There are two things I find suspect with the thermocouple and 125 °C ideas. The first is that thermocouples are typically used to probe larger objects and coupling is always an issue with smaller objects. In Gamers Nexus's final tests, the thermal couple was coupled to the back of the PCB. This can only measure the stead-state temperature of the back side. To extrapolate that number to actual VRM temperatures, one would have to assume a steady-state thermal gradient and do a few multiplications. One way to obtain that, would be the use of a thermal imaging device.
Which brings us to our second issue, these VRM packages are like any other silicon chip - the core is much hotter than the plastic packaging since the heat is generated in the silicon parts underneath. The manufacturer of the VRM chips usually assumes a given amount of cooling and hence an expected thermal gradient range, which is why the VRM as a whole package is specified for Tcase temperatures, when really, the MOSFETs are the main sources of heat. It is up to the GPU maker to implement such a thermal gradient. Given this thermal gradient, it is not clear what the core temperature of these VRM chips are. This is further complicated due to the power MOSFETs in the VRMs being typically specified for a steady-state thermal impedance and a transient thermal impedance (and in general the whole package features this distinction as well, due to thermal capacitance of any finite-sized object). A small burst workload (<1 second) close enough to the maximal internal channel temperatures at already high duty may trigger a transient over-temperature condition (in the silicon parts and not measurable as steady-state increases of temperature outside the package) and cause eventual failure through aggregated damage if not immediate.
For the longevity and reliability of these VRM chips, I would think a good design would be to keep the MOSFETs several 10s of degrees below the 125°C maximum to ensure actual MOSFET channel temperatures are always below the maximal channel temperatures specified. To do so, you would need to ensure the thermal gradient across the VRM chips to be as low as possible, which means effective cooling coupling in the form of a heatsink (that lowers the thermal impedance of the VRMs to the surrounding air).
