Wow, the 'waterfall economy' concept you describe is just brilliant and so insightful! It realy clarifies the apparent contradiction with depreciation. I'm curious tho, from a practical standpoint, what does the software orchestration look like for efficiently repurposing these high-end GPUs as they get demoted? Must be a wild puzzle to solve.
Yessss.... It is super hard. Fantastic question BTW. Thank you. So.. I may write a short note to update. It's as challenging as it is fairly questionable to ask - is it possible not every datacenter will have the ability to repurpose? Hyperscalers better equipped.
Good analysis and you bring excellent points, in particular to accounting involving judgement calls based on company-specific patterns.
The question that I am asking, however, is whether companies have sufficient "using engineering data and internal analysis" to support their conclusions. I might be wrong, but it sounds to me that, given the short history of GPU-based workloads, the data might be limited. In this case, the assumptions might change as more data becomes available., leading to shorter depreciable lives than the current 6-years.
I believe it's likely with the hyperscalers, Olga. More so Microsoft. Google. Amazon. But we have no visibility on the private companies (too soon) - Big AI labs etc, who seem to be vertically integrating w hardware now. Yes limited on data and disclosures. Your concerns are fair.. Good analysts covering public companies will be asking these questions at their briefings. Investors of the private AI labs should be doing the same. To risk mitigate
Wow, the 'waterfall economy' concept you describe is just brilliant and so insightful! It realy clarifies the apparent contradiction with depreciation. I'm curious tho, from a practical standpoint, what does the software orchestration look like for efficiently repurposing these high-end GPUs as they get demoted? Must be a wild puzzle to solve.
Yessss.... It is super hard. Fantastic question BTW. Thank you. So.. I may write a short note to update. It's as challenging as it is fairly questionable to ask - is it possible not every datacenter will have the ability to repurpose? Hyperscalers better equipped.
Please see the following paper on energy efficiency in new versus old technologies. This was briefly pointed out in the article.
https://arxiv.org/abs/2310.07516
Good analysis and you bring excellent points, in particular to accounting involving judgement calls based on company-specific patterns.
The question that I am asking, however, is whether companies have sufficient "using engineering data and internal analysis" to support their conclusions. I might be wrong, but it sounds to me that, given the short history of GPU-based workloads, the data might be limited. In this case, the assumptions might change as more data becomes available., leading to shorter depreciable lives than the current 6-years.
I believe it's likely with the hyperscalers, Olga. More so Microsoft. Google. Amazon. But we have no visibility on the private companies (too soon) - Big AI labs etc, who seem to be vertically integrating w hardware now. Yes limited on data and disclosures. Your concerns are fair.. Good analysts covering public companies will be asking these questions at their briefings. Investors of the private AI labs should be doing the same. To risk mitigate