The AI stack · Explainers · Amit Patnaik

What sits behind an AI answer?

This explainer maps the layers behind modern AI systems: power, fabs, chips, server systems, CUDA and kernels, frameworks, model weights, inference serving, and applications.

Hardware: power, semiconductor process, packaging, HBM, GPUs, accelerators, and networking.
Software: kernels, CUDA, PyTorch, JAX, TensorRT-LLM, vLLM, SGLang, and serving runtimes.
Bottlenecks: memory bandwidth, KV cache, packaging capacity, cluster networking, and power interconnection.

Sources and dates

Published April 2026. This interactive reference explains concepts; its vendor comparisons, market estimates, and purchasing examples are historical and have not received a complete current-source review. Check the original vendor or official source before acting on a specification, price, or legal statement.

Concept corrections, 26 September 2026. Sources: GPT-4 technical report · Mixtral paper · InstructGPT paper · DPO paper · DeepSeek-R1 paper · PyTorch 2 paper · CUDA guide · Triton documentation · TSMC CoWoS.

Additional references: Agent design · Serving and memory · Speculative decoding · Quantisation · NCCL communication · Chip manufacturing · Power supply.