Wells says TurboQuant cuts memory load, echoes Jevons Paradox for AI adoption

Wells Fargo notes that Google Research (GOOGL) published a blog post on a new quantization method, TurboQuant, that compresses LLM Key Value caches to 3 bits with zero accuracy loss without retraining resulting in up to an eight times speedup with its 4-bit vs. unquantized 32-bit keys significantly cutting memory overhead. While the firm can appreciate that major algorithmic innovation like this could be viewed as negative for memory makers, it is also reminded of Jevons Parodox in that the industry is continually focused on driving greater AI efficiency, or unlocking exponential improvements in token unit economics, as a driver of greater adoption / deployment. Publicly traded companies in the space include Micron (MU) and SanDisk (SNDK), which are underperforming this morning, and Dell (DELL) and HP Inc. (HPQ), both trading higher.

Claim 55% Off TipRanks