It was only GPU use. It’s come down a little to around 70-72 now (27W, close to 100% GPU use, close to 0% CPU use). This seems much better, though I don’t know if it’ll stay here, or there’s just some variance in it and it’s low now.
I’ll leave it at 2100 for now and keep an eye on it. If I can get the field diagnostics installed (they fail - see my comment in DGX Spark GB10: partnerdiag PowerStress reproducibly hard-powers-off the box — acpitz 88→97.8°C in 5s, all other field tests pass - #7 by DannyTup) I will try running that, although I’m not sure if it does anything that would detect any issue I have here.