Marvell unveils second-generation Structera X memory expanders for AI inference
The new 3406 and 3506 models bring six-channel CXL 3.2 support, up to 9 TB per controller, and inline compression aimed at the AI memory wall
Marvell Technology (NASDAQ: MRVL) has a new answer to one of AI’s least glamorous problems: running out of memory. On October 12, 2026, the company unveiled its second-generation Structera X CXL memory expanders, the models 3406 and 3506.
Marvell says they are the first devices in the industry to offer six memory channels for CXL 3.2. AI models keep getting hungrier for memory, so a bigger plate seems like a reasonable thing to build.
What Marvell actually announced
The headline upgrade is the channel count. The previous generation used a four-channel design, and the new parts move to six, a 50% increase in channels.
Capacity is where the numbers get large. Each controller supports up to 9 TB of DDR4 memory or up to 6 TB of DDR5.
The devices support four-host connectivity over PCIe 6.0 x16 links, with maximum bandwidth of 256 GB/s. Up to four separate host systems can tap the same pool of expanded memory.
The new Structera X parts include inline LZ4 compression, which Marvell says boosts usable memory capacity by 2 to 3 times.
The devices are expected to enter sampling by Q1 2027. Before that, Marvell plans live demonstrations at the OCP Global Summit, running from October 12 to October 15, 2026.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Why AI inference needs this
The target here is AI inference, the phase where a trained model actually answers questions and generates output.
Large language models lean heavily on something called the KV-cache, which holds information about the context the model is working with so it does not have to recompute everything from scratch. Marvell is pitching the new Structera X lineup as a way to meet those rising KV-cache demands.
CXL lets servers attach additional memory beyond what sits directly alongside the processor, and Marvell’s devices are built to comply with the CXL 3.2 version of that standard.
Building on an existing product line
The company started its CXL product line in 2024, with the goal of easing bandwidth and capacity limits in cloud and AI workloads.
The first-generation Structera X models, the 2404 and 2504, have already shipped to hyperscalers. Product updates for that generation are continuing into 2026.
The Structera portfolio covers expansion (the X line), acceleration (the A line), and switching (the S line). The new 3406 and 3506 slot into the expansion side of that lineup as successors to the parts that already made it into hyperscaler deployments.
What this means for Marvell and the AI hardware market
According to the research findings, CXL technology is shifting from evaluation toward production.
The new devices are only expected to sample by Q1 2027, and sampling is an early step before volume production. The 2 to 3 times effective capacity from LZ4 compression will also depend on how compressible a given workload’s data is, so real-world results at customer sites will be the true test.
The OCP Global Summit demonstrations offer an early look, while the sampling window in Q1 2027 will show whether hyperscalers are ready to move from interest to orders.