StepFun — AI supply-chain exposure

StepFun is a young Shanghai frontier lab founded by a former Microsoft executive, building the Step series of very large, trillion-parameter models with a deliberate multimodal focus — systems meant to handle text, images and other formats in one model rather than language alone. It sits in the models layer of the chain, both pretraining its own systems and serving them for inference.

Its structural position is that of a challenger inside China's frontier cohort. A lab like this is a consumer of the layers beneath it — accelerators, memory and the data-centre capacity to train at scale — and its leverage comes from differentiation, here a bet that multimodal breadth rather than pure text scale is where frontier value accrues. The model weights it as structurally central because a capable domestic model house anchors demand for everything upstream, even as its own moat is still being built.