Broadcom’s $8.2B AI Chip Forecast and TurboQuant’s 6x Memory Cut Challenge NVIDIA
Google’s TurboQuant compression claims 6x lower memory requirements and 8x faster AI inference, triggering up to 6.4% selloffs in SK Hynix and Kioxia shares. OpenAI’s new partnership with Broadcom targets custom AI accelerators, with Broadcom forecasting semiconductor revenue to reach $8.2 billion in 2026, challenging NVIDIA’s market leadership.
1. Google Unveils TurboQuant to Slash AI Memory Needs
Google researchers introduced TurboQuant, a compression algorithm that reduces memory requirements for large language models by at least sixfold and accelerates inference speeds by up to eight times. This announcement prompted SK Hynix shares to drop 6.4% and Kioxia to fall by the same margin as investors worry about potential demand headwinds for memory components used in AI accelerators.
2. Analysts Cite Jevons Paradox to Predict Continued Memory Demand
Despite initial selloffs, analysts argue that greater efficiency will amplify overall memory consumption in line with the Jevons Paradox. JPMorgan analysts note that hyperscalers may capitalize on lower training costs to expand AI workloads, suggesting only short-term profit-taking rather than a sustainable demand decline.
3. OpenAI Partners with Broadcom to Develop Custom AI Accelerators
OpenAI has selected Broadcom to co-develop custom AI accelerators, marking a strategic move away from reliance on NVIDIA’s GPUs. The partnership leverages Broadcom’s chip design capabilities to tailor semiconductors for OpenAI’s specific application requirements, intensifying competition in the AI hardware landscape.
4. Broadcom Projects AI Revenue Doubling to $8.2 Billion
Broadcom projects its AI semiconductor revenue will climb to $8.2 billion in 2026, more than doubling expected figures and positioning it as a formidable challenger to NVIDIA. This revenue outlook underscores the broader industry trend toward specialized AI chips and could pressure NVIDIA’s unit shipments and pricing power.