Marvell has acquired Swiss photonics specialist Polariton Technologies, adding plasmonics-based optical modulation to its data center connectivity portfolio as hyperscale customers prepare for 3.2-terabit links. Terms were not disclosed.
The strategic logic is straightforward. AI clusters are driving a sharp increase in east-west traffic inside and between data centers, forcing network architects to rethink how quickly optical interconnects can scale without blowing out power budgets. Marvell already sells DSPs, electro-optics, switching silicon, and custom infrastructure chips. Polariton fills a device-level gap. Its technology aims to improve how optical signals are modulated, a critical function as the industry moves beyond today’s 800G and 1.6T deployments toward denser, lower-power interconnects.
That matters because optical performance is becoming a system constraint. Training and inference clusters need faster scale-across links. Data center interconnect demand is rising as operators spread AI workloads across campuses and regions. In both cases, the challenge is the same. Move more data with less energy and less space.
Polariton’s pitch is that plasmonics can outperform conventional silicon photonics on density and energy efficiency, particularly in compact coherent modules and DCI applications such as ZR and ZR+. For Marvell, the acquisition is less about buying revenue than securing a differentiated building block and the engineering team behind it. This is a technology tuck-in designed to protect roadmap control.
The broader industry context is consolidation around the optical stack. As AI infrastructure budgets expand, chip and networking vendors are racing to own more of the interconnect bill of materials, from DSPs and lasers to packaging and photonic integration. Vertical control offers two advantages. It can shorten development cycles and improve margin capture in a market where performance gains are increasingly won through co-design.
The open question is execution. Advanced photonics acquisitions often hinge on whether promising lab-scale performance can be translated into manufacturable, high-yield products at hyperscale volumes. Still, Marvell is making a clear bet that the next competitive frontier in AI infrastructure sits between the GPU racks, not inside them.
Source: Company press release and Acquire.fyi's proprietary data
