Why We Chose .NET 10 Over Python for an AI Product

Python is the default for AI, but we chose .NET 10 for AIML.chat because we are not training models — we are calling them over an API and building reliable, concurrent, multi-tenant infrastructure around them, where static typing, performance, and a long support window matter far more than the ML library ecosystem. The conventional choice is right for research and wrong for our kind of product.

The model is an API call

The case for Python in AI rests on its ML ecosystem — the training frameworks, the notebooks, the research tooling. That ecosystem is decisive if you are training or fine-tuning models. We are not. AIML.chat calls a hosted model through an API for chat and embeddings, which any language can do. Once the model is a network call, the Python advantage mostly evaporates, and the real question becomes which runtime builds the surrounding system best: the ingestion pipeline, the vector retrieval, the streaming, the billing, the multi-tenancy. That is an infrastructure question, not a machine-learning one.

Where .NET wins for this job

Static typing is the biggest day-to-day win. In a multi-tenant system where a leaked tenant boundary is a security incident, having the compiler enforce contracts catches whole categories of bugs before they ship, and makes refactoring safe as the product changes weekly. .NET's performance handles concurrent SSE streaming without tuning gymnastics, and the LTS release gives a young product a stable foundation rather than a runtime that shifts underneath it. These are unglamorous properties that compound into shipping speed and fewer incidents.

Closing the ecosystem gap

The one real gap — AI library maturity — is closing fast. Microsoft.Extensions.AI provides clean, provider-neutral abstractions for chat and embeddings, so we get the integration we need without leaving the platform, and without binding ourselves to a single model vendor. For the operations a product like ours actually performs — embed, retrieve, prompt, stream, meter — the .NET AI tooling is more than sufficient, and what it lacks in research breadth it makes up in production fit.

Choosing for the work you actually do

The lesson is to choose the stack for the work in front of you, not the work the industry is famous for. If your product trains models, Python's ecosystem is hard to beat. If your product is reliable infrastructure that calls models, the criteria flip toward type safety, performance, and supportability — and there .NET 10 is an excellent, slightly contrarian choice. We optimized for the system we are actually building, and the unconventional pick has paid off in fewer bugs and faster, more confident iteration.

Getting started

Try what the stack produced on the AIML.chat free plan: index your content and see grounded answers stream in. Read building an AI agent platform on .NET 10, using pgvector for AI embeddings, and compare tiers on the pricing page.