Google’s TurboQuant Cuts Memory Needs 6x, OpenAI Taps Broadcom for AI Chips
Google unveiled its TurboQuant compression algorithm, reducing memory requirements for large language models by sixfold and boosting inference speed eightfold, triggering a selloff in memory chip stocks of up to 6.4%. OpenAI's partnership with Broadcom to co-develop custom AI accelerators could reshape cloud hardware competition and pressure Google Cloud.
1. Google Unveils TurboQuant Compression
Google’s AI research team introduced TurboQuant, a compression algorithm that cuts memory requirements for running large language models by sixfold while accelerating inference speeds eightfold, aiming to lower training costs and improve efficiency for hyperscalers and enterprise users.
2. Memory Stock Selloff and Analyst Response
Shares of key memory suppliers including SK Hynix and Kioxia slumped as much as 6.4% after the announcement, with Micron and Sandisk also falling; however, analysts at JPMorgan and Morgan Stanley cited the Jevons Paradox, arguing that efficiency gains will fuel higher, not lower, memory demand.
3. OpenAI-Broadcom Partnership Shifts Hardware Landscape
Separately, OpenAI partnered with Broadcom to co-develop custom AI accelerators, with Broadcom expecting AI semiconductor revenue to rise to $8.2 billion in 2026, a move that broadens AI hardware competition beyond Nvidia and could pressure Google Cloud’s GPU-based offerings.