Fundamentally, a limit on doing it at scale is that it for efficiency it requires the heat to be consumed near the production - and the bulk of large power intense data centres are not located in the midst of high density residential neighbourhoods with a demand for heat.
If you're going to be doing cogeneration on-site, it can be worth considering a Combined Heating, Cooling, and Power system - but the benefits tend to be fairly marginal.
In the future, there's a possibility of extending the "lights out datacentre" concept, and going fully automated. If you don't need to accommodate humans, you can run much higher temperature gradients.
With a gradient of around 200 K, for instance, you might expect to recover somewhere between 30-35% of your total energy input.
The UK actually ought to be a good choice for an experiment along those lines, given its long history with gas-cooled nuclear systems - sadly, the engineers involved have mostly all retired by now...
The only way you could make this viable would be to change silicon processes to those capable of running at significantly over 100C. This incurs a big efficiency penalty, but then you can start boiling water directly off the die and letting the steam move itself, perhaps to some sort of Stirling engine condenser cycle, with a lot less pumping losses.