I don't assume computing power is constant, just that the rate of content generation is more or less constant (I think a billion devices publishing a billion new pieces of content every second is a pretty reasonable upper bound). OP asked about the odds of collisions occurring by accident due to the sheer volume of content generated and published, not about attacker scenarios.

Obviously we wouldn't use the same hash algorithm and setup for 12 million years, but the sheer absurdity of that length of time and that pace of content production shows this method will last, at least until flaws are found in SHA256.