Google's new algorithm claims to "6x compress KV cache," causing the U.S. stock storage sector to weaken.

On Wednesday, after the U.S. stock market opened, the storage sector weakened against a backdrop of decent market sentiment. By the close, Micron Technology fell 3.40%, SanDisk dropped 3.50%, and Western Digital and Seagate Technology also declined.

Regarding today’s market fluctuations, multiple sources have pointed the finger at Google. The AI giant earlier introduced a compression algorithm called TurboQuant that may reduce memory requirements for AI systems.

According to Google, the TurboQuant compression technology aims to reduce memory usage for large language models and vector search engines. The algorithm primarily addresses the bottleneck issue of key-value caches used for storing frequently accessed information in AI systems. As context windows expand, these caches are becoming a major memory bottleneck.

TurboQuant can compress key-value caches to 3-bit precision without the need to retrain or fine-tune the models, while essentially maintaining the model’s accuracy. Tests on open-source models including Gemma and Mistral show that the technology can achieve approximately 6 times the memory compression effect for key-value caches.

Additionally, test results on NVIDIA’s H100 accelerator show that, compared to unquantized key vectors, the algorithm can achieve up to approximately 8 times performance improvement. Researchers also noted that the application of this technology is not limited to AI models but includes supporting vector retrieval capabilities for large-scale search engines.

Google plans to showcase the TurboQuant technology at the International Conference on Learning Representations (ICLR 2026) in April.

It is evident that while the application prospects of this technology still raise questions, the market has already begun to trade on expectations of a shift in memory demand.

Regarding the latest changes, Wells Fargo TMT analyst Andrew Rocha interpreted, “As the context window continues to expand, the data storage size in KV caches is experiencing explosive growth, thereby increasing the demand for memory capacity. TurboQuant directly compresses this cost curve. If this technology can be widely adopted, it would be beneficial for the memory cost curve.”

Rocha also stated that this technology could affect future demand assessments for memory capacity specifications.

He wrote, “If the memory specifications required for these AI applications are significantly reduced, the market will quickly reassess how much memory capacity is actually needed.”

However, Rocha also pointed out that it is still unclear whether this technology is only applicable to Google’s own ecosystem or if it can be extended to other AI labs. Additionally, there is uncertainty about whether test results in a lab environment can be successfully translated into performance in real production environments.

It is worth mentioning that, as the disruptor of the storage sector, Google has not gained any benefits. The company’s stock price briefly fell below $290 on Wednesday, down nearly 17% from the historical high of $349 set in early February, just a stone’s throw away from the critical psychological level of 20%.

Eastmoney Illustration · Adding Practical Insights

(Source: Caixin)