Marvell brings AI memory and interconnect tech to SC26 in Chicago

Marvell Technology is heading to Supercomputing 2026 with a clear message: AI’s biggest problem may not be compute. It may be getting data to the compute fast enough.

The company will exhibit at SC26, which runs November 15 to 20, 2026, in Chicago. Its showcase at booth 4003 centers on memory capacity and optical connectivity, two areas where AI infrastructure keeps running into walls.

What Marvell is putting on display

The headliners are two memory technologies: Photonic Fabric memory modules and Structera CXL memory solutions. Both target the growing appetite for bandwidth in high-performance computing workloads.

Photonic Fabric is Marvell’s approach to shared memory over optical links. The platform supports memory pools of up to 32 TB spread across multiple racks.

Marvell claims Photonic Fabric delivers 2-3x throughput gains for inference workloads. That’s a company claim, not an independent benchmark, so buyers will want to see it hold up in real deployments.

On the CXL side, Marvell recently announced its second-generation Structera X 3406/3506 CXL 3.2 memory expanders. They use six-channel designs, are compatible with PCIe Gen 6, and support up to 9 TB of memory.

CXL, short for Compute Express Link, is a standard for attaching extra memory to processors over a high-speed connection.

Those expanders are expected to sample in the first quarter of 2027. So what visitors see in Chicago is a preview, not something rolling into data centers this year.

The supporting cast

Marvell isn’t stopping at memory. The booth will also feature SSD controllers, DPUs, 1.6T/800G active electrical cables and advanced switching silicon.

Marvell’s Bravera SC6 SSD controllers are expected to sample in Q4 2026. Its Teralynx switches, by contrast, are already deployed in hyperscale AI clouds.

Preet Virk, Marvell’s SVP and GM, is scheduled to speak at SC26 on November 18.

Why memory has become the bottleneck

The timing of this push connects to a specific workload trend. Marvell is aligning these products with rising KV cache demands in agentic AI inference.

A KV cache is essentially a model’s short-term memory during a conversation or task. It stores intermediate results so the model doesn’t redo work each time it generates the next piece of output.

Agentic AI, where models carry out multi-step tasks rather than answering a single prompt, tends to keep that short-term memory busy for longer. The longer and more complex the task, the more memory it can consume.

That is where capacity numbers like 9 TB and 32 TB become relevant. Marvell’s strategy treats memory and interconnect as linked problems, which explains the pairing of CXL expanders with optical links and high-speed cabling.

What this means for Marvell and the AI hardware market

Marvell is trying to establish its CXL-based products and optical architectures as core building blocks for hyperscale cloud and enterprise environments. Key pieces like the second-gen Structera expanders are still months away from sampling.

Sampling is also only an early step. It means customers get chips to test, not that volume orders are locked in.

There are risks worth watching. The 2-3x inference throughput claim for Photonic Fabric needs validation outside Marvell’s own materials. Timelines for Q4 2026 and Q1 2027 sampling could slip, as hardware schedules often do.

The fact that Teralynx switches are already running in hyperscale AI clouds gives Marvell an existing foothold with exactly the customers it wants to sell its memory products to.

The key signals to track after SC26: whether Structera X samples arrive on schedule in early 2027, whether Photonic Fabric’s performance claims show up in customer deployments, and whether any named partners emerge for the shared-memory approach.