Intel's Hidden AI Play: Fab + Packaging + Photonics — The Triple Moat
The market prices Intel as a dying CPU company. It doesn't price in a working foundry, a silicon photonics monopoly, and the SambaNova inference alliance.
The Setup Nobody's Discussing
While Wall Street obsesses over NVIDIA's training dominance, a fundamental shift is happening in AI economics. The training era — where CUDA lock-in made NVIDIA untouchable — is giving way to the inference era.
Training an AI model is a one-time event. Inference — running that model billions of times for users — is the recurring revenue. Today, training accounts for ~30% of AI compute spend and inference ~70%. By 2027, inference will be 90%+.
Training requires NVIDIA GPUs because of CUDA — 15 years of libraries, 4M+ developers, and an ecosystem that's nearly impossible to replicate. But inference? Inference is about cost per token. And here, the competitive landscape opens wide:
| Solution | Cost/Token | CUDA Required? | Beneficiary |
|---|---|---|---|
| NVIDIA GPU | $$$ | Yes | NVDA |
| Intel Gaudi | $$ | No (PyTorch) | INTC |
| Google TPU | $$ | No (JAX) | GOOGL |
| Intel Xeon CPU | $ | No | INTC |
| Qualcomm NPU | ¢ | No | QCOM |
The dirty secret: 90% of AI inference today already runs on Intel Xeon CPUs, not GPUs.