Unlike Google's Bristlecone or IBM's Qbit computer, do simulators like Q# or Alibaba really use quantum mechanics anywhere in their physical chips? Are they just defining properties using a classical computer and trying to achieve quantum simulations ?
There is a distinction between what you use to write a program (the SDK), and what you use to run it (the backend).
The SDK can be either a graphical interface, like the IBM Q Experience or the CAS-Alibaba Quantum Computing Laboratory. It could also be a way of writing programs, like Q#, QISKit, Forest, Circ, ProjectQ, etc.
The backend can either be a simulator that runs on a standard computer, or an actual quantum device.
Simulators use our knowledge of quantum theory to construct the simulation program, but no actual quantum computing happens. It is just the standard chips of your own computer, or of a supercomputer they let you use, running standard classical programs.
This approach is something we can do for small quantum programs, but the runtime will become unfeasibly long for large ones. So if you notice that your job takes longer and longer to run as you add more qubits, you know that it is being classically simulated rather than run on a real device.
The only actual quantum devices that can be used are those by IBM, Rigetti and Alibaba. To write programs for these you can use the Q Experience, QISKit or ProjectQ for the IBM devices, Rigetti's Forest for their devices, or the Alibaba graphical interface for their device.
Microsoft are making hardware, and they hope that it will one day be used as a backend in Q#. But they have not yet gotten a single qubit, so we might have to wait a while. Until then it will be only simulators that can be used (or other companies hardware).
Quantum simulators don't rely on quantum-mechanical effects in the physical chips; instead they simulate certain aspects of quantum state and operations on it using only classical compute.
Universal simulators simulate full quantum state of the system, performing linear algebra transformations on it. They support universal set of quantum operations, but the memory required grows exponentially with the number of qubits simulated, thus the system size is limited to 30-40 qubits (depending on the hardware and the simulator).
Specialized simulators can focus on simulating certain aspects of larger systems at the cost of supporting a smaller set of operations. For example, CHP simulator (described in this paper) evolves only the stabilizer information in a matrix tableau instead of the full quantum state, and the required memory growth is quadratic in the number of qubits. The set of gates supported is limited to CNOT, Hadamard and phase gates, but it is sufficient to study quantum error correction protocols.