← AI Terminology
NVIDIA
NVIDIA is an American semiconductor company that designs GPUs and AI computing platforms — the dominant supplier of hardware for AI training and inference worldwide, with an estimated 70–95% market share in data centre AI accelerators.
Its H100 and A100 GPUs are the primary compute substrate for frontier AI model training.
Its H100 and A100 GPUs are the primary compute substrate for frontier AI model training.
Why It Matters in AI
CUDA (2007) gave GPU-accelerated computing a software ecosystem; the 2012 AlexNet deep learning breakthrough cemented GPUs as the AI compute standard. NVIDIA's first-mover advantage in AI hardware — GPUs + CUDA + cuDNN + the full software stack — created an almost insurmountable moat. Training GPT-4 required ~25,000 A100 GPUs; Llama 3 required ~16,000 H100s. NVIDIA is effectively the picks-and-shovels of the AI gold rush, with revenue growing from $27B (2023) to ~$130B (2025).
Key Points
| Aspect | Description |
|---|---|
| CUDA | Proprietary GPU programming platform — 17+ years of ecosystem lock-in (libraries, frameworks) |
| NVLink | High-bandwidth GPU-GPU interconnect (3.2 TB/s on NVLink 4) — critical for model parallelism |
| GPU lines | H100/H200 (Hopper, data centre AI), A100 (Ampere), GB200/B200 (Blackwell, 2024–25), RTX 4090 (consumer) |
| Market cap | Became a $3T+ company in 2024 — one of the most valuable companies in history driven by AI demand |
| Competitors | AMD (Instinct MI300X), Google (TPU v5), Intel (Gaudi 3), Cerebras, Groq — all far behind |
| Software stack | cuDNN, NCCL, TensorRT, Triton Inference Server, NeMo — end-to-end AI development platform |
Simple Analogy
The oil refinery of the AI era: just as the 20th century industrial economy ran on oil, the 21st century AI economy runs on GPU compute — and NVIDIA controls the refinery. Every major AI company (OpenAI, Anthropic, Meta, Google DeepMind) pays NVIDIA for the compute that powers their models.
Common Usage Examples
nvidia-smi— monitor GPU utilisation, memory, temperature during training- CUDA:
model.to("cuda")— move PyTorch model to NVIDIA GPU - H100 cluster: 8× H100 SXM5 = 640GB HBM3, 3.2 TB/s NVLink — standard frontier training node
- TensorRT:
trtexec --onnx=model.onnx --fp16 --saveEngine=model.trt— NVIDIA inference optimisation NCCL_DEBUG=INFO— debug NCCL (collective communications library) for multi-GPU training
Summary
In short: NVIDIA is the dominant provider of AI compute hardware and software — its GPUs and CUDA ecosystem power essentially all frontier AI training, making it the most strategically important company in the current AI wave.