大摩眼中的DeepSeek：以存代算、以少胜多！

Core Insights - DeepSeek is revolutionizing AI scalability by utilizing a hybrid architecture that replaces scarce HBM resources with more cost-effective DRAM, focusing on smarter design rather than merely increasing GPU clusters [1][5] Group 1: Technological Innovation - DeepSeek's innovative module, "Engram," separates storage from computation, significantly reducing the need for expensive HBM by employing a "Conditional Memory" mechanism [1][3] - The Engram architecture allows for efficient retrieval of static knowledge stored in DRAM, freeing up HBM for more complex reasoning tasks, thus enhancing overall efficiency [3][5] Group 2: Cost Structure and Economic Impact - The shift from reliance on HBM to DRAM is expected to reshape the hardware cost structure, making AI infrastructure more affordable [5][7] - A 100 billion parameter Engram model requires approximately 200GB of system DRAM, indicating a 13% increase in the use of commercial DRAM per system compared to existing setups [5][7] Group 3: Competitive Landscape - Despite hardware limitations, Chinese AI models have rapidly closed the performance gap with leading global models, demonstrating strong competitive capabilities [6][8] - DeepSeek V3.2 achieved an MMLU score of approximately 88.5% and coding capability of around 72%, showcasing its efficiency in reasoning and performance [6][8] Group 4: Future Outlook - The upcoming DeepSeek V4 model is anticipated to leverage the Engram architecture for significant advancements in coding and reasoning, potentially running on consumer-grade hardware [8] - This development could lower the marginal costs of high-level AI inference, facilitating broader deployment of AI applications without reliance on expensive data center GPUs [8]