Also, the Reuters article says that they have a depreciation schedule built into the contact. Compute hardware can depreciate pretty quickly, if I remember correctly major tech companies have around 3 years planned for the depreciation.
Historically yea, though now I think it's being stretched to 5+ years as they see hardware last longer in production (but you're atiop correct, it's very quickly relative to a lot of other loan collateral, still a short term deal even if it's 7 years)
Personally I think the bigger risk is software innovations making CPU training (and/or inference) sufficiently viable that it's cheaper to train models on a commodity CPU cluster than on some proportionally expensive GPU cluster. I don't know enough about the space to say whether that's likely, but it seems like a low risk, since pretty much any parallel algorithm will always be faster on GPU than CPU - it's just a question of the marginal benefits and cost (e.g. maybe it takes more CPU to train same model in same time, but cost of CPU is so much lower that it's worth buying more of them).
There's also the fact that if you are not training LLM, you can get a better deal using some older hardware.
But some value as long as computing power can be sold above running costs(power, cooling, etc.). Good question is when the newer model is so much more efficient it makes sense to replace them.