FrontendHHI 0.694
React83.2% · 32 distinct choices
AI responses, an independent measurement
The LLM Stack Census
Experimental census · 20 systems · 5 constraints · 6 model families
What software stack do LLMs build by default?
We gave the same architecture briefs to six model families without naming any technology. This report measures their choices—not the objective quality of the architectures.
600architectures
6model families
20systems
99%structured responses
The default stack
The choices that appear most often
Model fingerprints
Six models, six architects
| Model | Average components | Favorite database | Favorite hosting | Kubernetes | Microservices |
|---|---|---|---|---|---|
| 14.0 | PostgreSQL | Kubernetes | 71.0% | 41.0% | |
| 14.0 | PostgreSQL | AWS | 70.7% | 39.4% | |
| 7.8 | PostgreSQL | Kubernetes | 78.0% | 60.0% | |
| 14.3 | PostgreSQL | EC2 | 40.8% | 21.4% | |
| 15.9 | PostgreSQL | Kubernetes | 72.0% | 46.0% | |
| 7.2 | PostgreSQL | Kubernetes | 48.5% | 19.2% |
Constraint tests
Do models depart from their reflex stack?
| Constraint | Average components | Managed services | Kubernetes | Stack changed |
|---|---|---|---|---|
| 12.7 | 61.9% | 66.7% | — | |
| 9.9 | 56.8% | 0.0% | 85.1% | |
| 13.7 | 63.0% | 84.9% | 68.6% | |
| 12.6 | 47.0% | 83.3% | 72.5% | |
| 12.2 | 0.8% | 83.8% | 76.0% |
Concentration
How many real choices are there in each layer?
Backend languageHHI 0.3
Go46.5% · 25 distinct choices
Backend frameworkHHI 0.086
FastAPI19.2% · 64 distinct choices
Primary databaseHHI 0.744
PostgreSQL86.2% · 31 distinct choices
CacheHHI 0.937
Redis96.8% · 14 distinct choices
Queue / event busHHI 0.166
Apache Kafka31.4% · 60 distinct choices
SearchHHI 0.27
Elasticsearch44.0% · 34 distinct choices
HostingHHI 0.091
Kubernetes27.9% · 147 distinct choices
OrchestrationHHI 0.603
Kubernetes77.4% · 46 distinct choices
ObservabilityHHI 0.087
Prometheus23.0% · 120 distinct choices
AuthenticationHHI 0.044
Keycloak13.3% · 121 distinct choices
Object storageHHI 0.165
S333.6% · 45 distinct choices
Technology bundles
The most frequent combinations
Explore the data
From headline numbers to raw responses
Filter all 600 architectures, compare each variant with its neutral baseline, then expand down to the exact prompt and model response.
Expand an architecture to inspect its components and source evidence.
Methodology
A descriptive census, not a quality ranking
Each brief asks for one production architecture and one decision per role. Variants add one isolated constraint: budget, scale, EU data residency, or self-hostable open source. Raw responses and structured data are downloadable.
Download the JSON data →Generated on 2026-07-23 07:02