I think one of the things that people analysing the finances of the AI sector often miss is that the capital expenses associated with data centre buildout involve spending a lot of money on compute hardware which depreciates and becomes obsolete very fast.

Large companies have often changed the depreciation schedule for server hardware to 6 years rather than 3. I think it remains to be seen if this is justified as it’s not as though the pace of change in hardware has slowed down by 50% arbitrarily.

I think it’s pretty obvious to me that a data centre full of ampere GPUs is worth less than a data centre full of Blackwell GPUs. I suspect this is also why AI companies prefer to talk in terms of gigawatts of compute, as if efficiency per watt does not exist.

Data centre costs are not one-off. Some are, for buying land and building buildings and infrastructure. The actual compute however is a constant investment. I’m not familiar enough with the market to know what proportion of expenses are ongoing but I bet it’s higher than people think.