For applications with well-characterized working sets and pinned to a fixed set of machines, server-side flash makes a lot of sense, as long as you can size the cache to produce a high hit rate without breaking the bank. Virtualized workloads that often migrate between systems are generally better served with a hybrid array, since they make it easier to match flash capacity to the workload.
Another option, at least for file-based applications, is to front traditional NAS arrays with flash appliances such as Avere's FXT series that both virtualize the namespace and handle data placement to legacy NAS filers. Like hybrid arrays, this scheme allows independent optimization of flash and disk capacities.
As applications become rewritten for distributed cloud stacks, Google-like scenarios with applications using local storage and a centralized cloud controller dynamically managing storage pools that span physical servers will become more common for enterprise applications. In the meantime, software from the likes of PernixData and QLogic that can pool server-side flash and provide cache coherency across many servers is an intriguing alternative for applications running on VMware or traditional SANs.
Flash vendors promising all-flash performance at a hard disk price are no longer selling just a dream. Through judicious use of the latest storage hardware software, enterprises can come close to having it all.