SemiAnalysis Weekly The AI Silicon Shortage Explained: TSMC, Nvidia CPO, Memory Crisis & What Comes Next
6 snips
Mar 27, 2026 Dan, a SemiAnalysis contributor who reports conference takeaways and explains CPO and optical interconnects. Ivan Chiam, an analyst modeling AI compute demand and memory dynamics. Sravan Kundojjala, a foundry and memory market analyst tracking TSMC allocation. They map the AI silicon shortage, TSMC capacity choices, memory/HBM pressure, GPU rental pricing, and Nvidia's co‑packaged optics roadmap in short, punchy conversations.
AI Snips
Chapters
Transcript
Episode notes
Smartphone Reallocation Won't Solve GPU Shortage
- Reallocating smartphone N3 wafer demand is an insufficient release valve: even 25% reallocation yields only ~700k Rubin GPUs or ~1.5M TPU v7s.
- Ivan's modeling showed modest smartphone-to-accelerator reallocation barely moves the supply needle given accelerator scale.
HBM Uses Far More Wafer Area Creating Structural Tightness
- HBM is structurally tight because HBM uses 3x–4x more wafer area per bit than commodity DRAM and per-chip HBM capacity keeps rising.
- Ivan added that NVIDIA's demand for very high pin speeds (e.g., 11 Gbps) strains suppliers and keeps HBM scarcity persistent until 2H2027 capacity adds.
GPU Rental Prices Inflected And Then Jumped
- GPU rental prices stopped falling and rose since October 2025, with H100 hourly rates moving from ~$170 low to higher levels amid tight supply.
- Dan described rental-price inflection plus rapid 15–20% increases early in the year and volatile wide price ranges from low liquidity in the market.