I can neither confirm nor deny that AWS is working on an API call for placing a fleet of satellites in orbit.

On a more serious note, I'm excited to see what happens to the NTP pools in the area, since GPS is often a timing source for them. It should be a non-event, but it could get interesting.

Oh, and what happens to CDMA cell towers that use GPS to sync? Do they go off-line?

Anywhere that GPS is used in a timing-critical application, they'll be using a timing receiver (like http://www.trimble.com/timing/thunderbolt-e.aspx) to handle their time reference; those use the GPS signal (when they have it) to discipline a (ovenized) quartz crystal, so that they have good performance during "holdover" (loss of GPS) especially for short-term outages. In its default configuration, the Thunderbolt-E linked above, for example, will be within +- 8us after being in holdover for 24 hours.

A Rubidium reference gets you down to the 1us/day range, cesium gets you down to a tenth of that. It all depends how much you're willing to pay.

±8µs can be rather bad, e.g. for financial applications.

There is certainly hardware available that can do better than this, but the parent asked about cell sites in specific (and cell sites are one of the top users of the Thunderbolt). You can get GPS hardware based on atomic references, for example, that will do better over these kinds of timespans.

(You really need GPS or some other precision time transfer mechanism for long-term accuracy, though ... there's no getting past that.)

Well...if you're in that position, you probably have money to spend on failover config. Pretty accurate hardware clock modules were available even before GPS. Not cheap, not simple, but reliable - if you have the need and the money.

Anything relying on GPS for timing is going to revert to an onboard crystal timer for the outage duration.