The conversation every technical founder has
At some point, every company building on top of AI APIs asks the same question: should we just build our own model? The reasoning sounds compelling. Control over the model. No API costs at scale. No dependency on OpenAI or Anthropic or Google changing their pricing.
The reasoning is almost always wrong.
The cost you're not calculating
Training compute is the headline number. It's also the smallest long-term cost. The real costs: a dedicated ML team to maintain and retrain the model as your data evolves. Evaluation infrastructure to catch regressions. Serving infrastructure sized for your peak load. Safety and alignment work to prevent the model from doing things your users will blame you for.
API providers have teams of hundreds working on these problems. For most companies, competing with that investment makes no sense.
The model is the easy part. Everything around the model is where the real work lives.
When it does make sense
Fine-tuning on your own data — yes, often. Running open-source models self-hosted for privacy or cost at scale — sometimes. Training from scratch — almost never, for almost any company not named Google or Meta.
Use the APIs. Invest in your data layer, your prompting strategy, your evaluation pipeline. That's where your competitive moat actually lives.