DeepSeek is rewriting the expansion rules of AI: the next generation of AI's decisive point is no longer simply stacking larger GPU clusters, but rather using smarter hybrid architectures to replace scarce HBM resources with more cost-effective DRAM.
According to news from the Wind Trading Desk, Morgan Stanley's latest research report released on January 21 shows that DeepSeek is changing the way large language models are constructed through an innovative module called &#34;Engram.&#34; The core breakthrough lies in separating storage from computation, significantly reducing the demand for expensive and scarce high-bandwidth memory (HBM) by introducing a &#34;Conditional Memory&#34; mechanism, and instead utilizing lower-cost ordinary system memory (DRAM) to handle complex reasoning tasks.
Morgan Stanley analyst Shawn Kim and his team believe that DeepSeek demonstrates the philosophy of &#34;Doing More With Less.&#34; This technological path of separating storage from computation not only alleviates the AI computing power constraints faced by China but also proves to the market that efficient hybrid architectures are the next frontier for AI.
This architecture, which has garnered significant attention from Morgan Stanley, originates from the team of DeepSeek founder Liang Wenfeng and collaborators from Peking University, who published a groundbreaking paper titled “Conditional Memory via Scalable Lookup” on January 13. In this paper, the team first proposed the &#34;Engram&#34; module.
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/01fc47f2-6706-4ee7-9030-b4761c1ba67c.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/01fc47f2-6706-4ee7-9030-b4761c1ba67c.png"/>
<h2>Goodbye to Brutal Computation: Engram Module and &#34;Conditional Memory&#34;</h2>
Morgan Stanley pointed out in the report that current Transformer models are extremely inefficient when it comes to remembering and retrieving simple static facts. For example, processing a simple query like &#34;London is in the UK&#34; requires expensive computations through multiple layers of attention mechanisms and feedforward networks in traditional models. This approach wastes valuable GPU computing resources.
DeepSeek's solution is to introduce the principle of &#34;Conditional Memory,&#34; namely the Engram module.
The core of this architecture is to separate static pattern storage from dynamic reasoning. DeepSeek no longer loads all information into expensive HBM at once, but instead offloads the model's &#34;library&#34; or &#34;dictionary&#34; (static knowledge) to the CPU or system memory (DRAM), retrieving it only when needed.
Morgan Stanley analysts emphasized in the report: &#34;DeepSeek separates 'Conditional Memory' from computation, unlocking new levels of efficiency for large language models (LLM). Engram is an efficient method to 'lookup' basic information without overloading HBM, thereby freeing up capacity for more complex reasoning tasks
This design directly addresses the most expensive bottleneck in current AI infrastructure—HBM. By reducing the reliance on HBM, DeepSeek demonstrates that efficiency can be improved under existing GPU and system memory architectures, effectively reducing the need for costly hardware upgrades.
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/a261b804-5c36-4c2b-b702-858771c1d8a6.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/a261b804-5c36-4c2b-b702-858771c1d8a6.png"/>
<h2>Infrastructure Economics: Reducing HBM Dependence, Amplifying DRAM Value</h2>
The most direct impact of this technological transformation is the reshaping of hardware cost structures. Morgan Stanley points out that the Engram architecture minimizes the need for high-speed memory (HBM) by separating static pattern storage from dynamic computation. This means that infrastructure costs may shift from expensive GPUs to more cost-effective memory (DRAM).
The report breaks down this data impact in detail: “Although the paper does not specify, a 100 billion parameter (100B) Engram (assuming 2 bytes per parameter under FP16/BF16) means a minimum requirement of about 200GB of system DRAM.”
In contrast, NVIDIA's Vera Rubin system is already equipped with 1.5TB of DRAM per CPU. Analysts calculate that DeepSeek's architecture implies “each system will increase its usage of commoditized DRAM by about 13%.”
The investment logic behind this shift is very clear:
<ol>
<li>
