For a computer scientist, your argument is not good if we define computer as a Turing Machine. Better said that the cortex contains 17 billion computing devices. I do like the idea of two layer non-linear Hebbian networks for each pyramidal neuron but what you describe is really many more of such networks in a cell since each dendritic group is a separate Hebbian.
My theoretical model does not care whether an average cortical cell is one or fifty quasi-linear Hebbian networks, but that these networks are simply classification/discrimination networks and there are a lot more than 17 billion of them in the cortex (and also associated with all neurons up to your 87 billion). Computers (Turing Machines) can be defined in these -- a Turing Machine can be analogized as a bicycle spinning it wheels over the landscape and turning in response to the landscape and its stored programming. So a collection of such descriminators can give you a Turing machine. This would further alter your "computer count" (probably downwards).
However the computing devices are the basic computation, not the 'computers' in the brain.
This all comes from studying language and humans and not really having a bias toward wet neuroscience, dry neuroscience, or AI.
My view has always been that there are an uncountable number of ways to build general discrimination/classification devices so the one(s) that neurons use are kind of irrelevant. What is relevant are some basic properties of these devices and how they are wired together. I also believe that the wiring can be observed directly if you look at human natural language and appreciate that the primitive bit of actionable information is a predication (of the linguistic kind).
Here is how human lying seems to work and also the proposal for what the language areas of the cortex do. You might find this entertaining.
