Marvell Unveils Concept of Optically Connected Data Centers Spanning Thousands of Kilometers

Marvell has presented its vision for the future of data centers, where optical connections will enable the unification of computing resources across thousands of kilometers. According to CEO Matt Murphy, current architectures are limited by physical constraints, as processors, accelerators, and memory are tightly bound to local systems due to the limitations of copper cables. Optical technologies will eliminate these barriers, creating a global infrastructure with dynamic workload distribution.

In the near future, Marvell will begin shipping new solutions for inter-data center connections, including Colorz 1600 modules with 1.6 Tbps throughput based on 2-nm DSP. The company also announced Teralynx T100 switches supporting 102.4 Tbps and 512 ports at 200 Gbps. These technologies will allow hyperscalers to synchronize AI clusters, distributed storage, and perform operations without strict latency requirements.

Optical networks will enable the creation of virtual systems with separated pools of CPU, GPU, and memory. This will enhance infrastructure flexibility, allowing operators to assemble configurations tailored to specific tasks while avoiding redundant hardware costs. For example, memory-intensive tasks will no longer require additional accelerators—sufficient HBM capacity can be allocated from a shared pool.

While interconnecting data centers across thousands of kilometers is not suitable for latency-sensitive tasks, Marvell’s technologies will enable hyperscalers to efficiently synchronize infrastructure across continents. This is particularly critical for AI workloads, where data consistency and rapid scalability are essential.