The trade, stated without ideology
APIs (OpenAI, Anthropic, Google): the best models available, zero infrastructure, per-token pricing, and your data crossing to the vendor's servers, bounded by enterprise no-training agreements if you pay for them. Self-hosted (open-weight models like Llama, Mistral, Qwen on your GPU infrastructure): data never leaves your walls, no per-token fees, full control, and you own the hardware, the ops, and the quality gap at the top end. The integration practice builds both; this page is the decision.
When self-hosted genuinely makes sense
- Data that cannot leave, legally or contractually: regulated industries, classified internal data, NDAs that prohibit third-party processing, where even enterprise API terms are insufficient. This is the category where self-hosting is the answer, not an option.
- Volume where per-token economics invert: very high, steady workloads where flat hardware cost beats metered tokens, real at scale, and the math must include ops salaries and GPU refresh cycles, not just the invoice.
- Latency and availability control: on-prem inference has no network round-trip and no vendor outage, matters for some embedded and real-time uses.
- Fine-tuned proprietary behavior: models adapted to your domain, kept out of shared infrastructure.
When APIs win, the common case, stated honestly
For most business features, chatbots, RAG over documents, content tooling, extraction, the frontier API models are simply better than anything self-hostable today, and the enterprise no-training tiers resolve the privacy concern at a fraction of self-hosting cost. The self-hosted top models trail the frontier; the gap costs quality exactly where AI features earn their keep. Add the ops reality (GPU capacity, model updates, security patching) and the API wins for most use cases most of the time. The integration architecture keeps the choice open either way: the abstraction layer means a self-hosted endpoint is a configuration change, not a rewrite, build on APIs now, graduate to self-hosted for specific classes when the data or volume case is proven.
The middle paths most businesses actually use
Hybrid by data class: APIs for public and low-sensitivity work, self-hosted smaller models for the sensitive tier. Dedicated instances from vendors (private capacity, data isolation) without owning hardware. And the readiness checklist first, because the data classification that decides this question is work you needed anyway. The decision runs on your data classes and volumes: the brief starts the evaluation with both options priced, honestly.