Memory is becoming the most expensive real estate in the data center. Marvell Technology’s new Structera CXL controllers, announced Tuesday from Santa Clara, promise to squeeze up to 3.64 times more data into every gigabyte of DRAM. For companies running AI workloads, that math changes the economics of computing.
The problem is straightforward. AI models eat memory.
They eat it faster than chipmakers can fab it. Server operators buy more sticks, fill more slots, hit physical limits. Then they buy more servers.
The cost spirals. Marvell’s approach compresses the data while it sits in memory, then decompresses it on the fly when the processor needs it.
The system never knows the difference. What makes this different from existing CXL controllers is the compression itself. Most CXL controllers on the market move data around.
They extend memory pools, they let different processors share memory. But they don’t shrink the data. Marvell built a dedicated hardware block called the Compression-Decompression Block, or CDB, that lives on the controller.
It compresses data as it writes to DRAM and decompresses as it reads. All in real time.
All without touching the host CPU. That last part matters. Software compression works, but it burns processor cycles.
On a system running large language model inference or training, every CPU cycle is already spoken for. Offloading compression to dedicated silicon means the host CPU stays focused on the AI workload.
The compression happens in the background, invisible, at full bandwidth. The immediate consequence is that a server with 256 gigabytes of DRAM can behave like it has 900 gigabytes. Or more, depending on the data.
Marvell’s spec says up to 3.64x compression. That is not a theoretical number. The CDB adheres to specifications submitted to the Open Compute Project, meaning it follows an open standard other companies can adopt.
Marvell claims it is the only CXL memory controller offering inline compression as an in-production feature. For the broader tech industry, the timing is sharp.
Memory shortages are widespread. AI demand is accelerating. Cloud providers, hyperscalers, enterprise data centers — all of them face the same wall.
You cannot just pour more DRAM into a system. Physical slots are finite.
Power budgets are finite. Cooling capacity is finite. Compression lets operators make existing memory work harder.
The two new controllers are called Structera X and Structera A. They are part of Marvell’s larger CXL family. But the real product here is not the chip.
It is the compression block. That block is what lets a data center operator look at a rack of servers and see 3.64 times more usable memory without buying a single additional DIMM.
There are limits. Not all data compresses equally. Random data, encrypted data, already-compressed data — none of it shrinks much.
The 3.64x number is a ceiling, not a guarantee. But for the kind of structured data AI workloads process, the gains are real.
Model weights. Embedding tables. Intermediate activations.
These compress well. The consequences ripple outward. Memory vendors see their product effectively multiplied.
Server makers see a way to pack more capability into the same footprint. AI companies see a path to larger models without proportionally larger hardware budgets.
And the Open Compute Project sees another building block for open, standardized infrastructure. Marvell is betting that the market will pay for compression in silicon rather than leaving it to software. Given the price of DRAM and the hunger of AI, it is a bet that looks safe.
{“@context”: “https://schema.org”, “@type”: “NewsArticle”, “headline”: “Marvell’s New CXL Accelerators Aim to Cut AI Memory Costs”, “description”: “Memory is becoming the most expensive real estate in the data center. Marvell Technology’s new Structera CXL controllers, announced Tuesday from Santa Clara, promise to…”, “datePublished”: “2026-06-27T13:42:00+00:00T09:30:00+00:00”, “author”: {“@type”: “Person”, “name”: “Erika Johnson”}, “publisher”: {“@type”: “Organization”, “name”: “InfoPulse Today”, “url”: “https://newsfeed.infopulsetoday.com/”}}
