Assuming you want to define the goal, "AGI", as something functionally equivalent to part (or all) of the human brain, there are two broad approaches to implement that.
1) Try to build a neuron-level brain simulator - something that is a far distant possibility, not because of compute, but because we don't have a clear enough idea of how the brain is wired, how neurons work, and what level of fidelity is needed to capture all the aspects of neuron dynamics that are functionally relevant rather than just part of a wetware realization
OR
2) Analyze what the brain is doing, to extent possible given our current incomplete knowledge, and/or reduce the definition of "AGI" to a functional level, then design a functional architecture/implementation, rather than neuron level one, to implement it
The compute demands of these two approaches are massively different. It's like the difference between an electronic circuit simulator that works at gate level vs one that works at functional level.
For time being we have no choice other than following the functional approach, since we just don't know enough to build an accurate brain simulator even if that was for some reason to be seen as the preferred approach.
The power efficiency of a brain vs a gigawatt systolic array is certainly dramatic, and it would be great for the planet to close that gap, but it seems we first need to build a working "AGI" or artificial brain (however you want it define the goal) before we optimize it. Research and iteration requires a flexible platform like GPUs. Maybe when we figure it out we can use more of a dataflow brain-like approach to reduce power usage.
OTOH, look at the difference between a single user MOE LLM, and one running in a datacenter simultaneously processing multiple inputs. In the single-user case we conceptualize the MOE as saving FLOPs/power by only having one "expert" active at a time, but in the multi-user case all experts are active all the time handling tokens from different users. The potential of a dataflow approach to save power may be similar, with all parts of the model active at the same time when handling a datacenter load, so a custom hardware realization may not be needed/relevant for power efficiency.