In order to build substantial quantities of stuff, molecular manufacturing systems will probably rely on the ability of systems to fabricate more systems.
(Note that this does not mean that tiny self-contained robots will have to build more tiny self-contained robots. Small manufacturing systems will not be self-sufficient, simply because that would be substantially less efficient and harder to design.)
The point of molecular manufacturing is to use rather intricate molecular machines to fabricate stuff. There are other ways to fabricate stuff, such as bulk manufacturing and self-assembly. Bulk manufacturing is what we do today in human-scale factories, so I won't say more about it.
Self-assembly has a limited and difficult design space for fairly fundamental reasons. If you shake a bunch of parts together and the correct assembly comes out, then the parts must have incorporated in them, in one way or another, the information about how the final product should turn out. In effect, each component not only has to function as a component in the final assembly, it also has to include a purpose-designed manufacturing (joining) tool that is different from every other tool in the other parts, and that functions within a very tight band of constraints (energy level near thermal noise). Although protein and DNA demonstrate that such systems can make a wide range of products, the difficulties of protein design and the physical weakness of DNA reflect the practical physical limitations that self-assembly must confront.
Mechanically guided assembly spans a range from enzymes to (so far theoretical) nanoscale robot arms manipulating molecules under direct computer control. On the enzyme side of the scale, enzymes that make copies of themselves from carefully engineered feedstock have already been produced. This is of course too limited for general-purpose product making.
Ned Seeman's DNA fabricator is a step toward generality. It is a machine made of DNA, programmed by DNA, that can make one of four DNA strands. Its product space has a size of four. The machine itself is far more complex - there are vast numbers of DNA sequences of comparable size, and probably thousands of devices that we would recognize as machines. So the machine can't come close to making a copy of itself.
The special-purpose manufacturing machines that Drexler has been arguing for, over on the Metamodern blog, have a related problem. Each machine has only a tiny fraction of the operational flexibility needed to make a complex machine, and so a factory of such machines that could build any machine in the factory would have to contain an immense number of different machines. Even if special-purpose machines built parts that were assembled by general-purpose machines (which is what Drexler is arguing for), there would still be a large number of parts, each needing its own machine.
We can chart a continuum from a system using some special-purpose and some general-purpose machines, to a system using all general-purpose machines. A general-purpose machine need not be much more functionally complex than a special-purpose machine, so it may be that the earliest manufacturing systems will be more on the general-purpose side.
A general-purpose machine such as a robot arm must have at least one actuator, several kinds of structure and bearings, and perhaps some sensing (though the most efficient versions will move deterministically, with no sensing required). It must also have inputs from a control computer, which probably means direct drive of its actuators. This is quite a lot of functionality. All these pieces of functionality will have to fit together and work together.
Either a general-purpose molecular manufacturing system can be made a lot simpler than anyone thinks (including me), or it will have to be designed in modular pieces, just as modern software is written in modular pieces. The pieces will have to be well understood, and will have to play well together in a variety of configurations, because the designer of one piece probably will not understand every single detail of every other piece that they will have to interact with.
If a highly functional mechanical system can be designed from modular, well-understood components, that implies that other systems can be designed from the same components. In other words, the design space is probably pretty large. A wide variety of things that are simpler than (for example) a robot arm can presumably be built by the same general-purpose system that builds the robot arm. For that matter, a number of different robot arms can probably be built. And products that are combinations of simpler products, re-using their functionality in new ways, can probably be built: for example, several robot arms can be linked into a shape-changing truss.
More generally, it is more likely that the design space of an engineered system that can build all its components is quite a bit larger than needed, than that the design space is only exactly as large as needed (which would imply massive and useless efforts toward elegance on the part of the designers).
At the end of a short chain of reasonable suppositions, I reach the conclusion that even a primitive molecular manufacturing system will probably be able to make a broad and useful range of products, with each product requiring significantly less design work than went into the manufacturing system itself. To me, that says that even a primitive molecular manufacturing system has the potential for substantial economic impact on a rapid time scale.