a day ago
- AI compute supply is seriously struggling to meet demand, with Anthropic recently experiencing severe uptime issues and degrading product features to free up compute.
- Alibaba Cloud also reports inability to keep pace with customer demand, showing extreme contention for inference resources.
- Agentic AI models are accelerating token demand as they handle complex tasks, with penetration still in low single digits among knowledge workers.
- If adoption reaches just 5% of knowledge workers, the industry cannot handle the token demand due to DRAM supply constraints capping infrastructure growth until 2027.
- DRAM price rises indicate supply tightness, and no amount of investment can overcome physical hardware limitations.
- Users should avoid single-provider lock-in and consider long-term contracts for businesses, as inference supply constraints will worsen through 2026-2027.
- Hardware engineering is a more telling indicator of AI sustainability than financial engineering narratives about an AI bubble.