AI Startups
AI Infrastructure Stack Blueprint
Reference architecture for an AI startup at seed → Series A → Series B. GPU sizing (L40S vs H100), inference runtime (vLLM vs TensorRT-LLM), vector DB selection, RAG patterns, observability.
24 pages· PDF· Founder, CTO, ML Engineer
What's inside
The wrong AI infrastructure decision at seed becomes a six-figure mistake by Series A. This blueprint is the reference architecture we ship to AI-native startups across the funding curve — what to provision at seed, what to evolve to at Series A, what to harden at Series B. GPU sizing, inference runtime, vector DB, RAG patterns, and observability — all with concrete vendor recommendations.
Table of contents
- GPU sizing: L40S vs H100 vs MI300 — when each makes sense
- Inference runtime selection: vLLM, TensorRT-LLM, llama.cpp
- Vector DB shortlist: latency, scale, operational realities
- RAG patterns: chunking, embeddings, retrieval ranking
- Observability stack: traces, evals, regression detection
- Seed → Series A → Series B evolution path