When do we get a real hardware paradigm shift instead of just scaling the GPU/RAM triad?

Right now, our entire path to AGI relies on throwing gigawatts at a 1945 Von Neumann bottleneck. Almost 80% of data center energy isn't even spent on compute, it's wasted just moving bytes between HBM/DRAM/SRAM and the cores. Meanwhile, we force naturally analog, probabilistic physics into strict 0s and 1s just to simulate neural weights in software. ​Incremental node shrinks and software optimizations are hitting a physical wall. What is the "discovery of electricity" level breakthrough that finally ditches the RAM+GPU triad for native AI hardware? Neuromorphic, photonics, or thermodynamic compute? What do you think?