Groq — AI supply-chain exposure

Groq designs inference accelerators it calls LPUs — chips purpose-built to serve already-trained models rather than to train them. Where a general-purpose GPU juggles many workloads, Groq's silicon is a fixed-function ASIC engineered for one thing: pushing tokens through a model with deterministic, predictable timing. That determinism is the pitch — latency you can count on, request after request — and it matters most in the serving half of the chain, where a model is run over and over rather than built once.

Its structural hook is the shift in where AI compute is spent. Training is a one-off burst; inference is the recurring cost that scales with every user query, and it is becoming the larger workload. A challenger optimised for low-latency serving is a genuine wedge into that market. The catch is incumbency — Groq competes against an entrenched GPU ecosystem and the software developers are already locked into, so its centrality rests on inference specialising away from general-purpose hardware.