NVIDIA — AI supply-chain exposure

NVIDIA sits at the gravitational centre of the AI build-out. It designs the GPUs that train and serve nearly every frontier model, but it makes none of them itself — the silicon is fabricated by TSMC and only becomes a usable accelerator once it is fused to its high-bandwidth memory through advanced packaging. That makes NVIDIA both the chain's biggest source of pricing power and one of its most exposed buyers: its ability to ship depends on foundry slots, on CoWoS packaging capacity, and on a memory supply that has been the tightest link in AI for over a year.

Its real moat is less the chip than the system around it — the CUDA software stack developers are locked into, and the networking that lashes tens of thousands of GPUs into one machine. When the model reads NVIDIA as a producer sitting on binding constraints, that is why: it owns the reference accelerator, so demand for everything beneath it in the chain ultimately routes through NVIDIA's roadmap.