US memory chip stocks lost ~$100B in market value this week, led by Micron's 15% drop, after Google Research detailed its TurboQuant compression algorithm
New research suggests AI data centres will need much less memory than investors had bargained for — US memory chip stocks …
Context & Ripple Effects
The sell-off interrupts a memory-market narrative built around AI-led scarcity: in February, Micron said it could meet only roughly half to two-thirds of demand for certain key customers. That backdrop made memory suppliers especially sensitive to evidence that models could use less memory.
The immediate market interpretation was subsequently contested. Analysts and researchers later argued that more memory-efficient LLMs could expand, rather than shrink, memory demand, shifting the question from memory required per model to total AI workload volume.
First-order effects
- Micron and other US memory-chip stocks are repriced against a lower implied memory requirement for AI data centres; Micron bears the sharpest immediate pressure after its 15% decline.
- Google Research's disclosure puts TurboQuant at the center of investor scrutiny over whether compression changes the hardware intensity of AI deployments.
Second-order effects
- Memory suppliers' AI-driven sales and capacity assumptions face closer challenge, particularly where valuations depend on sustained data-centre memory intensity.
- AI infrastructure buyers gain a potential route to lower memory needs per workload, increasing pressure on suppliers to demonstrate that aggregate deployment growth can offset efficiency gains.
Third-order effects
- The episode points to a more volatile AI-memory capex cycle, in which software-level efficiency advances can quickly reset expectations for component demand.
- If efficiency enables materially broader model deployment, the eventual effect may be demand expansion rather than contraction; the later analyst view that TurboQuant could increase total demand underscores that this remains an adoption-and-volume question, not a simple per-model memory calculation.
The trend: AI infrastructure economics are increasingly being shaped by the tension between lower resource use per workload and higher total workload volume.