The strongest argument for possibility is that general intelligence already exists: human brains are physical systems, so the property is not supernatural. Unless cognition depends on something non-physical, it can in principle be reproduced in another substrate. The open questions are about engineering, scale and time, not about whether the target is coherent.
The case against
Sceptics make three arguments. First, that intelligence is inseparable from a body and a social world, so disembodied computation will always be missing something. Second, that current methods, which interpolate over data, cannot produce genuine understanding no matter how much compute is added. Third, that the economic and energy constraints will bind before the technical ones are solved.
The case for
Brains are proof of concept, and they run on about twenty watts. Nothing in the physics says twenty watts is the floor for cognition. The argument is about whether we can find the right architecture, not whether one can exist.
What would settle it
A system that acquires a genuinely new skill after deployment, in an unfamiliar domain, without a retraining run, and can explain how it did it. That single demonstration would move the debate further than any benchmark score.