RTX is basically a marketing term.
The DXR/VKRT APIs and DirectML / Khronos equivalent expose new abstractions that can be implemented on current GPUs, but also allow for new HW architectures to accelerate them.
DirectML (API) and NV Tensor cores (HW) allow massive matrix multiplications to be expressed succinctly (API) and executed efficiently. Building specific HW specifically for this task allows a 10-100x speed up over using general purpose shading cores for the same task. So they're staying. NV is as much an 'AI' / deep learning company as a graphics company now (things like self-driving cars are driving their architecture choices as much as what gamers want). Apple has already made this kind of specialized matrix crunching hardware (aka "neural hardware") standard in their products. It's staying around. We can use it for image filters, but other apps on the system will be using it for voice recognition, facial recognition, spatial reasoning, etc, etc...
As for DXR, there's two major new HW developments that are required to accelerate it properly.
1) Fine grained thread scheduling. This emerged first in Volta. Traditionally with GPU work we want to Batch, Batch, Batch, and have the CPU send large workloads. In recent years with indirect draw/dispatch the GPU has been able to feed itself somewhat, but you still want to batch large amounts of threads together. However, with DXR, and surface might randomly want to emit a SINGLE ray, which enqueues a SINGLE ray vs scene query, which results in a SINGLE hit shader thread invocation. It was mentioned earlier in the thread that CPUs are better at RT for this reason... Well that's the hype/newsflash now! These new GPUs are built specifically to tackle this! Details are scarce so far, but they must be able to re-pack many individual thread requests into thread-groups for optimal execution. This is paving the way to entirely new execution models on GPU general purpose shader cores. It's basically CPU style delegates.
2) Fixed function hardware for ray-triangle intersection and BVH traversal, added now with Tesla. If we're going to be doing these operations a lot, then fixed function HW is going to beat shader cores. If your GPU doesn't have this HW, DXR can implement these operations using the shader cores, but on Tesla, you'll get ray-triangle intersection "for free"... This will stick around for the same reason that GPUs still have fixed function rasterization HW, depth buffer HW, and ROPs. It's also worth noting that while the API does allow drivers to provide 'black box' BVH implementations (in order to allow it to be baked into Silicon), you can also provide your own custom BVH and intersection shaders to be run on compute cores (which benefit from innovation #1, above).
That out of the way, in the future, the flexible GPU threading model (#1) above is going to have knock on hidden benefits EVERYWHERE. All sorts of workloads besides ray-tracing will be able to be mapped onto these future GPU's much more efficiently now, if this trend continues.
RT specifically will have slow uptake because the best-of-the-betst stupidly expensive HW can only manage something like 4spp at 4k 30Hz, and all other hardware can likely do <1spp... But it's on the roadmap now. We'll be slowly creeping towards the day where more and more of our images are built from ray queries instead of raster queries.
Games have already shipped using hybrid Raster/RT shadows (throwing a few ray queries in to improve quality / remove aliasing -- and this is well before specialized RT hardware existed!!!), so those kinds of tricks will continue.
NV has shown off a nice adaptive AA technique, where problematic edges are detected and a few extra samples can be collected in those areas on demand by casting some late rays.
Area shadows can benefit greatly - again, probably a hybrid shadowmaps + ray approach, where better estimations of the penumbra size can be obtained with a few ray queries.
Reflections are obvious and will be in all the tech demo's... However, even things like baking light probes for "traditional" reflections can benefit...
It's been mentioned that full path tracing will be too slow - that's true, but, I've worked on games where I've done a 30s light bake of the scene during the loading screen, to allow semi-dynamic scenes (e.g. user created, but static in gameplay). Tesla can do 300 billion ray queries in that time - probably enough for a nice path traced light bake