Introduction

DeepSeek's latest model release highlights how architectural efficiency can dramatically lower memory demands during AI inference, signaling a shift in how hardware requirements are calculated.

What Happened

The V4.1-Flash variant requires only one-quarter the high-bandwidth memory and one-eighth the solid-state storage for its key-value cache compared to the prior generation. This reduction applies specifically to the KV cache, which stores attention state while processing tokens, and not to the model weights or training data.

Why This Matters

For investors in Micron Technology and Sandisk, the implications are significant. Lower memory intensity per token could expand AI accessibility, driving broader workload adoption. However, the reduction covers only inference-time cache storage, not the broader memory footprint of model weights or training cycles. Still, as token volumes grow, even small per-token savings translate into substantial hardware demand shifts.

Key Takeaways

  • DeepSeek V4.1-Flash activates just 8 billion parameters for input and 16 billion for output from a 552-billion-parameter architecture.
  • KV cache optimization cuts HBM need by 75% and SSD need by 87.5% relative to the previous generation.
  • Micron's recent quarterly results showed strong cloud memory and data-center revenue, with billions in HBM4 shipments.
  • Sandisk's data-center revenue surged over 100% sequentially, driven largely by pricing power rather than volume growth.
  • Hedge fund activity reflected mixed sentiment, with increased positions in Micron but notable short interest in Sandisk.
  • Architectural efficiency may stimulate wider AI adoption, potentially balancing memory demand against per-token savings.

Conclusion

DeepSeek's update reframes the relationship between AI efficiency and hardware demand. While the memory-intensity thesis faces a recalibration, the long-term outlook for Micron and Sandisk depends on how broadly these efficiency gains are adopted and whether expanded usage offsets reduced per-token costs.