This is distinct from classical computers, where you can describe a bit without needing to describe the other bits in the computer. You cannot describe a qubit (in a way that's computationally useful, at least) without describing all of them.
To describe a set of classical bits completely, you need a probability distribution over the whole system (including possible classical correlations induced by the crosstalk that OP spoke about). That probability distribution is a stochastic vector that has 2^n real positive components if we have n bits.
To describe a set of qubits completely, you need at least a "quantum probability distribution", i.e. a ket that has 2^n complex components (I am skipping the discussion of density matrices).
Both the classical and quantum description of the bits requires exponentially many real numbers. This exponential behavior on its own is not enough to the explain "quantum computational advantage". The exponential behavior is well known in plenty of classical contexts (e.g. under the name "curse of dimensionality") and classically solved with Monte Carlo modeling.
Scott Aaronson's lecture notes cover that quite well.
At the end, the issue of crosstalk is modeled in a very similar way in the quantum and classical case, and if it forbids the existence of a quantum computer, it should forbid the existence of classical ones as well. In both cases it is the use of linear binary (or stabilizer) codes that solves that issue.
If I have understood your argument correctly I don’t think the conclusion follows from the premise because crosstalk is not fundamental to classical computing. By that I mean that for a given logical cpu design, crosstalk can be (and is) minimized through physical design characteristics (eg grounding, wire spacing etc) without reducing the size of computation. The same cannot afaik be said of quantum computers, where the interaction between qbits is essential to computation
Of course, qubits are drastically more difficult, but the principle is the same.
Basically what I mean is that classical computing can split things up during storage or routing to reduce coupling, but quantum can’t do this because the qbits must stay entangled to be useful
And the way we mathematically express and model classical correlations in classical error correcting codes is virtually the same as the way to model quantum correlations (entanglement) in quantum error correcting codes.
All of this with the usual disclaimer: the quantum case is much more difficult, engineeringly speaking.
For example in PCIe each lane can only cary one bit at a time in each direction. In typical consumer equipment, there are at most 16 lanes of PCIe (eg a graphics card socket) meaning there can only be at most 16 bits (2 bytes) on the wire at any given time, but the bits are sent at a very high frequency allowing for high transfer rates. This only works because taking those 256 bytes and sending them one by one (or 16 by 16) over the wire doesn't lose information.
- Both classical probabilistic bits and qubits need exponential amount of memory to write down their full state (e.g. stochastic vectors or kets). The exponential growth, on its own, is not enough to explain the conjectured additional computational power of qubits. This is discussed quite well in the Aaronson lecture notes.
- Entanglement does not have much to do with things being kept physically in contact (or proximity), just like correlation between classical bits has little to do with bits being kept in contact.
- Nothing stops you from sending/storing entangled bits one by one, completely separate from each other. If anything, the vast majority of interesting uses of entanglement very much depend on doing interesting things to spatially separate, uncoupled, disconnected, remote qubits. Sending 1000 qubits from point A to point B does not require a bus of width 1000, you can send each qubit completely separately from the rest, no matter whether they are entangled or not.
- Not even error correction requires you to work with "big" sets of qubits at the same time. In error correcting codes, the qubits are all entangled, but you still work with qubits one by one (e.g. see Shor's syndrome measurement protocol).
- I strongly believe your first sentence is too vague and/or wrong: "What makes qbits valuable is that they scale superlinearly the more of them you have entangled together". As I mentioned, the exponential growth of the "state descriptor" is there for the classical probabilistic computers, which are believed to be no more powerful than classical deterministic computers (see e.g. derandomization and expander graphs). Moreover, Holevo's theorem does basically say that you can not extract more than n bits of information from n qubits.
- Another quote: "Since they have to remain entangled to do valuable work, you can't physically separate them to prevent interference" -- yes, you can and very much do separate them in the vast majority of interesting applications of entanglement.