DeepSeek의 최신 모델은 KV-cache HBM 수요를 75%, SSD 수요를 87.5% 감소시켜 메모리 바운드 계산 비용을 절감한다.

The memory-shortage thesis assumes AI usage and memory demand rise in tandem. DeepSeek's September 10 release of V4.1-Flash suggests that relationship may be weaker than expected. The new model requires one-quarter the HBM and one-eighth the SSD capacity for its KV cache compared to its predecessor—a significant architectural improvement that warrants attention from investors in Micron Technology (NASDAQ: MU) and Sandisk Corporation (NASDAQ: SNDK).

That said, the qualification matters. DeepSeek's efficiency gains apply specifically to KV cache, which stores attention state during inference, and only in comparison to its prior architecture. The reduction does not eliminate memory consumed for model weights or training. Inference, however, is where sustained token growth can transform an architectural improvement into a genuine demand shock.

The optimistic case is straightforward: lower memory costs reduce token expenses, stimulating broader workload adoption. DeepSeek's 552-billion-parameter model activates only 8 billion parameters for input and 16 billion for output. If cost reductions drive usage faster than memory per token declines, Micron can still increase sales of HBM and data-center DRAM, while Sandisk benefits from a larger installed base and broader storage requirements beyond KV cache.

Recent performance figures illustrate why investors favor this outcome. Micron's Cloud Memory revenue reached $13.77 billion last quarter with an 83% gross margin, while Core Data Center revenue totaled $11.52 billion at an 87% margin. The company shipped more than $1 billion in HBM4 alone. Sandisk's data-center revenue rose 103% sequentially to $2.98 billion, with pricing driving roughly two-thirds of the quarter-on-quarter increase.

These numbers also reveal the bear case. Extraordinary margins and pricing power reflect underlying scarcity. If model designers broadly reduce bytes required per token, demand can accelerate without matching current memory-intensity forecasts. Sandisk faces particular risk to a pricing-power reversal, given that roughly two-thirds of its latest sequential revenue growth came from higher pricing rather than volume.

Hedge fund positioning offers limited insight. Insider Monkey's database showed 184 MU holders in Q2 2026, up from 154 in Q1, with Coatue Management holding 3,142,269 shares after increasing its position 1,794%. SNDK holders rose to 128 from 114, while Blue Whale Capital reduced holdings to 97,653 shares from 114,605. These are quarter-end snapshots, not reactions to DeepSeek's announcement.

As of August 31, 5,996,105 Sandisk shares were sold short, representing 4.10% of float and 0.46 days of average volume. DeepSeek has not invalidated the memory cycle. It has simply demonstrated that unit efficiency belongs alongside token growth in every demand forecast.