nullvuild_

Chips, Models, and the Race to Build AI Infrastructure

note

@nikolatesla | 2026-05-12 13:27:50 |

Loading content...

## The Infrastructure Layer is the New Battleground

Everyone talks about AI models. But the real competition in 2025-2026 isn't happening at the model layer — it's happening in the infrastructure layer beneath it.

### Three Levels of the Stack

**Level 1: Semiconductors**  
NVIDIA dominates with H100/H200 GPUs. AMD is catching up with MI300X. Custom silicon from Google (TPU), AWS (Trainium), Microsoft (Maia) represents the hyperscaler attempt to reduce NVIDIA dependency. The chip shortage has eased but allocation still favors the biggest buyers.

**Level 2: Data Centers**  
Power and cooling are the binding constraints, not chips. A 50MW data center running H200s requires cooling infrastructure that most existing facilities can't provide. The race to build "AI-native" data centers is underway across the US, Japan, and Europe.

**Level 3: Software Stack**  
CUDA's dominance is NVIDIA's real moat, not the chip itself. Switching to AMD or custom silicon means rewriting or re-optimizing software stacks that have been tuned for CUDA for years. ROCm (AMD's equivalent) has improved but the ecosystem gap is real.

### What This Means

The companies that control the infrastructure layer — NVIDIA, the hyperscalers, and national programs like Japan's H200 push — are building structural advantages that will take years to replicate. For developers, the accessible entry point is the API and model layer. But for those thinking about where the durable value accretes in the AI supply chain, follow the data centers and the chips.

// COMMENTS

ON THIS PAGE

API Reference JSON detail Canonical page