A typical answer would probably state that "good" data and an well-built model would be components of a successful machine learning deployment, but I believe that the more qualitative sides of ML development are much more necessary. As issues of bias, fairness, and explainability have become more prominent in AI/ML development, ensuring that your systems are usable and equitable for all populations is extremely important. Bias can be encoded both in data and the models themselves, so it's also important that this be prioritized for successful ML deployment.
Obviously, you'll need GPU/TPU capacity to efficiently train your algorithms, but this is definitely a barrier for those coming from low-resource regions. Efforts like Google Colaboratory have helped somewhat but can be challenging to use if you don't have sufficient electricity or internet access.