So one question I is what kinds of applications are people developing
with GAE? It looks like the pricing of computational units (I forget
the exact naming) makes it plausible to stage compute intensive
applications.
I could for example see the utility of providing solver
libraries...that way obvious users of say Linear Eq. solver wouldn't
have cost of transferring library to GAE, just parameters.
Also, have you done performance comparison of GAE & Amazon's service?

The most salient design factor is that GAE has slow, blocking writes
and very fast, massively parallel reads. This makes sense if you think
of Google's general architecture: massive crawling operations that
fill the BigTable, massive MapReduce operations that write indices,
and massive web delivery. We only get to use the latter for now

To design accordingly one approach is to precompute your data and
indices and bulk upload. Dynamic uploads are expensive. For example,
there's a project to implement MapReduce on GAE, but I doubt if that
will be cost effective because it implies large dynamic uploads.

Two other ways to get data into your application, both rather low
bandwidth:
Writes at the scale of an interactive user are ok, though the user may
experience the delay.
Fetch a URL, especially through Google APIs, and usually through XHR
calls to catch potential timeouts.

Not yet, because GAE is still early. I think Google hasn't tuned the
service or price models.