You Can't Buy a Data Center
9 hours ago
- #AI Infrastructure
- #Data Centers
- #Supply Chain
- AI data center demand is surging, causing supply chain bottlenecks across GPUs, HBM memory, advanced packaging, power transformers, cooling systems, networking, and critical materials.
- NVIDIA dominates AI chip production with 80% market share, but AMD is gaining ground with a $100 billion deal with OpenAI.
- High Bandwidth Memory (HBM) is in critical shortage, with production sold out through 2026, and prices up 246% in 2025.
- Advanced packaging (CoWoS) is monopolized by TSMC, with NVIDIA alone needing 60% of global capacity in 2026.
- Power transformers now have a 2.5 to 4-year lead time, up from 6-8 months pre-AI boom, due to custom designs and limited manufacturing capacity.
- Liquid cooling is essential for high-power AI racks, with demand surging 156% year-over-year and lead times of 6-9 months.
- Networking demands are skyrocketing, with each AI rack requiring up to 1,526 fiber connections, driving $150 billion in fiber needs.
- Critical materials like T-Glass are bottlenecked, with production controlled by a single Japanese company and new capacity not expected until late 2027.
- The AI supply chain is highly concentrated geographically, with 90% of advanced chips made in Taiwan and 62% of HBM memory from SK Hynix in Korea.
- Geopolitical tensions, like the US-China chip war, are exacerbating supply chain issues, with export controls and tariffs disrupting production.
- Power infrastructure is the most critical bottleneck, with data center electricity demand projected to grow 165% by 2030, outpacing grid capacity.
- Hyperscalers like Microsoft and Amazon are turning to nuclear power to meet energy needs, but solutions won't scale until the 2030s.
- Supply chain shortages are expected to persist through at least 2028, with HBM memory and power infrastructure remaining tight until 2033.