There are two extremes regarding the design of self-driving vehicle infrastructure and dependent systems, completely independent and connected.
The first extreme is seen by the Google Car which originally aimed for NO STEERING WHEEL. Each vehicle is autonomous and independently operated, similar to a human. Though in this case the human intelligence is replaced with "artificial" intelligence.
However, there are several problems. It's not clear that their "exists" enough data to properly train the vehicles. The sensor hardware technology is still lacking, underdeveloped, and expensive. Furthermore, completely ignoring vehicle to vehicle communication completely ignores the potential that a network of communicating vehicles offers, especially at high speeds and with platoons. These vehicles also must continuously contact with Google's centralized services, incurring a large latency and bandwidth cost.
The benefits are mainly due to security. The "trusted" centralized service removes the need to communicate and "trust" other vehicle manufacturers.
The other alternative is a distributed vehicular network utilizing vehicle to vehicle communication. Vehicles can notify each other of their intentions such as braking, turning, etc. Computations can be shared and collective intelligence is leveraged.
However, the downside is the security risks. Trusting other vehicles computations or even safety messages invites a host of spoofing and tampering attacks.
So which way will the future unfold? Maybe not to either "extreme" though a compromise somewhere in the future. It's clear that independence is useful, though at the same time the notion of collective intelligence scales very quickly especially with the number of vehicles on the road.