Sandisk shares sold off following Google's introduction of TurboQuant, an AI compression technology that could significantly reduce memory requirements for AI models. The announcement prompted investors to reassess long-term demand assumptions for memory-related stocks. TurboQuant directly impacts Sandisk's core business of flash storage products serving data centres, PCs and mobile devices. The memory industry has been closely tied to AI infrastructure buildouts, where demand for fast, dense storage has been a key growth driver. Trading at $615.83, Sandisk's shares are approximately 20% below the analyst target of $770.32. The stock recorded a 30-day return of -3.07% following the news. Debate continues over whether more efficient compression might ultimately expand AI usage overall, potentially supporting memory volumes through increased model counts.

Micron Technology has been identified as a semiconductor stock to consider buying during the current memory supercycle. The company, one of three major DRAM producers alongside SK Hynix and Samsung, generates approximately 80% of revenue from DRAM and the remainder from NAND memory. Demand for high bandwidth memory has created supply constraints across the DRAM market, driving prices higher. Last quarter, Micron's revenue surged nearly threefold to $23.9 billion, whilst gross margins jumped to 74.4% from 36.8% year-over-year. The company projects margins will reach 81%. Despite strong performance, Micron trades at a forward P/E of four times fiscal 2027 estimates due to the memory market's historically cyclical nature. However, the company is securing longer contracts, including a five-year deal, providing greater revenue visibility.

Memory stocks fell Wednesday after Google unveiled TurboQuant, a compression algorithm that could reduce memory requirements for AI systems. SanDisk dropped 5.7%, Micron Technology fell 3%, Western Digital declined 4.7%, and Seagate Technology slid 4%, despite broader technology sector strength. TurboQuant compresses key-value cache in large language models to 3 bits without requiring training, whilst maintaining accuracy. Testing on open-source models including Gemma and Mistral showed a 6x reduction in memory size and up to 8x performance increase on H100 GPU accelerators. The technology uses PolarQuant compression and the Quantized Johnson-Lindenstrauss algorithm to eliminate errors. Google will present the findings at ICLR 2026 and AISTATS 2026. Memory stocks have rallied significantly year to date, making them vulnerable to demand-reducing developments.