Daniel, Startups can take advantage of Google Machine Learning to build some amazing new apps for various vertical markets. We are making TensorFlow available via Open Source, and the Machine Learning APIs available for no cost or very low cost.
I think the most lucrative markets will be big data areas like Legal research, Patent research, Product research, medical/pharma research. In all cases there are mountains of existing data that no single company has tried to analyze. Think about clinical trials data, or years of past legal cases, etc. ML will have lots of uses for consumer products too, like the email, maps, and other things I mentioned. But, I think those consumer apps will take a long time to monetize. It is more likely that high value commercial projects will generate the early revenue.
The ML Vision APIs can be used to identify parts in a production process to spot defects. They can be used to review thousands of Xrays or MRIs or retina scans for medical purposes. It could be used for facial recognition in security/military intelligence/border security type applications.
If I were a startup I would focus on very complex big data in high value industries. Security, fraud detection, error detection, legal or medical research, etc. The self driving cars or rockets or other billion dollar projects will probably be the domain of huge multi-national companies.