The data centres aren't inherently spikey, in general use their consumption is reasonably predictable.

However, if a DC detects that the _grid_ is wobbly (voltage or frequency deviations) the DC will disconnect without warning, and switch to its batteries and generators.

The grid complains because it's suddenly lost hundreds of MW of load. For the DC to have isolation capability, it would need a load-sink which can consume roughly the same power as the DC in normal operation, and can take in that load at a moment's notice.

It's a hard problem to solve, and probably better managed at grid-level than DC-level.

My intuition is that there would be a fairly stable base load, but doing something like switching on a new training run of a frontier model would be incredibly spiky, thousands of GPUs going from somewhat idle to 100% in seconds.

>it would need a load-sink which can consume roughly the same power as the DC in normal operation, and can take in that load at a moment's notice.

Liquid rheostat. A big one.