unknown

NVIDIA is not just a GPU company — it operates a five-layer AI infrastructure platform: silicon (GPUs), networking (NVLink/InfiniBand via Mellanox), platform software (CUDA), framework integration (PyTorch/TensorFlow/JAX), and cloud services (DGX Cloud). No competitor owns more than two of these layers. CUDA, launched in 2006, is the core moat with 4M+ developers and 20 years of optimization depth. FY2025 revenue hit $130.5B, with data center at $115.2B (91% of total). Three key strategic bets drove this: investing in free software (CUDA) to lock in hardware sales, annual GPU architecture self-cannibalization to stay ahead of competitors, and the $6.9B Mellanox acquisition to own cluster interconnects. Threats are real but 2-3 years from mattering: ROCm, OpenAI's Triton, and hyperscaler custom silicon are all attacking the stack. Practical advice: use NVIDIA today, but design infrastructure for hardware portability.

What Does NVIDIA Actually Do?

Paolo Perrone

<p>Great read, thanks for sharing (and writing) this one! I liked the wording and how the article was structured — overall it was really easy to read this one</p>
