Without relying on stacking hardware, AMD acquires MEXT to solve the AI memory shortage.

This week, American chip giant AMD announced the acquisition of a startup called MEXT. The business of this company is very interesting. They are trying to break down the "DRAM memory wall" in front of the AI industry with a pure software solution.

Since 2026, DRAM has been becoming a strategically scarce resource, and there is no solution in the short term. According to data from institutions such as TrendForce, HBM memory has already consumed about a quarter of the global DRAM wafer production capacity. The price of general - purpose server DRAM has soared by about 3.5 times in the past two quarters, and the cost of memory alone accounts for more than half of the hardware cost of a server.

What's more troublesome is that even if companies bite the bullet and buy enough DRAM, the unique operating mode of large AI models will cause a large number of memory pages to be actually in a cold state, but the traditional memory management mechanism will not automatically move them out. This results in companies spending a large budget on upgrades, while the actual utilization rate may not increase much.

Compared with replacing more expensive CXL memory modules, MEXT has developed a pure software path. Their core product is called the predictive memory engine, which is essentially an AI - driven memory hierarchical manager operating at the system level. It can solve the problem of insufficient DRAM in an extremely flexible way.

Generally speaking, the system will first continuously scan the memory access behavior of the workload, identify the cold memory pages that are not currently in active use, and then unload these pages to NAND flash memory. Immediately afterwards, MEXT's AI engine comes into play. This engine can predict which of the unloaded pages will be used by learning the behavior pattern of the workload. Then, before the actual request occurs, the engine will "actively" retrieve those pages back to DRAM.

In other words, just as a large language model can learn to predict the next word in a sentence, MEXT's AI engine can predict the next page in memory. When the pre - fetch hit rate is high enough, data can seamlessly "shuttle" between DRAM and flash memory.

MEXT claims that this technology only needs to run on a single CPU core for both training and inference. It can complete predictions within microseconds and continuously optimize itself. It also fully supports local deployment or cloud environments.

Most importantly, it does not require changes to existing hardware and can be installed within five minutes. This feature enables MEXT to claim that it can reduce infrastructure costs by 50% and increase the available memory capacity by 2 to 4 times. Currently, several companies have implemented applications in real - world scenarios.

Back to AMD, what core problems can the acquisition of MEXT solve?

Firstly, the most direct purpose is to reduce costs. AMD's data center business has long revolved around EPYC processors and Instinct accelerators. However, in the entire solution for cloud service providers, the powerful performance of the chips is accompanied by huge memory overhead. At this time, if MEXT's predictive memory engine can be integrated into AMD's data center product portfolio, it means that customers can run larger models with the same configuration, which is extremely attractive to hyperscale cloud providers such as Amazon and Google.

Secondly, MEXT fills in the missing memory scheduling logic in AMD's full - stack closed - loop. This itself requires a team that understands both the memory system architecture and the working principle of AI to achieve. The addition of MEXT just fills this gap. The technical accumulation and innovation ability of this team will help AMD continuously overcome various problems in the memory architecture of modern data centers and provide talent support for subsequent technological iterations.

Looking at the entire industry, AMD's acquisition of MEXT sends a clear signal - in the trend of limited DRAM production capacity, the model of simply relying on hardware stacking to increase memory capacity has reached a bottleneck. Software optimization technologies such as AI - based memory hierarchy and intelligent scheduling are becoming the mainstream direction for the industry to break through the situation.

Different from the heavy - asset model of traditional hardware upgrades, pure software solutions like MEXT have the characteristics of light weight, high adaptability, and low cost, and can quickly cover the server market. This is also an important reason why the capital market highly approves of this acquisition. Therefore, after the news spread, AMD's stock price soared.

Of course, MEXT's solution is not a panacea. For unpredictable scenarios, its benefits may be discounted. So, as AMD has officially launched the integration work of MEXT technology, it will continue to optimize the functions in the future.

This article is from the WeChat official account "Meku.com". Author: Meku.com. Republished by 36Kr with permission.