Marvell Technology has launched a new portfolio of memory solutions for AI infrastructure, covering server-class AI storage, rack-scale Compute Express Link (CXL) memory expansion and pooling, and multi-cabinet shared memory over optical interconnects. The company said the offering is aimed at addressing rising memory capacity and bandwidth constraints in Agentic AI inference.
Marvell said larger AI models, longer context windows, and growing demand for KV Cache are exposing the limits of traditional architectures that keep compute and memory tightly coupled. By using memory disaggregation, the company said, memory resources can scale more independently from compute resources, improving GPU utilization while cutting data movement latency.
The products announced
The new lineup includes three parts:
- Bravera SC6 PCIe 6.0 SSD controller: built for AI inference storage workloads and designed to help cloud service providers move more KV Cache onto high-performance SSDs to improve infrastructure efficiency. Marvell said the controller uses an architecture compatible with NAND from multiple suppliers and is expected to begin sampling in the fourth quarter of 2026.
- Marvell Structera X memory expansion solution: based on CXL technology, it supports rack-scale memory expansion and resource pooling, allowing data centers to share and allocate memory resources more flexibly while lowering AI infrastructure costs.
- Marvell Photonic Fabric optical memory solution: uses optical interconnect technology to build a shared memory architecture across multiple cabinets. Marvell said it can support up to 32TB of warm KV Cache offload and help AI inference clusters raise throughput.
Photonic Fabric performance claims
Marvell said the Photonic Fabric offering can deliver up to 2x to 3x higher token throughput within existing data center space and power constraints, while supporting larger models and AI applications with longer context windows.
Executive comment
Will Chu, a Marvell executive, said AI infrastructure is moving away from single-server architectures toward systems in which compute, memory, and connectivity operate together. In his view, memory will need to scale more independently in the future to improve resource utilization and token efficiency.
Marvell also said that as demand keeps rising for AI agents and large-model inference, memory capacity, bandwidth, and data transfer efficiency are becoming a new competitive focus in AI infrastructure after compute.