NVIDIA

NVIDIA is the dominant supplier of AI accelerators, whose data-center GPUs and the CUDA software ecosystem underpin the training and inference of nearly every frontier model. Its H100, H200, and Blackwell-generation chips became the scarcest strategic resource of the AI era, and the company's market capitalization crossed into the multi-trillion-dollar range as demand outstripped supply. Its influence extends through investments across the entire AI stack.

Founded in 1993 by Jensen Huang, Chris Malachowsky, and Curtis Priem, NVIDIA built its business on graphics processing units for gaming before its parallel-computing architecture proved ideal for deep learning. The CUDA platform, launched in 2006, created a software moat that has proven far harder for rivals to overcome than the hardware itself.

Data-center GPUs are now the bulk of NVIDIA's revenue. Its annual cadence — Hopper (H100/H200), then Blackwell (B200/GB200), then Rubin — sets the tempo for the whole industry, and full-rack systems like the GB200 NVL72 sell compute at datacenter scale. NVIDIA also sells networking (from its Mellanox acquisition), CPUs (Grace), and inference software (NIM, Dynamo).

Beyond selling chips, NVIDIA has become a central financier of AI demand, taking stakes in and signing supply deals with OpenAI, xAI, CoreWeave, Mistral, and many neoclouds. Its main challengers are AMD, custom silicon from Google, Amazon, and Broadcom-designed accelerators, and inference specialists like Groq and Cerebras — but its combination of hardware, networking, and software keeps it the default choice for frontier training.