Many AI startups build on external foundation-model APIs. That can accelerate product development, but it also creates dependency on providers whose pricing, capabilities and policies the startup does not control.

Map provider concentration

Determine what percentage of critical workloads depend on one provider and whether alternatives are technically viable. A backup provider that has never been tested is not a real fallback.

Understand switching cost

Prompts, tool schemas, embeddings, safety behavior and evaluation results can be model-specific. Estimate the engineering and quality cost of migration rather than assuming APIs are interchangeable.

Stress-test pricing

Model how gross margin changes if inference prices rise, discounts disappear or usage shifts toward more expensive modalities. Dependency is more dangerous when pricing power cannot be passed to customers.

Ask where differentiation lives

If a better base model immediately erases the startup’s advantage, defensibility may be weak. Stronger companies add workflow integration, proprietary data, distribution, identity, evaluation or domain-specific systems around the model.

Review reliability architecture

Rate limits and outages can become the startup’s outages. Look for routing, queues, graceful degradation and observability.

Consider policy dependency

Provider terms, safety policies or supported regions can change. Companies serving sensitive or regulated workflows need contingency plans.

This risk belongs in the broader moat analysis described in AI startup moats.

Using external models is not inherently a weakness. The investment question is whether the startup has designed the business so that model providers remain inputs rather than becoming the entire product advantage.