GPUs sit idle waiting on memory more often than most cloud operators say out loud. Marvell used the Flash Memory Summit on August 4 to put a number on the fix: three products aimed at letting memory scale independently of the server it once had to sit inside, so a cluster can add capacity without adding another accelerator card next to it (Marvell, 2026).

The lineup covers three altitudes. The Bravera SC6 is a PCIe 6.0 SSD controller built to move key-value cache data out of high-bandwidth memory and onto flash without stalling the GPU behind it. Structera X expands memory at the rack level over Compute Express Link, an interconnect standard hyperscalers call CXL. Photonic Fabric goes furthest: an optical layer that lets processors reach a shared pool of up to 32 terabytes across racks up to 50 meters apart, treating memory as a utility rather than something bolted to a specific server.

Memory Is the New Constraint

Agentic AI workloads keep a running memory of every tool call, retrieved document and intermediate reasoning step in a session. That key-value cache grows with context length, and it grows fastest in exactly the workloads enterprises are shipping now: agents that hold state across a long task instead of answering one prompt and forgetting. Hyperscalers sized server-attached memory for shorter conversations. It runs out first, long before the GPU next to it does.

Marvell treats that as an architecture problem. Detach memory from compute, let cloud providers scale each independently, and a GPU that would have stalled waiting on local memory instead reaches into a shared pool sized for the job.

Will Chu, who runs Marvell's custom cloud solutions group, frames this as infrastructure moving past isolated servers toward systems where compute and memory operate as one connected pool rather than fixed pairs (Marvell, 2026).

Marvell Is Repeating the Interconnect Playbook

Readers of this site have seen this pattern before. NVIDIA's $2 billion stake in Marvell in April bought into interconnect capacity through a minority investment. The optical interconnect race that followed, most recently Lumilens raising $700 million this month, has been about getting data between chips faster as copper runs out of headroom inside the rack.

Photonic Fabric extends that same logic to memory itself. Once light carries data between racks instead of just within them, the physical boundary of a server stops defining what a GPU can reach. Compute and network already ran on this premise. Memory joins them now: distance inside a data center is a bandwidth problem before it is a cabling problem.

Unknowns and Uncertainties

Structera X sits on CXL, an open standard multiple vendors support, which limits lock-in. Photonic Fabric runs on Marvell's own optical architecture, and Marvell has not said whether other vendors' processors or network interface cards can address that shared memory pool without Marvell silicon in between. If no other vendor's silicon can read that pool, the arrangement functions as a new form of lock-in dressed in disaggregation language.

Marvell has not disclosed pricing for any of the three products, and the Bravera SC6 will not begin sampling until the fourth quarter of 2026, which likely pushes production deployment into 2027.

Sources: Marvell Technology, Inc. "Marvell Advances AI Memory Infrastructure Portfolio to Accelerate Agentic AI Inference." Marvell, 4 Aug. 2026, www.marvell.com.