600 AI Architectures: What LLMs Default to When Designing Systems
Six LLM families produced 600 system architectures from identical briefs, exposing default tech choices, low consensus in key layers, and constraint-driven shifts.
A study prompted six LLM model families with 20 identical architecture briefs—without naming any specific technology—and collected 600 resulting system designs. The goal was to measure each model's default technology reflexes, not the quality of the architectures themselves.
The results reveal sharp asymmetry across layers. PostgreSQL (86%) and Redis (97%) emerged as near-universal defaults, while backend frameworks, hosting platforms, observability tools, and authentication systems showed over 100 distinct choices each with very low concentration—meaning there's no real consensus among models in these areas.
When a single constraint was added—budget, scale, EU data residency, or self-hosting—most models substantially altered their stack, with Kubernetes usage dropping to near zero in some scenarios. Microservices adoption also varied widely across models, ranging from 19% to 60%.
For engineers, the takeaway is twofold: teams relying on unreviewed LLM suggestions risk unconsciously converging on the same database and caching choices, while other architectural layers remain highly dependent on which specific model was consulted.