Marvell VP pushes for DDR4 recycling for use in CXL memory, amid the worst DRAM shortage in years — company introduces three-tier AI memory infrastructure

Memory will account for roughly 30% of hyperscaler capex this year, up from about 8% in 2023 and 2024. Conventional DRAM contract prices rose 90% to 95% in a single quarter, and Meta is already running recycled DDR4 behind CXL across millions of servers, cutting server counts by up to 25% for some inference workloads. Into that market, Marvell has brought a three-tier “AI memory infrastructure” portfolio, announced at FMS 2026 in Santa Clara on August 4 and pitched to EE Times last week, where product marketing VP Khurram Malik put DDR4 reuse at the top of the CXL use-case list.

Only one piece of it is new: the Bravera SC6 PCIe 6.0 SSD controller, which samples in Q4. Structera X has been shipping since 2024, the Structera S switch was announced at OFC in March, and the Photonic Fabric optical memory tier came with the Celestial AI acquisition in February. Meta, the reference customer for the use case Marvell is selling, did it with an ASIC of its own design rather than anything from Marvell.

What’s new (and what isn’t)

The Bravera SC6 is a PCIe 6.0 x4 controller with 16 NAND channels, eight chip enables per channel, a 3,600 MT/s ONFI and Toggle interface, and 12 Arm Cortex-R82 cores plus three Cortex-M7s, according to Marvell’s product blog.

Marvell says it doubles the Bravera SC5, supports NAND from multiple suppliers, and uses a host-managed flash translation layer, meaning cloud operators control write placement, garbage collection, and wear leveling themselves, with KV cache spillover from HBM to SSD as the primary workload. Sampling starts in Q4 2026, which puts drives in 2027 at the earliest on the timeline Tom’s Hardware Premium laid out for the PCIe 6.0 controller field, where Phison and Silicon Motion are chasing the same Gen6 drive launches.

Structera X 2404 and X 2504, the DDR4 and DDR5 expansion controllers, were announced in July 2024, and Marvell’s FMS blog says both are in hyperscaler deployments, with inline LZ4 compression that Malik told EE Times yields roughly 2 to 2.5 times the effective capacity.

The Structera S 30260 switch, announced at OFC in March, carries 260 PCIe 6.0 and CXL 3.x lanes, connects 16 or 32 CPUs or GPUs to up to 48TB of shared memory at 4 TB/s aggregate bandwidth and under 460ns round trip, and begins sampling this quarter. Photonic Fabric came in with the $3.25 billion Celestial AI acquisition that closed in February, with Marvell now describing it as a shared-memory tier reaching up to 50 meters with up to 32 TB of warm KV cache and a claimed 2 to 3 times token throughput inside existing footprints. Marvell also took a $2 billion investment from Nvidia in March, tied to NVLink Fusion, the proprietary scale-up fabric CXL pooling has to sit alongside.

DRAM contract prices have tripled since Structera launched

Conventional DRAM contract prices rose 90% to 95% quarter-on-quarter in Q1 2026, the steepest increase on record for every DRAM category. Q2 added another 58% to 63%, and server DRAM is set to climb a further 13% to 18% in Q3, held in check mainly by the long-term agreements U.S. hyperscalers signed to cap pricing through 2027 and 2028. Server DRAM buyers in the U.S. and China were receiving about 70% of their orders late last year as suppliers diverted wafers to HBM.

Samsung, SK hynix, and Micron have all been winding down DDR4 output since last year, and TrendForce put the legacy memory rally at up to 50% for DDR4 in Q1 alone. A module pulled from a decommissioned 2021 server is now worth a multiple of what it was when Structera X was announced, and the DDR5 that would replace it has roughly doubled in price twice. When Marvell first pitched DDR4 reuse in 2024, it was a sustainability line item, but two years later, the memory crunch has turned it into a capex line item.

Meta’s recycling numbers

Meta’s Vistara ASIC, presented at ISCA 2026 in late June, is a CXL 2.0 Type-3 expander on a PCIe 5.0 x16 link that bridges two DDR4 channels to a host and runs at 128GB per chip using 32GB modules recovered from retired machines. Each MemServer pairs a 158-core AMD EPYC Turin with 768GB of local DDR5-6400 and 256GB of CXL-attached DDR4-2400, and Meta claims idle round-trip latency of around 50ns on the controller path.

The paper states that around 40% of Meta’s fleet is memory-capacity bound, that its servers last three to five years while the DRAM inside them is good for seven to 10, and that the deployment cuts server counts by up to 25% for disaggregated ML inference, average latency by 29% for distributed caching, and job failures by 33%, The Register from the paper ahead of the talk.

Malik told EE Times, “The first and foremost important use case within CXL is the recycling of DDR4,” and Marvell’s FMS release says Structera X was “developed in close collaboration with leading hyperscalers” without naming one. Meta’s paper describes an in-house ASIC, an in-house MemServer chassis, and an in-house Linux page-placement stack, the full path one would take when volume justifies its own silicon, while Astera Labs’ Leo controllers reached Microsoft Azure’s M-series preview in November last year.

CXL’s deployment gap

SemiAnalysis declared CXL dead for AI in March 2024 because Ethernet-style SerDes such as NVLink carry roughly three times the bandwidth per millimeter of die edge as PCIe 5.0 or 6.0, so every CXL lane on an accelerator costs scale-up bandwidth that could have gone to the GPU-to-GPU fabric instead.

Marvell’s CXL pitch accordingly targets CPU-side capacity and KV cache staging rather than the GPU-to-GPU path where you’ll find NVLink Fusion. Yole estimated two-thirds of servers sold in Q1 2025 were CXL-capable and expects more than 90% to be by the end of this year, yet it puts the share of servers actually using CXL at close to zero today and only 13% by 2030.

Marvell’s benchmarks for the Structera S show a 16TB pooled DRAM tier delivering 4.8 times the inference throughput and an 82.7% cut in time to first token in GPU configurations, attributed to keeping KV cache in DRAM instead of recomputing or reloading it. Those are vendor numbers with no disclosed model or baseline, and the switch only starts sampling this quarter. Nvidia, AMD, and the SSD makers all offer their own routes for the same cache spillover.