It's not that it's a local maxima. There really isn't any economic incentive to care about investing in technologies that can't be turned into a profit in a few years. Capitalism is doing a fine job optimizing, it's just optimizing the wrong function.
>humans are bad at predicting the future, so attempting to allocate resources perfectly is probably a waste of resources.
It's hard to do worse than "random" or "not at all", which is basically what our current system is. Sure it brought us GPUs. But that's pure coincidence that what video games require happened to be what AI requires. If we weren't so lucky that two completely unrelated industries happened to require the same technology, we would never get AI.
Imagine tomorrow someone make a superior graphics chip. That's somehow highly specialized to 3d rendering in video games and does nothing else. Dropping all the general purpose computation and linear algebra stuff. Gamers would all switch to it and the market for GPGPUs would die and AI would stop advancing, or even reverse.
And GPUs aren't that weird or surprising of an invention to invest in. Increased computing power benefits many scientific fields besides AI and would be an obvious investment for a hypothetical central planner.