NVIDIA has a monopoly on AI training GPUs — there is no real alternative and they set the price at whatever they want

An H100 GPU costs $25,000-40,000. A DGX H100 system (8 GPUs) costs $350,000-500,000. Training a frontier LLM requires 10,000-30,000 H100s ($250M-1.2B in GPU costs alone). NVIDIA's data center GPU revenue was $47.5B in FY2024 with 80%+ gross margins. AMD's MI300X exists but has 30-40% less software ecosystem support (CUDA vs ROCm). Google's TPUs are not sold externally. Intel's Gaudi is 2 generations behind. Every AI company, from OpenAI to a university lab, is dependent on a single vendor with monopoly pricing power. So what? NVIDIA's monopoly means: (a) GPU costs are the #1 expense for AI companies, consuming 60-80% of total funding, (b) startups cannot compete with incumbents because they cannot afford GPUs, (c) AI research is concentrated at wealthy institutions that can pay NVIDIA's prices, and (d) NVIDIA captures most of the economic value of the AI revolution — not the companies building AI products. Why does this persist? CUDA. NVIDIA built CUDA in 2006 and spent 17 years building the software ecosystem (cuDNN, TensorRT, NCCL, Triton). Every ML framework (PyTorch, TensorFlow, JAX) is optimized for CUDA. Switching to AMD requires rewriting kernels, debugging ROCm compatibility issues, and accepting 10-30% performance regressions. The switching cost is so high that even companies that hate NVIDIA's pricing cannot leave. The moat is not the hardware — it is the software ecosystem.