There are ways to achieve fault-tolerance, even for a centralised system (e.g, by maintaining an active/passive replica). In cases like this, where there is such a great gap between the performance of current systems, you can always "waste" another node(s) for fault-tolerance and still operate with less cost, if you want.
I agree that some types of computations (e.g., multiple/distributed sources) may not be benefited by a centralised approach and I definitely don't claim that this is a solution for everything. However, the point is to criticise the design choices that we make for a streaming system. Streaming support, even for popular systems today, is something like an extension (sometimes a hack) on top of the core of the system, hidden beneath multiple layers of abstraction. In addition, modern systems try to do many things at the same time (support AI, batch & stream processing, connectors to publish-subscribe systems, multiple wire protocols) and they end up doing most of them poorly.