Marvell's Optical Memory Pool Gives 16 GPUs 32TB of Shared RAM

Breaking the GPU memory tax with light

Optical fabrics let datacenters scale capacity without buying extra accelerators.

In modern AI clusters, High Bandwidth Memory (HBM) sits barely 1.5 millimeters from the processor die. If a large language model outgrows that local pool, operators face a painful reality: the only way to add capacity is to buy more GPUs. That effectively turns half-million-dollar accelerator racks into the world's most expensive memory controllers.

Marvell wants to break that link. Built on foundational designs from Celestial AI, the company's Photonic Fabric Memory shifts memory into an independent box across the row, replacing lossy copper traces with high-bandwidth optical fiber. Rather than running out of room during massive multi-turn agent chats or reasoning workloads, accelerators tap into a communal external reservoir.

The targeted pain point is the key-value (KV) cache. When context windows balloon, servers either evict old tokens and recompute them later or page them out to sluggish storage tiers. A shared optical pool lets one processor read the pre-computed context of another across nodes without stalls.

Marvell Photonic Fabric Memory Appliance specifications

Inside the 32TB appliance architecture

The marquee 32TB figure does not belong to a single card, but rather to a complete 16-module Photonic Fabric Memory Appliance. Each individual sled carries a dedicated Photonic Fabric ASIC, 7.2Tbps of optical I/O, 72GB of HBM3e, and eight standard DDR5 RDIMM slots.

Populating each module with 256GB sticks provides 2TB of DRAM per sled. When 16 modules are joined together via a passive all-to-all fiber shuffle, the appliance yields 32TB of pooled DDR5 backed by 1,152GB of HBM3e running as a write-through cache in front of the DRAM.

Connecting up to 16 hosts over fiber avoids the thermal and spatial limits of co-packaged sockets, but the approach arrives during an awkward economic cycle. With enterprise DDR5 RDIMM prices hovering near 5.5 times their pre-crisis levels in the US, filling 128 DIMM slots with high-density modules will carry a steep bill.

Nvidia plays both partner and rival

Marvell is not operating in a vacuum, and Nvidia has prepared for this shift. In March 2026, Nvidia took a $2 billion stake in Marvell focused on NVLink Fusion and photonics. Yet Nvidia continues building rival tiers for context-heavy workloads.

At GTC 2026, Nvidia introduced BlueField-4 STX, a flash-powered context tier aimed squarely at offloading the KV cache from primary accelerator memory. Meanwhile, the upcoming Vera Rubin NVL72 rack integrates 54TB of LPDDR5X across its unified footprint, outstripping Marvell's pooled appliance in pure raw volume. High-end clusters like the banned Nvidia B300 deployments highlight just how desperately datacenters need scalable memory architectures.

Availability and rollout timeline

Enterprise buyers will have to wait before slotting optical memory appliances into production racks. Marvell confirmed that the technology remains pre-silicon hardware, with customer sampling slated for 2027.

Commercial pricing has not been disclosed, and total deployment expenses will remain heavily tied to broader DDR5 component pricing across enterprise distributors in North America and Europe.

Optical fabrics could finally loosen the grip that co-packaged memory holds over AI capital budgets. Whether Marvell can turn its promising physical modules into an industry standard before proprietary interconnects dominate the datacenter will be the critical test over the next two years.