The Great AI Silicon Shortage
9 hours ago
- #TSMC-N3
- #HBM-shortage
- #AI-compute
- The AI compute shortage is intensifying due to skyrocketing token demand and agentic workflows, with companies like Anthropic adding $6B ARR in a month.
- TSMC's N3 wafer capacity is a major bottleneck, with AI accelerator demand set to dominate N3 usage by 2026, squeezing out smartphone and CPU demand.
- Nvidia, Google, AWS, and others are transitioning to N3 for AI accelerators, causing a demand shock that TSMC's current capacity cannot meet.
- Smartphone demand may soften, potentially freeing up N3 capacity for AI accelerators, but memory constraints (especially HBM) remain a critical issue.
- HBM demand is growing rapidly, consuming more wafer capacity than commodity DRAM, with higher pin speeds and stack heights exacerbating supply challenges.
- CoWoS packaging constraints are easing, but front-end capacity (N3 wafers) is now the dominant bottleneck in AI chip production.
- Hyperscalers like Google are doubling capex expectations for 2026, but silicon supply remains the limiting factor for AI infrastructure expansion.