Whether P = NP is a question about the asymptotic growth of computational costs of algorithms as functions of input sizes. It may provide hints about concrete computational cost estimates of algorithms for specific input sizes, but doesn't provide answers.
In the eyes of the asymptotic setting, an $O(n^{10000})$ cost is ‘smaller’ than an $O(1.0001^n)$ cost because for sufficiently large values of $n$, past some boundary $n_0$, all values of $1.0001^n$ exceed the corresponding values of $n^{10000}$.
Does that mean an instance of a cryptosystem that costs $128\;\text J$ energy to compute and costs $128^{10000}\;\text J \approx 10^{20000}\;\text J$ to break is less secure than a cryptosystem that costs $128\;\text J$ to compute and costs $1.0001^{128}\;\text J \approx 1.01\;\text J$ to break? Certainly not!
For example, finding $k \in \{0,1\}^{128}$ given $\operatorname{AES}_k(0)$ takes $O(1)$ time because the input size $n$ is a constant, but that fact implies nothing about its feasibility in practice. The asymptotic growth of the NFS and ECM costs guide us toward parameter growth curves for, e.g., RSA, but they require concrete analysis to attain confidence in concrete choice of parameters.
For a particularly goofy example of this, see
Daniel J. Bernstein, Nadia Heninger, Paul Lou, and Luke Valenta, ‘Post-quantum RSA’, PQCrypto 2017, Springer LNCS 10346, 2017, pp. 311–329,
which shows how to take advantage of a merely quadratic difference between a user's asymptotic costs and an attacker's asymptotic costs to make a cryptosystem, with a recommendation for specific input sizes guided by concrete use and attack cost estimates, that is just barely usable by rich users but likely out of reach of attackers even if they had access to a large quantum computer.
So would it matter if there turned out to be an asymptotically polynomial-time algorithm to solve every NP-complete problem in $O(f(n)^{10000})$ time when testing a solution takes $O(f(n))$ time for polynomial $f$? No, not really.
Could it matter, if there were a really amazing trick to use a low-degree polynomial that is, in concrete terms, cheap to compute? Maybe, but at this point that seems implausible—and cryptography might survive nevertheless by high-degree polynomial asymptotic gaps and exorbitant concrete gaps between user and attacker costs even if they're not exponential.
