcomb can replace multiple discrete laser sources with one device producing dozens of wavelengths. The company has demonstrated its technology in a rack-mounted system and says the investment will support industrialisation, manufacturing readiness, and commercial deployment.

Marcello Girardi, chief executive of Solinide Photonics, said: “A key priority now is preparing for commercial readiness, launch and real-world deployment.” That transition is difficult because data-centre operators do not purchase components on laboratory performance alone.

New optical systems have to meet standards, survive thermal and operational stress, integrate with existing switching and server architectures, reach viable cost levels, and be available in volumes sufficient for infrastructure deployments measured in thousands of components.

Efficiency becomes a system problem

Solinide’s proposition reflects a broader change in AI infrastructure economics. Early phases of the investment cycle concentrated on obtaining enough processors; the next constraint increasingly concerns how efficiently the complete system surrounding those processors operates.

Network bandwidth, electricity, cooling, memory, and utilisation all determine how much useful computing an operator receives from an expensive cluster. An improvement in processor performance can therefore be undermined if moving information into and out of the hardware consumes too much energy or leaves accelerators idle.

Photonics companies have consequently moved closer to the centre of the AI supply chain. Optical connections have long been essential to telecommunications and data centres, but the bandwidth requirements of AI training are pushing optical technology towards shorter connections between racks, boards, and eventually individual chips.

Solinide says the same technology can address telecommunications and other high-performance computing markets, providing potential revenue outside AI. That diversification may matter because data-centre infrastructure expenditure is large but concentrated among relatively few hyperscalers, equipment manufacturers, and specialist operators whose qualification processes can take considerable time.

The company also faces established optical-component suppliers and a growing collection of silicon photonics businesses pursuing different ways to increase bandwidth while cutting power consumption. Some are integrating optical systems more closely with processors and switches, while others are developing co-packaged optics, alternative transceivers, or different materials.

The €4m round therefore buys Solinide an opportunity to prove that its microcomb technology can move from research into repeatable industrial production rather than scale itself. Cost, reliability, supply, and integration will determine whether the efficiency achieved in controlled demonstrations survives inside working AI infrastructure.

As compute clusters become larger, optical networking is shifting from supporting equipment towards part of the economic limit on AI expansion. Operators can keep installing accelerators, but their value increasingly depends on whether the data required to keep them busy can move without consuming an ever-larger share of the power and capital budget.