Share
AMD has a new pitch for data center buyers: keep almost all of your server’s performance while buying half the memory.
The catch is a software license. The payoff, according to AMD, could be server costs falling by as much as 83 percent.
The technology comes from Mext, a startup AMD acquired on June 15, 2026. Mext’s software uses AI models to predict which chunks of memory a server will need next. It then shuffles data between expensive DRAM and much cheaper NVMe flash storage before the operating system or applications notice anything happened.
AMD’s headline claim is that Epyc servers running the software can replace half their DRAM with flash and still hold onto up to 99% of their performance. Memory is one of the priciest line items in a modern server, so a claim like that tends to get attention.
How Mext’s predictive memory works
Traditional computers already have a fallback when memory runs short. Linux calls it swap, Windows calls it a page file, and both move data to slower storage when RAM fills up. The problem is that this usually happens reactively. By the time the system realizes it needs a page of data, the user is already waiting for it to come back from disk.
Mext’s approach is to use AI models to anticipate those requests. Pages get loaded back into DRAM before they are called for, so the slower flash tier stays mostly invisible to the workload. The startup calls this Predictive Memory.
The economics are the real selling point. NVMe flash is about 50 times cheaper per bit than DRAM and draws roughly 30 times less power per bit.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
We respect your privacy. Unsubscribe anytime.
The numbers AMD is putting on the table
Gary Smerdon, Mext’s former CEO, now serves as AMD’s corporate vice president of business development. He is the one making the case that the technology could trim enterprise server costs by as much as 83 percent, largely by swapping costly RDIMMs, the registered memory modules used in servers, for flash.
AMD has also started testing the idea on its own infrastructure. An initial internal deployment spanned 86 to 145 hosts and added around 58 terabytes of usable capacity.
That rollout cost approximately $601,000. AMD says getting the same capacity through equivalent DRAM expansions would have cost $3.3 million, which works out to savings of 82%.
The company plans to push further. AMD is targeting cumulative savings that could exceed $60 million by the end of 2027 as it scales the technology across thousands of servers. That is a target, not a result, and it depends on the rollout going as planned.
On the performance side, AMD’s figures suggest mixed setups combining DRAM with Mext software improve performance per dollar by 1.9 to 3.5 times. Physical DRAM can drop by more than 50% while performance stays in the 95% to 99% range on certain workloads.
Why memory has become the pressure point
Mext is a young company. It was founded in 2023 and only launched its Predictive Memory software in April 2026, which means AMD bought it roughly two months after its flagship product hit the market.
Memory capacity is increasingly a bottleneck for AI and high-performance computing. Large models and data-hungry workloads want enormous pools of fast memory, and DRAM prices have been climbing alongside that demand.
