Marvell Technology and Penguin Solutions announced a strategic partnership on November 16, 2026, to jointly address AI memory requirements. This collaboration combines two complementary technologies: CXL based memory expansion, which enables data centers to utilize existing DDR4 memory to expand server memory capacity; And advanced photon technology, aimed at achieving high bandwidth, low latency memory sharing across multiple racks.
AI reasoning drives memory demand
As AI workloads shift from model training to large-scale inference and agent-based AI, memory capacity, bandwidth, and latency are becoming key constraints on application performance and token economy. The continuously growing model size, longer context windows, and expanding KV cache requirements are driving the memory demand per node to reach hundreds of GB and increasingly move towards TB level. These demands are accelerating the shift from traditional server attached memory to memory expansion, pooling, and decoupling technologies such as CXL and optical shared memory architectures, enabling memory capacity to be expanded independently of computation.
Two level memory expansion scheme
Marvell and Penguin Solutions are developing two complementary memory expansion technologies for server and pod level AI infrastructure requirements.
Direct Additional Memory Expansion: The plan is to integrate the Marvell Structera X memory expansion controller into the Penguin Solutions SMART CXL add-on card, allowing data centers to recycle existing high-density DDR4 memory modules from retired servers and directly connect them to servers through PCIe 5.0 and CXL 2.0, thereby freeing up additional memory capacity. This method provides scalable, low latency memory expansion in a standard form factor, aimed at wider industry adoption.
Multi rack photonic memory pooling: By integrating memory modules and network cards based on Marvell Photonic Fabric optical interconnect technology into Penguin Solutions Memory AI memory devices and Compute AI systems, AI data centers will be able to support low latency, high bandwidth, multi rack decoupled memory sharing across longer distances, enabling PB level memory pooling for context rich proxy AI inference.
Upgrade path from direct attachment to multi rack sharing
The two technologies are highly complementary. Directly adding memory expansion solves today’s server level instant capacity requirements, while multi rack photonic memory sharing aims to achieve high bandwidth, low latency memory sharing across multiple racks, supporting increasingly demanding AI workloads. Both methods work together to provide customers with greater flexibility, expanding memory capacity and AI infrastructure as demand evolves, offering a single upgrade path from direct add-on expansion to decoupled multi rack sharing without the need to re architect infrastructure at every step. The additional advantage of the standard AIC appearance is the support for memory expansion beyond direct server connections. The Penguin SMART CXL add-on card driven by Structera X can be used to build dedicated CXL memory devices, allowing operators to pool and recycle DDR4 capacity for decoupled allocation across multiple hosts, not limited to a single server.
Product availability time
The Penguin SMART CXL add-on card is driven by the Marvell Structera X memory expansion controller and is expected to be available for certification and enterprise customer deployment by the end of 2026. The joint development of memory expansion servers integrated with Photonic Fabric is expected to produce samples in the second half of 2027. Marvell will showcase relevant information at booth 4003 at SC26 (McCormick Place, Chicago, November 16-20).
Jihoun Han, Senior Technology and Market Analyst at Yole Group, stated that the demand for AI memory is growing rapidly, and adding new memory capacity requires time and significant investment; The rise in memory prices has also increased the cost of expanding AI infrastructure. This requires a solution to bridge the gap by better utilizing existing memory resources. Memory expansion and pooling, combined with optical interconnect to help manage latency during cross rack expansion of memory pools, provide a promising option for meeting growing capacity demands and managing costs in the coming product cycles.
Preet Virk, Senior Vice President and General Manager of Marvell’s Photonics Fabric Business Unit, stated that in order to expand AI inference, tokens need to be delivered with higher economic efficiency; By combining CXL based memory expansion with an optical memory sharing roadmap, Marvell and Penguin Solutions provide a practical path for CIOs and AI data center leaders to improve accelerator utilization, directly increase token revenue, alleviate HBM and DDR5 supply constraints, and build a more flexible AI memory architecture without the need for overall infrastructure replacement.
Kash Shaikh, President and CEO of Penguin Solutions, stated that as inference and agent-based AI workloads accelerate in data centers, memory is becoming increasingly critical to the performance and economy of AI factories; The collaboration with Marvell is advancing its memory innovation roadmap based on CXL memory and advanced photonics solutions to address memory bottlenecks.
About Marvell
To deliver data infrastructure technology that connects the world, Marvell is building solutions on the strongest foundation: partnerships with customers. For over 30 years, Marvell has been trusted by world leading technology companies to move, store, process, and protect world data through semiconductor solutions designed for customers’ current needs and future goals. Through deep collaboration and transparent processes, Marvell is changing the way enterprise, cloud, and carrier architectures transform tomorrow.
About Penguin Solutions
Penguin Solutions is an AI factory platform company. The company designs, builds, and manages next-generation data centers for enterprises, sovereign AI initiatives, and AI cloud providers. With deep design expertise in the intersection of data center AI infrastructure and memory solutions, Penguin’s Full Stack AI Factory platform combines differentiated infrastructure software, advanced memory, computing systems, end-to-end services, and industry-leading partner technologies to help customers accelerate deployment, optimize token economy, and maximize AI investment returns.



