For what its worth, Anthropic posted this in their corporate blog. So if its a joke, its coming out of vetted corporate PR.

Given its current state, this is a singularly bad idea.

By customer is easy: they’re each renting specific resources. A fractional cloud instance (excepting the sma burst able ones) is tied to specific CPUs and GPUs. And there are records of who rented which one when being kept already.

You might not be able to break out specific individual queries, but computing averages is completely straightforward

You can produce a remarkably good estimate by looking at CPU and GPU utilization out of procfs and profiling a handful of similar machines power use with similar utilization and workloads.

Network is less than 5% of power use for non-GPU loads; probably less for GPU.

Data center demand has created huge backlog of gas turbine orders. They’re not planning on renewables for the next big expansion

They know exactly what the power consumption of that hardware is though. This isnt tough to figure out just because you use a cloud provider

I’m guessing that there are security guards.

They also burn methane gas, not gasoline, which means that sugar probably isn’t an effective sabotage technique.