This should resolve yes, Llama-3.1-405B is open sourced and confirmed to have trained on 16K H100s for over 6 weeks according to their paper.
Just in GPU running costs alone that is over $15M at even the very conservative estimate of $1 per hour rental cost for an H100. And 6 weeks is 1008 hours.

So 16K GPUs, multiplied by over 1008 hours, multiplied by a very conservative estimate of $1 per hour per H100. Equals $16M USD

(note: these are conservative estimates, the average market rate per hour for an H100 is actually closer to $1.5-$2 per hour, and this is also not taking into account all the potential R&D costs that led upto the training of the model, and also doesn't include the capex of the GPUs themselves.)