It's complicated

• Full-stack players: Google: TPU + Gemini + apps. SpaceXAI: Colossus compute + Grok + X. OpenAI: models + apps + custom chips + Stargate. The biggest players are vertically integrating in both directions.
• Partnerships everywhere: Anthropic ⨯ SpaceXAI (Claude capacity on Colossus 1) · Palantir ⨯ NVIDIA (sovereign AI) · SK Telecom ⨯ NVIDIA/OpenAI (AI cloud + consumer distribution) Model labs, infrastructure owners, and distributors are wiring themselves together.
• The chokepoint: TSMC. Over 90% of leading-edge chips, including NVIDIA, Apple, AMD, and custom silicon, are fabricated by one company in Taiwan. Capacity is sold out. HBM is a separate chokepoint: SK Hynix, Samsung, and Micron make the stacked memory; TSMC packages it with the processor.
• US vs China: chip export controls cut NVIDIA's China share from ~95% to near zero, and China answered with world-class open models, some trained entirely on domestic chips.