I wanted to reply point-by-point to this, but no doubt the post would be far too long, so I'm just going to summarize things (sort of). VA has already asked my opinion on this and I'm going to give some of that same info here as well as add a few new points...
You have to remember that almost all the work at optimization has gone into the texture side of things, both in rendering performance and memory usage. It's not really going to be any better than it is now, and is in fact going to get much worse. Remember that normal maps aren't easily compressed, since the compression utterly destroys quality in the case of a normal map. That means most/all normal maps are NOT going to be compressed. Adding a normal map like that to a fighter is going to, in the best case, more than double the memory requirements for that model. Normal maps are going to comprise from anywhere to 65-80% of the models texture memory requirements. That is a complete killer considering the already extremely high memory usage that we already have from textures. All of the work that can be done to improve memory usage has already been done, there isn't much more that we can get out of it. And memory usage is only going to increase in the future. Shaders require memory, new animations require more memory, more explosion types require more memory, render targets require more memory, FSAA requires more memory. And we also have possible future support for light maps and height maps which is also going to increase memory usage that much more. Memory usage is our killer, it's only going to get worse, and everyone has to not only realize that but also plan for it.
A normal map is also not the same as any other map since it requires much more processing work to render. It means per-pixel lighting, and that costs GPU performance, and it also requires more CPU time to compute the tangent space coords for each and every object that gets rendered, every single frame. A diffuse/glow/spec map, combined, cost about as much performance as adding a normal map. It does add more realizism, but it most certainly is not coming free, even if you exclude the large memory requirements. A normal map isn't going to add much in the way of detail to a fighter than an extra 1-2k of polys will. The difference is that using the normal map will actually be slower than going with the extra polys. So don't depend on normal maps to save you from anything, because they won't. They will add extra detail to a model, but if it's not on a large ship then that is going to be wasted memory and wasted processing time. Shaders will allow more efficient use of normal maps, but it's not going to do anything for the memory usage. For every normal map that you add you are taking something away, whether that be better explosions or more weapons effects or whatever, it's going to cost you.
I gave VA some basic poly counts that I want to see in any current fighter and bomber. Those in the 8-10k range for fighters, and 9-12k for bombers. My counts for larger vessels are up in the air, but I fully expect 25-40k to be the norm before long. I mentioned already that optimizations have been done to increase the performance of texture handling, but almost nothing has been done for geometry yet. There are numerous areas in the model rendering code which are creating bottlenecks. The collision detection code is also rather crap with regards to large poly models. Object culling and render order are also something that is horribly wrong with the code. These are things that are going to get fixed. Even if the hi-poly models are a little bit of a strain now, future code upgrades probably won't even flinch at them. I already use 60k+ models for testing, so handling large poly models is possible even with the current code. And it is only going to improve. Vastly.
We can always tie in better detail control to allow users to tone down the poly hit if needed. But what we can't do is take a lower-poly model and make it look better later on. Use textures efficiently and creatively, and then increase poly counts. We added the code to handle textures better, but the biggest problems are in the model rendering and collision detection code. Once we work those issues out the performance should increase significantly. So plan ahead and give your models that extra 2-3k of poly detail now (provided it's not wasted obviously) and the code will catch up to it. Even your basic video card these days can easily handle the poly counts that we are using, it's the code that has trouble, but that we can do something about.
Go with the the larger poly counts, texture your models efficiently, use detail boxes, avoid poly waste, avoid using normal maps unless it truely adds to the model. These are all basic things that should not only improve the model overall, but also improve how that model works in-game. Lighting will be enhanced with more polys, texture mapping (if done properly) will be enhanced with more polys, and general "coolness" will as well. If you are making a model now then you might as well plan for the future, because we are. We are planning for more models in a single frame, and more efficient collision detection, and more efficient rendering, and much improved user detail control. If you only make your model for what you can do now then you are only hurting yourselves, because in another year those models will be low quality in comparison to what you will be able to get away with.