MU: TurboQuant Is Not Another DeepSeek Moment

On March 24, 2026, Google published a TurboQuant blog post, bringing back into focus a piece of work that had already been public since April 2025. At its core, TurboQuant is an aggressive compression scheme targeting inference-time KV cache and vector search. The headline numbers—“≥6× KV cache reduction, up to 8× speedup, with near-zero accuracy loss”—were enough to trigger an immediate market reaction, quickly reframing it as a potential negative for HBM, DRAM, and even SSD demand.

That reaction follows a familiar playbook. If inference-time memory footprint can be materially compressed, does that imply a structural reset in memory demand? Extend that logic one step further, and it starts to resemble the “DeepSeek moment” narrative from early 2025, when efficiency gains in models sparked fears of a valuation reset across the compute stack. The market response this time looked similar—SanDisk sold off as much as 8.1%, while Micron dropped another ~5.8% on top of prior weakness.

---

The problem is, this linear extrapolation is likely wrong—again.

Start with first principles. KV cache is a mechanism used during transformer inference to store historical token key/value tensors, avoiding recomputation of attention projections for past tokens. It is an inference-side construct, not a training one. Conceptually, it trades memory footprint and bandwidth for compute reuse. As such, it primarily impacts the working set during inference: GPU memory utilization, memory bandwidth pressure, supported context length, batch size, and user concurrency. It does not touch model weights, nor broader storage categories such as training datasets, checkpoints, object storage, logs, or embedding corpora.