From compute tools to collateral: the rapid financialization of GPUs is creating a new asset class, but the infrastructure to trade, value, and hedge them is still being built
Hey, great read as always. That line about 'GPUs are evolving from mere compute tools into strategically significant financing components' really hit different. It perfectly captures the wild speed of AI development and how traditional finance is trying to catch up. So insightful!
In the longer term, I expect some (many?) of cloud providers to be impaled by this logic: financing with long-life assumptions on a short-life asset. But I think that disaster takes a long time to develop...perhaps +3 years or more from now. The reason is severely constrained supply vs demand for the foreseeable future.
Sure Rubin is gonna make today's Blackwell chips obsolete... but that word has different meanings:
- If we say "obsolete" means "outpaced by substantially better newer products," then Blackwell will be obsolete within 2026.
- If we say "obsolete" means "no longer economically very useful" then Blackwell will enjoy a long lifetime in the sun before obsolescence arrives; maybe til 2032 or longer.
For evidence we can look back to Ampere-class chips, which (per Nvidia) are still popular and highly utilized. Those came out in 2020. It's hard to look at a 6-year life assumption as wrong, for a thing still so highly utilized 6 years later. Do we really think all those get switched off and tossed in the landfill within the 12 months? or the next 12 after that?
Someday this thing saturates, the music stops, and some cloud lenders will get burned, badly. But I'm not betting on that in 2026. Or in 2027.
Sometimes you just have to take the risk. We don’t know how the software or the applications running on these GPUs will evolve over time. Edge computing may have different needs. Individual PCs running personalized LLM in the cloud may have different compute needs. Deepseek demonstrated capabilities to build perfectly good working models on lower performance GPUs. Amazon is taking a highly conservative approach of writing down the value to zero in 5 years. I used to work for a company that wrote down the desktop PCs in 5 years but kept using those for 6-7 years. My point is useful life is very nebulous term, especially when it comes to brand new technology.
Hey, great read as always. That line about 'GPUs are evolving from mere compute tools into strategically significant financing components' really hit different. It perfectly captures the wild speed of AI development and how traditional finance is trying to catch up. So insightful!
In the longer term, I expect some (many?) of cloud providers to be impaled by this logic: financing with long-life assumptions on a short-life asset. But I think that disaster takes a long time to develop...perhaps +3 years or more from now. The reason is severely constrained supply vs demand for the foreseeable future.
Sure Rubin is gonna make today's Blackwell chips obsolete... but that word has different meanings:
- If we say "obsolete" means "outpaced by substantially better newer products," then Blackwell will be obsolete within 2026.
- If we say "obsolete" means "no longer economically very useful" then Blackwell will enjoy a long lifetime in the sun before obsolescence arrives; maybe til 2032 or longer.
For evidence we can look back to Ampere-class chips, which (per Nvidia) are still popular and highly utilized. Those came out in 2020. It's hard to look at a 6-year life assumption as wrong, for a thing still so highly utilized 6 years later. Do we really think all those get switched off and tossed in the landfill within the 12 months? or the next 12 after that?
Someday this thing saturates, the music stops, and some cloud lenders will get burned, badly. But I'm not betting on that in 2026. Or in 2027.
Sometimes you just have to take the risk. We don’t know how the software or the applications running on these GPUs will evolve over time. Edge computing may have different needs. Individual PCs running personalized LLM in the cloud may have different compute needs. Deepseek demonstrated capabilities to build perfectly good working models on lower performance GPUs. Amazon is taking a highly conservative approach of writing down the value to zero in 5 years. I used to work for a company that wrote down the desktop PCs in 5 years but kept using those for 6-7 years. My point is useful life is very nebulous term, especially when it comes to brand new technology.