Samsung announces a “milestone” in AI. It is … RAM that can be removed

Samsung has just decided to remind the world that in 2025 everything can be a revolution, if only it is talked about long enough and seriously enough. Indeed, the company’s latest press release announces a “fundamental transformation of AI infrastructure,” a “new era of superchips,” and a “milestone for the entire industry.”

The reason?
New RAM format.

LPDDR memory module, which – note – is not soldered permanently.

AI has changed. So the RAM at Samsung must too

At the outset, Samsung reports that the world of AI has undergone a dramatic metamorphosis. We used to train models, today we are “constantly inferring,” and this, as it turns out, changes absolutely everything. The biggest challenge is no longer performance, but energy efficiency, as data centers are beginning to resemble power plants.

And here comes SOCAMM2 – Small Outline Compression Attached Memory Module – a name that is long, serious and sounds exactly as if it is meant to hide the fact that we are talking about RAM in a different package.

Samsung explains that SOCAMM2 combines “the best features of LPDDR” with a “modular architecture,” offering more than twice the bandwidth of RDIMMs and more than 55% less power consumption. Sounds impressive, until you notice that LPDDR has been energy-efficient for years, and the difference is mainly that this time Samsung has allowed it to be disconnected without a soldering iron.

A revolution according to Samsung: memory that can be removed

One of the main “user benefits” is that the SOCAMM2 has a detachable design. Samsung describes this with devout focus, emphasizing easier maintenance and greater expansion flexibility.

That is, exactly what has been standard in the server world… for decades.

But in the world of LPDDR – which is usually soldered – this is indeed a novelty. The problem is that Samsung is communicating this change as if it has discovered a new branch of physics, rather than applying modularity where it didn’t exist before.

It’s a bit like a laptop manufacturer announcing: “We are introducing a battery that can be replaced without taking apart the entire chassis. This is a new era of mobility.”

Horizontal means better (because it does)

Another pillar of the SOCAMM2 revolution is that the memory modules are mounted horizontally, rather than vertically like RDIMMs. Samsung explains that this improves airflow. All of these sound sensible, but again: these are engineering decisions, not a technological revelation. They have, however, been wrapped in language that suggests that without SOCAMM2, data centers simply won’t survive the next wave of AI.

NVIDIA was, she said, blessed

It wouldn’t be a full-fledged AI announcement without the ritualistic invocation of NVIDIA. So Samsung reports “close cooperation,” “continuous optimization” and “ensuring the highest responsiveness.” A quote from an NVIDIA director appears, saying exactly what every director says in every announcement:

• “next-generation data centers,
• “exceptional power efficiency”,
• “complex reasoning and physical AI”.

This is not an accusation – it is simply the liturgy of modern technology PR. Without it, the product does not exist.

Standardization, that is: this is just the beginning

Samsung also adds that SOCAMM2 is so important that the industry is already working on its standardization under JEDEC. Which is true and acutely good news – but again is presented as evidence that we are witnessing the birth of a new era.

But to do justice: SOCAMM2 is not a bad solution. It is a sensible response to a specific market need. It can indeed reduce energy consumption and simplify servicing. But trying to sell it as a “milestone for the entire industry” is simply PR puffery that amuses rather than convinces.