An advance by Google is shaking the giants of computer chips and memory on the stock market. Tuesday March 24, the American company published a blog post on Turboquant, a set of algorithms allowing massive compression for large language models (LLM). In other words, Google is offering a solution to reduce the RAM requirements needed for LLMs like ChatGPT or Gemini to function.
It was enough for the markets to worry about manufacturers of chips, semiconductors and other storage solutions. Micron Technology, the largest American manufacturer of memory chips, saw its stock fall from the next day: it closed down 3.4% on Thursday. SK Hynix fell 6% and Samsung Electronics fell 4%.
Artificial intelligence has accentuated the shortage of RAM. This has pushed up prices for companies in the sector, as demand exceeds available supply. However, Google’s solution could reduce pressure on the RAM market.
« For the markets, the reasoning is immediate: if each AI unit consumes less memory, then the structural growth in demand for DRAM and NAND could slow down, which calls into question part of the narrative [récit] ultra-positive integrated into sector valuations », Explains, in a note, John Plassard, economist at Cité gestion.
Better understand TurboQuant
« TurboQuant is a compression method that provides significant reduction in model size without loss of accuracy, making it ideal for key-value (KV) cache compression and vector search », Specifies Google in its blog article. And all this, without losing AI performance.
