Hardware

Another video noted that previous generations used multiple isolated shunt resistors to feed sets of VRM phases. The current measurements in these shunts was used to balance the current in the phases. If any shunt showed no current (connector or pins unplugged) the GPU wouldnt even turn on. Previous generations would do ridiculous things to ensure proper current balance in the wires.

This is done with 6 12V wires on the 30 series. Pairs of wires go to a shunt so there are 3 shunts, which each feed maybe 4 phases. In the worst case scenario, you could somehow lose half the wires, 1 from each pair, and the GPU wouldnt notice. This would result in up to 2x overload on any individual wire. This is why the 30 series did not burn like the 40 and 50 series.

The 4090/5090 short all of the pins together before going through 1/2 shunts, then are shorted again after the shunts. This means no individual or pair of pin currents are measured, and you can cut 5/6 of the wires and have it still think theres a good connection. This results in up to a 6x overload on a wire (e.g. all 600W going through 1 wire).

Also sub in cut wire for bad connection on a pin.

This is a design issue with the board at minimum.