I think it's fair to assume that the investments google is making in huge new data centers across the U.S. are aimed more at hosting network services like gmail and goffice than at making search faster or more effective. There are power use implications in this - sign up for organization wide gmail usage, for example, and you no longer have to power up your own mail servers.
In the more general case it's possible to see the return of time share services like goffice as a replay of a different kind of power balance altogether: one involving the distribution of organizational power between users and some central data processing authority.
What mainly drove time sharing in the early seventies was the disjunct between what science based computing could deliver to users and what data processing actually did deliver. Thus what companies like Digital Equipment Corporation, Data General, Wang, Honeywell, and the rest promised the mini-computer (i.e. non-mainframe) customer was productivity through user empowerment and the application of computers to tasks ranging from word processing to job scheduling and communications - all applications outside the data processing tradition and therefore stuff data processing management dismissed as both trivial to do and undesirable.
Unfortunately for user management, however, few senior managers understand that the only thing data processing has in common with computing is the use of computers - and data processing managers were therefore generally able to leverage their existing budget clout and privileged positions within Finance to stop or slow user technology adoption where capital costs were significant. Time sharing answered that challenge by letting user managers sign-up for services without triggering internal capital cost approval processes - and that produced most of the bang behind the time share services explosion in the 1970s.
Today most big corporate computing looks, from a user management perspective, pretty much the way data processing did in the 70s: largely incompetent, terribly expensive, terminally arrogant, and dramatically out of touch with user needs.
As a result we're seeing a resurgence in time sharing: companies like salesforce.com fundamentally sell on the same promise of competence, responsiveness and user control over costs and processes that made money for GE, Honeywell, I.P. Sharp, and thousands of other time share services suppliers in the 70s.
From an organizational power perspective, therefore, those google data centers look like a replay in the making. Tomorrow's google may be yesterday's third party data service: responding to the same market in the same way - and in long run just as doomed. What happened then was that it took the better part of fifteen years and IBM's colossal misjudgment on the PC for time-sharing to die - today, however, the successor technology is already in the market and the handwriting therefore on the wall before google even finishes building its data centers and gets around to formalizing the product suite.
The key reason for that is simply that owning all of a resource always provides better control than renting a chunk of it - and that user opposition to service repatriation therefore fades as costs and risks go down and the reality of the time sharing option modifies the attitudes and behavior of those in charge of the centralized IT service.
So does this mean that google's data centers won't make money? No; but what it does mean is that google's success as the anti-Microsoft will last just as long as Microsoft's willingness and ability to support the user enemy it originally opposed and has now become: data processing.