Most closed source applications contain no differentiating technical expertise, this is true. By "expertise", I am referring to deep expertise in computer science domains where the state-of-the-art is rarely published. The areas where you see large gaps in capability are domains that are intrinsically computer science limited e.g. most things involving high-scale/high-performance data infrastructure, real-time sensor processing, database kernels, etc. There is a surprising amount of highly evolved algorithms and designs that are not in literature but nonetheless show up in a multiple closed source systems.

Customers pay for the capabilities and performance that expertise affords. Most of it just manifests as "speeds and feeds" i.e. orders of magnitude more throughput, scalability, etc. In much rarer cases, there are manifest capabilities that don't have an obvious solution in published computer science, never mind open source (e.g. large scale semantic models of physical reality).

This is better described as "niche" expertise than anything significantly 'deep' or 'high-end'. Basically, stuff that (1) is not truly business-critical, at least compared to the closest FLOSS equivalent, and (2) is so inherently niche that it would not derive any benefit from the typical network effects that drive people to standardize on open source solutions. ESR himself discusses this case in his book The Magic Cauldron; one example he provides of something where FLOSS is probably not called for, is software for calculating cutting patterns for sawmills. A different one, is the business-management "secret sauce" that's embedded in many ERP packages. Note that the "not truly business-critical" proviso is quite relevant; for something that is critical to your business, however "niche" it might be, not having access to the source code would be an unacceptable risk!