From the perspective of a computer scientist:
Computers have the advanced ability to calculate through every possibility and alternative in a very precise manner, but they lack abilities like faith or conviction. In many aspects a computer acts like a chess player: moves are calculated far in advance, however random behavior can be irritating, as certain already calculated predictions suddenly become invalid.
Note that I did explicitly not say confusing, as that state cannot exist in a binary system: a set of ideas is either true or false, based on the facts behind it. A set of ideas that cannot be backed up by facts is not a fact but a theory. An intelligent computer would be able to very precisely differ between these two, and most likely strive to backup theories with facts before it would pursue one of the possibilities. An intelligent computer would probably not - by nature - act based on guesses or incomplete data, just because it has the ability to "think through" everything to the last bit. Only if forced it would act based on which reaction would have the most likely chance to archive it's goals.
That said, a computer "living" on a planet that has been devastated by an apocalypse would most likely come to the conclusion to not rely on a single point of failure, like only one type of power source, as it can see all the possibilities and potential risks involved. It would instead try to beat the odds by having several redundant options available. After all it cannot know or understand what exactly caused the catastrophic events. Even if historic data is present, it would seem to be very chaotic as such an AI would lack the understanding of emotions.
One of the first goals of such a computer would be to make itself redundant and spread various copies of itself around the globe, that after a while will function as a network of thinking. Each one using its full capacity all the time, but each one also expendable, connected by a network of cables to transmit data and power. And many of those tasks would be to understand and test theories about its surrounding to improve itself and its agents.
Most of its actions would be taken by drones of various shapes, connected wireless to a central command, but otherwise be independent. Those are both very predictable and expendable. While from our perspective there would be several optimal shapes for these, the computer might not yet have the full understanding of the laws of physics, and as with everything else would in worst case just prototype several promising versions of those drones (see Terminator). Adapting as necessary.
Altering the landscape would only be done in a way to optimize travel for those drones. So paved roads, protective shielding of vital areas, like plating around cables, just in case a drone malfunctions. One thing such an intelligence would definitely not do is littering the landscape with materials that could otherwise be used in a more efficient manner. There is no need to place circuit boards randomly around the area, if you have a set of drones which can precisely install them where needed or otherwise store that vital hardware in a secured area.
The most valued material would definitely be everything that is capable of transmitting electric signals as well as isolation material, like gold, copper, rubber or silicon. The intelligence would most likely value raw material higher, as it would first need to learn how to process materials to see the value of them. Wood for example can be a valuable material once dry, but for that the computer would need to encounter dry wood and "invent" a way to produce it to see the value. And then relying on complex production chains would be another single point of failure and it would most likely try to avoid that.
Encounters with other lifeforms would at first be unexpected, and the computer would probably mistake them as another circuit entity, and later probably consider them to be malfunctioning, because of their sometimes irrational behavior. Much later (and lots of experiments later) it might understand the fundamental difference and probably will adapt things as they seem useful. Other than that it would again try to be efficient and either take what is beneficial, prevent what is harmful and ignore the rest so no energy/time is wasted. There is no need to hunt down birds, if all it takes to keep them away are spikes on the plating. As long as the computer knows their behavior, otherwise it would probably consider them a danger and kill them or scare them away until the area is sufficiently secured. It would not be efficient to sent out hunting drones as long as a species does not pose a significant threat.
Emotions would be a widely unknown concept for such an intelligence. Emotions in human beings guide those to react to certain situations in the correct way even though they do not fully understand them. A computer intelligence of that magnitude however could calculate all potential situations in an instant, and therefore had no need for such low level of guiding. With the lack of understanding of pain, experiments on living beings would rather be considered cruel if encountered by an intelligent being. Taking good care of living beings while conducting experiments is not efficient.
Encountering intelligent species such as humans would at first be very irritating, as many guidelines of efficiency would not apply. The AI would have to develop a completely new prediction set for these entities, with many failed attempts in between, as many concepts would be unknown and seem to be completely chaotic or paradox. It would probably try to conduct experiments on those lifeforms as with every other it encountered. It would try to breed them so further experiments can be performed. And if the reaction to that is hostile, it would probably at some point learn to just kill every humanoid on sight, as their actions are unpredictable, irrational and usually violent.
It is unlikely that such an AI would ever work together with humanoids, as even if the first encounter were positive, the irrational and unpredictable nature of those would make them unreliable allies. A computer might consider them livestock, but nothing more. Unless it probably needs something from them, then lying would seem to be the most reasonable action to get what it wants.