# Marvell Targets 2027 Customer Sampling for Optical AI Memory Module By Simon Yoon Canonical URL: https://www.tokenpost.com/news/technology/35628 Published: 2026-10-20T12:05:34.000Z Updated: 2026-10-20T12:05:34.000Z Section: Technology > The proposed system would share 32 TB of DDR5 among up to 16 hosts. Its 1.125 TB of HBM3E could serve as a cache or as additional pooled memory. Marvell Technology plans to sample an optical memory module with multiple customers in 2027, advancing a design that would let AI servers share memory across racks for inference workloads. Marvell displayed the module at its Oct. 6 investor day, after introducing the shared memory architecture in August. The proposed system is intended to give processors access to a larger pool of memory when the capacity attached to any one processor is insufficient. A proposed appliance would use 16 modules to share 32 TB of DDR5 among up to 16 hosts. It would also contain 1.125 TB of HBM3E, which could serve either as a cache or as additional pooled memory. Each module is designed for 72 GB of HBM3E, up to 2 TB of DDR5 and 7.2 terabits per second of optical fabric bandwidth. The shared tier targets key-value caches, the working data used during AI inference. As models and conversations grow, those caches can exceed the memory available beside an individual processor. Optical connections would allow processors to reach the shared pool across server racks. The architecture has a stated reach of up to 50 meters. A technical configuration describes rack-to-rack memory sharing across up to 30 meters and claims memory access latency under 350 nanoseconds. Those distances describe different aspects of the developing architecture, rather than a single established operating limit. Marvell projects up to two to three times higher token throughput within existing data center space and power limits. That figure is a projection for the proposed architecture. A separate evaluation used hardware emulation and AI-serving simulations, including simulations of time to first token; it did not demonstrate that throughput multiplier in a production installation. “As AI scales, memory must scale more independently of compute,” Marvell Executive Vice President and General Manager of Custom Cloud Solutions Will Chu said in August. Marvell acquired Photonic Fabric developer Celestial AI on Feb. 2. The optical module builds on Marvell’s work in AI memory expansion, which also includes its [Structera X CXL controllers](). The next announced step for the module is customer sampling in 2027. ## Links in this article - [Structera X CXL controllers](https://www.tokenpost.com/news/technology/30516)