ARCHIVED · RETIRED NOV 2026
Originally Published June 2026
Semiconductor Deep Dive
Intel's Hidden AI Play: Fab + Packaging + Photonics — The Triple Moat
The market prices Intel as a legacy x86 CPU vendor. It overlooks a Western 18A foundry, 3D Foveros packaging, a 21.5% silicon photonics share, and the SambaNova inference stack.
Live AI Inference & Foundry Basket Quotes
Live Spot & 1D Change
18A
RibbonFET + PowerVia
3
Vertical Moats
$17.8B
Photonics TAM 2035
90%+
2027 Inference Mix
The Setup Nobody Is Modeling
While Wall Street extrapolates NVIDIA's AI training lock-in indefinitely, AI compute economics are pivoting from model training (one-time capex gated by CUDA) to production inference (recurring cost-per-token):
| Compute Architecture | Cost / Token | CUDA Required? | Primary Beneficiary |
|---|---|---|---|
| Hyperscale GPU Cluster | $$$ | Yes | NVDA |
| Dedicated Inference ASIC / RDU | $$ | No (PyTorch / JAX) | INTC / GOOGL |
| Xeon AMX Host CPU | $ | No | INTC |
| Edge / On-Device NPU | ¢ | No | QCOM |
Enterprise inference cares about cost per million tokens and rack power density — opening the door to non-CUDA silicon stacks.
Current Active Flagship Theses & Hubs