The Build vs Buy Matrix

• Frontier APIs — rent the best models by the token. Fastest to market, zero ML headcount, costs that scale with usage. The catch: your margin lives on their price list, and your capability is everyone else's
• Open-weight models — download Llama- or Mistral-class weights and run them yourself or via a host. You own the model and your data never leaves your walls — but you now own GPU ops, latency tuning, and the upgrade treadmill
• Custom training — train or heavily fine-tune from scratch. Rational only when AI is the product, or your data enables something nobody can rent

The honest default for most products: start on APIs, move to open weights when volume or privacy math forces it, custom-train almost never.