Are quantum chips supposed to replace the current CPUs and GPUs? Or are they supposed to be "just" another component that we connect to the CPU, the same way we connect GPUs?
Almost certainly the latter. QC is very good at highly specific jobs, like factoring or searching. Similarly to GPUs, getting the data from the CPU to the GPU is costly. For now and presumably for a very long time, QC has that problem but way worse. (And it wouldn't give you speed ups yet even if the cost was free -- but Martinis asserts that won't take long.)
For a very select set of problems (factoring and discrete log), quantum computers are exponentially faster than classical computers. For a few (including np-complete problems), they are quadratically faster. For everything else, they're no faster. (When I say "faster", I really mean the runtime of the best known quantum algorithms is better.)
For the forseeable future, quantum computers will be much smaller than classical computers -- the article is about Google building a 49 bit QC and how that would be a breakthrough. So for the forseeable future, they'll be separate components, used for special cases.
A little further: factoring and discrete log aren't a complete set (also, that depends on algorithm development). There are a few academic problems that are also exponentially faster, and, more generally to factoring and discrete log: hidden subgroup problems (which are what killed non-supersingular isogeny Diffie-Hellman as a post-quantum key exchange).
We are so far from talking about this that your question is nearly meaningless. Quantum computing has not even been demonstrated yet. This "quantum supremacy" is for a highly contrived algorithm and set up that has very little if any real world use cases.
A generalized quantum computer able to run standard computing algorithms is very far in the future and so much basic research in computing science has to happen before it can be talked about meaningfully.
We actually have a few examples of an absolute speed up for quantum computers, though none of them are exponential and some of them are problems nobody cares to solve. For example, Grover's algorithm is superior (O(sqrt(n)) vs O(n)) to the best possible classical algorithm for the same problem.
If you are willing to accept comparisons of best known classical algorithms to best known quantum algorithms (without a guarantee that the existing algorithms are ideal) you can add others like factoring to the list, with exponential speed ups.
Only a handful of useful algorithms (currently) that exploit quantum computers (see https://en.wikipedia.org/wiki/Quantum_algorithm). Shor's algorithm for integer factorization is the one that would radically change the current crypto situation (goodbye RSA).