Marvell says it will demonstrate a new set of 2nm optical interconnects at ECOC 2026 in Málaga, Spain, as the optical-communications industry gathers to address a less visible constraint in the AI buildout: moving data between increasingly dense pools of compute.
The company’s September 21 announcement covers four demonstration areas: a 400G-per-lane optical PAM4 link, an 800G ZR/ZR+ pluggable with integrated MACsec security, 1.6T ZR and coherent-lite O-band technologies, and a 102.4-terabit co-packaged-optics platform. Marvell will show the portfolio at Booth 2002 during the exhibition. The company describes several of the demonstrations as industry firsts; those claims are Marvell’s and are not independent performance results.
The timing is relevant. ECOC Exhibition’s official day-one guide says more than 340 exhibitors are in Málaga this week, with sessions focused on optical circuit switching, co-packaged optics, 1,600G-plus components and AI interconnects. Those are not peripheral technologies in modern AI systems. As clusters add accelerators and distribute work across racks and facilities, the network can become a practical limit on bandwidth, power and system design.
Why 2nm optical interconnects matter
Inside an AI data center, a model workload does not stay on one processor. Training and inference systems move data among accelerators, processors, memory and storage; larger deployments also connect racks and sites. Each step raises the importance of the links that carry that traffic. Marvell’s basic proposition is that moving to a 2nm process node can support faster optical connectivity while managing power and density.
The technologies in this announcement address different parts of that network. PAM4, short for pulse-amplitude modulation with four signal levels, is used for high-speed links inside data-center fabrics. ZR and ZR+ refer to coherent pluggable optics designed for data-center interconnects over longer distances. MACsec is a link-layer security standard; its inclusion in the 800G demonstration is therefore about protecting traffic in transit, not a claim that the component solves wider data-center security risks.
The 102.4T co-packaged-optics platform is the longer-horizon part of the presentation. In co-packaged optics, optical components are placed very close to switching silicon or accelerators, reducing the electrical distance that signals need to travel. Marvell has previously argued that this can improve the power budget as copper links lose reach at higher speeds. But the approach also creates difficult questions about cooling, fibre management, repairability and manufacturing. The company’s own explanation says commercial viability depends on the use case and broader ecosystem readiness.
A showcase, not a rollout
The announcement is a roadmap and demonstration update, not evidence that the full set of products is shipping at volume. It does not disclose power-consumption measurements, pricing, yields, customer deployments or independently verified throughput. That distinction matters when a press release connects a component roadmap to the enormous capital spending now associated with AI infrastructure.
Marvell did provide useful context earlier this year. In March, it said its Libra 2nm coherent DSP and related COLORZ products were expected to begin sampling in the second half of 2026. The September release now says that the 800G ZR/ZR+ demonstration uses Libra and includes MACsec, but it does not update the sampling schedule or provide a commercial-availability date for the technologies being shown in Málaga.
For readers following the infrastructure stack rather than only the processor race, this is the key point. Networking is becoming a design constraint alongside chips, memory, power and cooling. Unhyd’s reporting on Qualcomm and Amazon’s AI-infrastructure collaboration described the same shift from a different angle: compute, optical connectivity and deployment economics are increasingly negotiated together.
What to watch after ECOC
The next meaningful evidence will be concrete rather than promotional. Watch for published power-per-bit figures, interoperability results, sampling or production dates, named systems partners and actual data-center deployments. It will also matter whether co-packaged optics moves beyond tightly engineered demonstrations into systems that operators can build, cool and service reliably.
Marvell’s ECOC demonstrations do not settle those questions. They do make the competitive direction clearer: as AI clusters grow, advances in semiconductors will be judged not only by what a processor can compute, but by how efficiently the rest of the system can move the data around it.
