The author mentions the CMU acquisition. I think their contention is that it's not enough, because Uber doesn't have access to the kinds of vehicle telemetry data its competitors do that would be necessary to train good self-driving car models. Google, and other companies with existing advanced vehicles on the road, have thousands of cars in the wild now with suits of sensors collecting telemetry and phoning home. Uber contracts drivers who drive their own cars, so it's not benefiting in a vehicle-telemetry sense from the drives it's arranging in the way it could be and its competitors do.

I think this is probably a solvable problem, though.