Memory Panic,AI Surge,Supply Shock

● Vera Rubin Memory Panic

Memory shortage seen in the AI data center ‘VeraRubin’: the key to ending the misconception of “cutting capacity in half”

Three conclusions you must read today (from here, grab the key takeaway)

“VeraRubin memory 1.5TB → 768GB (cut in half)” If you stop there, you’re done—but in reality, it’s interpreted that the total memory requirements actually increase by around 20% due to changes in the SOCAM (SoC-Aware Memory) structure.
The reason isn’t just specs; demand is structurally growing through the CPU/GPU ratio, data center design, and AI agent/long-context/video memory.
And based on OpenAI’s feature flow like Dreaming and Chronicle Research, it’s highly likely that memory demands will grow exponentially as it moves beyond “storing text” into image/video and contextual reasoning.

This article isn’t about using “the noise of falling stock prices” to make blanket claims about the memory industry; instead, from the perspective of data-center structure + an AI functionality roadmap, I’ll sort it out like news—on why memory cycles can’t help but keep running.

1) Fluctuations in memory stocks in the US/Korea market, and the starting point: the ‘VeraRubin memory cut in half’ report

News flow

Recently, as memory semiconductor-related stocks have been wobbling in the market, a report stating that “the memory going into VeraRubin is decreasing” has spread as material for the stock price drop.

Key claim (the frame the market adopted)

The memory capacity that will go into Nvidia’s next-generation data center platform “VeraRubin” will
shrink from 1.5TB to 768GB
meaning it would be interpreted that memory demand will decrease

But a rebuttal point has emerged

It’s said that within the side that wrote the report, a rebuttal appeared with the intent of “don’t use it too sensationally.”
The viewpoint was that comparing only “capacity numbers” can lead to misunderstanding.

2) Even if capacity drops, total memory volume can increase: the SOCAM structure is changing

What kind of system is VeraRubin? (structure summary)

With VeraRubin, there appears to be a direction where GPUs/CPUs are configured at large scale, and the CPU-to-GPU ratio is designed to be nearly close to 1:1.
This is connected to why people say, “CPU is important for running agent AI.”

The important point here is that “Nvidia also puts the CPU in”

There’s mention of a configuration where Vera CPU (a CPU for agent orchestration purposes) is directly mounted on the rack.

Meaning of SOCAM

SOCAM is fundamentally described as a structure for integrating, stacking, and connecting LPDDR (low-power mobile/laptop memory) modules in a module-like way.
In other words, instead of “fixing memory on the board by soldering,” the space efficiency arises from the ability to turn the SOCAM framework and insert more modules into the available empty space.

To sum up

If you look only at a simple spec comparison like “192GB drops by 50%,” it can look like memory shrinking.
But because the number of 96GB SOCAM modules increases, overall you could end up with calculations that
more LPDDR 5X volume may actually be needed.

3) A calculation that flips the “cut in half” misconception: an additional 10–20% of memory might be possible

The point that unsettled the market

The “capacity decrease news” that claims 1.5TB → 768GB was consumed too strongly.

Core of the reinterpretation

If you analyze the original report text more deeply, it’s argued that it’s not that capacity drops by half, but
that it may be calculated as an additional need of around 10% to 20%.

Points investors should remember (checklist)

“Capacity numbers” and “total module/total slot/total channel volume” may differ.
In data centers, during the design optimization process, to match space efficiency, power, and bandwidth, the memory configuration can change.

4) Why Nvidia can’t easily reduce memory: Jensen’s GTC message + data center demand pressure

Market logic

There’s mention of an existing flow where Nvidia was asked at GTC (Nvidia’s event) to further increase DRAM/HBM capacity (supply volume).

The claim that “cutting the volume in half doesn’t make sense”

If demand pressure is strong, the viewpoint is that the platform is more likely to “deliver more efficiently through structural optimization” rather than actually reducing the memory it needs.

