IN Brief:
- The $3.4 million programme targets external InP laser sources, DFB arrays, and semiconductor optical amplifiers for AI data-centre interconnects.
- Development centres on optical power, wavelength density, efficiency, and thermal stability as optics move closer to high-power compute silicon.
- Customer sampling and early production ramps are targeted for the second half of 2027.
Sivers Semiconductors and SemiNex have started a $3.4 million development programme for indium phosphide light sources intended for optical interconnects in AI data centres, with customer sampling and early production ramps targeted for the second half of 2027.
The programme covers high-power external laser sources for co-packaged optics, high-channel-count distributed feedback laser arrays for wavelength-multiplexed links, and semiconductor optical amplifier stages intended to extend optical reach as channel counts increase. The companies are concentrating on the light-generation layer rather than the switching or compute silicon, where higher aggregate bandwidth is increasing the demands placed on optical power, efficiency, and thermal stability.
Co-packaged optics moves optical interfaces closer to high-bandwidth switch and compute devices instead of relying entirely on pluggable modules at the faceplate. Shorter electrical paths can reduce the power required to drive very high-speed signals across a board, but the approach shifts more of the optical and thermal burden towards packages already occupied by high-power silicon.
The external laser source consequently has to maintain controlled optical output across several wavelengths while operating within a restricted thermal envelope. Distributed feedback lasers provide wavelength-selective sources that can be arranged into arrays for wavelength-division multiplexing, while semiconductor optical amplifiers add gain where the optical link budget becomes tighter.
Separating the laser from the hottest part of the co-packaged assembly can also give designers more freedom over cooling and serviceability. That separation does not make the source simple: coupling efficiency, spectral stability, output power, device ageing, and thermal drift still have to remain predictable across the complete operating range.
Sivers and SemiNex are therefore developing more than a higher-power laser die. As optical channels multiply, the source becomes part of the system power budget, and small losses in wall-plug efficiency are multiplied across large numbers of wavelengths and links. The thermal load created by the light source also arrives in an environment where accelerator, switch, and memory devices are already demanding aggressive cooling.
Sivers has been extending its semiconductor position across both optical and high-frequency communications. The company recently secured an $8.2 million Ka-band beamforming IC production order for satellite terminals, while the SemiNex programme adds a separate path into optical data-centre infrastructure.
SemiNex contributes an InP portfolio covering high-power laser diodes, DFB lasers, gain chips, semiconductor optical amplifiers, and external-cavity laser technology. That device range allows the programme to address arrays and gain stages alongside the underlying light source, rather than treating each optical element as an unrelated component.
The H2 2027 sampling target leaves a substantial qualification period. Before customer deployment, the partners will need to characterise optical power, wavelength accuracy, efficiency, ageing, thermal behaviour, packaging, and manufacturing variation, then demonstrate that the resulting performance can be repeated across production lots.
Data-centre qualification adds sustained utilisation to that list. Optical components supporting AI clusters are expected to operate continuously under high traffic loads, so laboratory peak performance is of limited value if wavelength stability or efficiency deteriorates under prolonged thermal stress.
Co-packaged optics has progressed far enough that bandwidth alone is no longer the awkward part of the specification. Producing the necessary light efficiently, predictably, and at sufficient volume is becoming a manufacturing problem as well as a photonics problem. Sivers and SemiNex are aiming to put their InP technology into that gap before customer sampling begins in 2027.