Cost Structure Shift: Infrastructure costs may shift from GPUs to memory.
</li>
<li>
Cost-Effectiveness is Key: Configurations with moderate computation but large memory may provide higher “performance per dollar” than pure GPU expansion.
</li>
<li>
Memory Value Reevaluation: The improvement in reasoning capability surpasses the gains in knowledge acquisition, indicating that the value of memory has extended beyond computation.
</li>
</ol>
<h2>“Constraint-Induced Innovation”: The Breakthrough Path of Chinese AI</h2>
Morgan Stanley's data shows that despite limitations in advanced computing power, hardware acquisition, and training scale, leading AI models in China have rapidly narrowed the performance gap with global frontier models (such as ChatGPT 5.2) over the past two years. DeepSeek V3.2 has performed excellently in standardized benchmark tests, with an MMLU score of about 88.5% and coding ability (SWE-Bench) of about 72%, demonstrating strong competitiveness in reasoning and efficiency.
The report attributes this phenomenon to “constraint-induced innovation.” The development of AI in China is no longer solely reliant on brute-force parameter growth but has shifted towards algorithmic efficiency, system design, and deployment pragmatism. Morgan Stanley notes: “DeepSeek is proving that the next leap in AI capability may not come from more GPUs, but from learning how to think under constraints.”
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/aff5f06d-9e11-43c8-9d17-8a944d192651.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/aff5f06d-9e11-43c8-9d17-8a944d192651.png"/>
Morgan Stanley analysts pointed out: &#34;Strategically, this indicates that the progress of AI in China may increasingly depend not on directly narrowing the hardware gap, but on algorithmic and system-level innovations that bypass hardware bottlenecks.&#34;
<h2>Looking Ahead: Can Consumer Graphics Cards Run Large Models?</h2>
The report gives an attractive forecast for DeepSeek's next-generation model V4. Morgan Stanley expects that with the Engram memory architecture, V4 will achieve significant breakthroughs upon release, particularly in coding and inference.
What draws more market attention is its reduction of hardware thresholds. Morgan Stanley wrote in the report: &#34;Like its predecessor, this model is highly likely to run on consumer-grade hardware, with consumer-grade hardware (such as the RTX 5090) potentially being sufficient.&#34;
This means that the marginal cost of high-level AI inference will further decrease, allowing AI applications to be deployed more widely without fully relying on expensive data center-level GPU clusters.
Based on the above technological trends, Morgan Stanley reiterated its optimism about the localization theme of memory and semiconductor equipment in China.
The report clearly states: &#34;By decoupling memory from computation, China is building LLMs that are not only smarter but also structurally more efficient.&#34; Although the scale of the AI market in China is still only a small fraction of that in the United States, its spending and adoption momentum suggest that the upside potential may be underestimated.
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/26236e3a-f80f-4ff0-a073-792c634f2179.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/26236e3a-f80f-4ff0-a073-792c634f2179.png"/>
<pre><code>
The above exciting content comes from [Wind Trading Platform](https://mp.weixin.qq.com/s/uua05g5qk-N2J7h91pyqxQ).

For more detailed interpretations, including real-time analysis and frontline research, please join the【 [Wind Trading Platform ▪ Annual Membership](https://wallstreetcn.com/shop/item/1000309)】

[![](https://wpimg-wscn.awtmt.com/3c4a713c-7a38-4582-9850-d0eabaf0e7ad.png)](https://wallstreetcn.com/shop/item/1000309)</code></pre>

DPSK

DeepSeek is rewriting the AI expansion law with the philosophy of "spending less money to get more done." Morgan Stanley believes that DeepSeek's unique Engram architecture separates computation from storage through "conditional memory," replacing scarce HBM resources with cost-effective DRAM. This not only reshapes the value of ordinary storage but also proves that the winning point for the next stage of AI is no longer the brute force stacking of GPUs, but rather an efficient hybrid architecture

Morgan Stanley's view on DeepSeek: using savings to replace calculations, winning with less!