AI infrastructure is increasingly constrained by how well memory, compute and networking can scale together beyond a single rack. Marvell Technology (NASDAQ: MRVL) is using Supercomputing 2026 to demonstrate interconnect, switching and memory technologies that it says address those bottlenecks as enterprises and cloud operators build larger AI and high-performance computing systems.

The company’s focus is on three architectural domains. In scale-up environments, optical interconnects can extend memory and compute across multiple racks. Scale-out networks rely on switching, pluggable optics and electrical interconnects. Scale-across links connect data centers through coherent pluggables. Marvell’s message is that these layers can no longer be treated as isolated components when workloads require more memory capacity and bandwidth than a single rack can provide.

At its SC26 booth, Marvell plans to show its Photonic Fabric memory module, CXL memory products, Structera add-in cards, near-memory compute and expansion or compression technologies. The company also listed partner demonstrations using Structera X memory expansion controllers with DDR4 and DDR5 memory. These products are positioned around CXL, the industry interconnect approach intended to give processors access to memory and accelerators beyond their directly attached resources.

For IT leaders, CXL is relevant because larger AI workloads often expose a mismatch between available compute and available memory. Adding memory capacity is not automatically useful if data cannot move through the system fast enough or if the topology introduces operational complexity. Marvell has not published customer performance results in this announcement, but the demonstrations signal how hardware suppliers are trying to close that gap with memory pooling, expansion and near-memory processing designs.

Networking is the other half of the story. Marvell said it will demonstrate multipath reliable connection using Teralynx 10 switch silicon, a data-center simulator using Teralynx T100, and active electrical cables based on Alaska AEC digital-signal processors at 1.6T and 800G speeds. It also plans to show virtual rack configurations built with Teralynx T100 switch trays and CXL solutions. The descriptions point toward systems in which bandwidth planning begins at the rack rather than only at the server.

The company also listed Bravera SSD controllers and OCTEON data-processing-unit-based storage solutions. Those technologies broaden the announcement beyond networking: storage paths, memory access and packet processing all affect how efficiently an AI factory or HPC cluster can keep expensive compute resources supplied with data. The source does not establish a specific utilization improvement, so enterprises should treat the demonstrations as product positioning rather than a measured return on investment.

SC26 runs November 15 through 20 at McCormick Place in Chicago, and Marvell will also participate in a panel on photonics for AI and HPC clusters. The timing reflects an industry shift from selecting a single accelerator to designing an integrated fabric around accelerated workloads. Older architectures could often scale by adding more servers; the next constraint is increasingly the links and memory tiers between them. Marvell’s competitive test is whether its components can help customers expand those domains without creating a more fragmented infrastructure stack. Procurement teams will also need to evaluate interoperability, management tooling and support boundaries alongside raw bandwidth specifications.