NVDANVIDIA Corporation

Nvidia sells a computing platform rather than chips alone — and it is the CUDA software layer that keeps customers in place.

The company began with graphics accelerators for gaming, but the bulk of revenue now comes from the data centre: accelerators for training and serving large models, networking gear acquired with Mellanox, and complete racks. Buyers are large cloud operators and model developers, so revenue is heavily concentrated — a handful of customers account for a substantial share of sales.

The real barrier to competitors is software, not silicon. CUDA and the library ecosystem built on it are what existing machine-learning code is written against; porting that code to a rival accelerator costs time and carries risk. That is why competitors offering comparable hardware at lower prices have not yet shifted demand.

The weak points are cyclicality and dependence on other people's capacity. Nvidia owns no fabs: wafers come from TSMC, and access to leading-edge process nodes and to HBM memory constrains output more tightly than demand does. A separate factor is US export control on shipments to China, which closes part of the market and forces cut-down versions of the accelerators.