Reference (interpretation)

The purpose of the visit to Korea may be corporate promotion/event-type, but more important than that is the bigger picture of “why supply/investment pressure is increasing.”

5) When AI functions change, memory demand grows too: Dreaming + Chronicle (storing video/image context)

The signal sent by OpenAI’s ‘Dreaming’

Existing: when a conversation starts anew, it operated close to initialization/reset.
Change: it’s evolving to remember users’ conversations for longer and in a broader way.

‘Quality metrics’ like improved memory accuracy

There are references to numbers suggesting that memory accuracy is improving significantly over time.
This also means that the structure for maintaining, searching, and using memories has become important—not just “memory capacity.”

The more frightening part of Chronicle (Chronicle Research)

It involves capturing the screen and
saving/assembling the real-time context in the form of images/video, and then
helping the AI understand much better when the user later provides prompts.

Conclusion here (from a memory perspective)

Text is relatively light, but video/image/multimodal data
delivers a message that capacity requirements can grow by tens of thousands to millions of times compared to text.
In other words, as AI expands “memory” and “context” into multimodal forms, the logic is that memory demand grows structurally—not just in a one-off spike.

6) If we reach a World Model: real-time generation can lead to “astronomical memory”

Google-grade (like Genie 3) world model flow

It’s mentioned that it’s moving toward modeling not only text but also images/video and the environment, and
constructing scenes at a level close to real time.

Why more memory becomes necessary

If you continuously “generate/update something” in real time, then
you need to store/reference more of past context + current state + inference results.
So the view is that both memory semiconductors (DRAM) + NAND flash (with characteristics of storage/cache) could face structural excess demand.

7) The ‘core message’ from an investment perspective on memory (a point not talked about as much elsewhere)

The real key takeaway I’m drawing here

When the market shakes just by looking at “a reduction in VeraRubin memory capacity,”
actually, total memory requirements may be maintained or even increase due to module stacking/space efficiency/system structural changes.
And as AI functions expand beyond conversation memory (Dreaming) into storing screen/video context (Chronicle), memory demand is likely to recur structurally, not “just once.”

Keywords investors should check (naturally connected)

Data center expansion cycles (next-generation platforms)
AI infrastructure demand (agents/long-context/multimodal)
Memory supply constraints (physical expansion limits)
Therefore, related sectors should be viewed from the standpoint of *** memory cycles ***, and losses can get bigger the more you get swayed by short-term news.

(As for the five SEO keywords naturally mentioned in the body and summarized together, the *** DRAM , HBM , data center , AI infrastructure , and memory shortage *** flows are connected to one another.)

Final conclusion: It’s not “it decreased”—it can grow because everything becomes denser + AI keeps consuming more

In VeraRubin, “capacity numbers” may look smaller, but with the SOCAM structural change increasing memory mass (module demand), the actual needed DRAM can be interpreted as an additional 10–20%.
As AI evolves toward directions like (Dreaming/Chronicle/World Model),
it expands from text into multimodal/video context, and
it has a high likelihood of keeping memory demand persistently and structurally.
Given the physical limits of supply-side expansion, the viewpoint is that the end result is that the memory shortage character can remain beyond short-term noise.

< Summary >

The “VeraRubin memory cut in half” news can be overly decisive if it’s judged only by spec comparisons
With the SOCAM structure, module mounting efficiency increases, and total memory needs may rise by as much as 10–20%
It clashes with the flow of Nvidia/Jensen and increased supply expansion pressure (e.g., the GTC message)
OpenAI Dreaming expands memory, and Chronicle increases memory requirements by storing screen/video context
If it goes as far as a world model, real-time generation could further explode memory demand
Conclusion: you should look at memory demand through data center structure + an AI functionality roadmap, not just the “capacity decrease” headline

[2 related article news links]

*Source: [ 월텍남 – 월스트리트 테크남 ]

– 메모리는 오히려 20%더 필요하다는 계산입니다.

NextGenInsight.Net

Like this:

Nasdaq Meltdown, Fed Shock, Bond Yields Surge