For clarification, I'm not implying that GPU-backed loans are the thing that causes 2008 to happen again just because (a) of the title and (b) me using a picture from the Big Short in the preview - that'd be a reach. What I am doing is using math to try to help explain the rise of GPU securitisation, and the risks it may present *if* the assumptions do not hold. As of now, assuming recovery ends up 60% or above then using GPUs as collateral is fine, especially if we're talking about senior secured loans (which most of them are).
If reader(s) want to believe this is 2008, then that's on them. I'm not writing this to influence people's views.
Can I ask where you got the Beignet Investor docs from? They’ve been hard to track down and I’ve only gotten bits and pieces from “a friend” at one of the institutional investors.
I’ve wondered a lot of the same questions you’re asking. That debt doesn’t look much like an A+ to me, but I’m not a rater.
Also, I get that land is cheap in Louisiana, but who thought it was a good idea to build a data center there?
Getting access to these docs is obviously hard. It took a lot of researching to put together. The next best alternative I could find was these pieces below.
I'm not a rater either but this is what S&P Global thinks it is. That rating may or may not be generous depending on where you stand. Apparently Meta thinks it's a good idea to build a data centre there - if I were to guess, it'd be for cheap land and access to water and electricity.
This gets wonky when you begin to think that GPUs might not depreciate in the time scale you’re proposing (e.g cost of renting h100 going from dollar ish to ~2.50 two years later and the Google cloud guy saying that they don’t decommission TPUs). If the constraint is flops for AGI then they might just keep appreciating until we hit AGI and get limitless everything.
That's exactly why I mentioned the math is stylised because a scenario could happen that the assumptions in my math (based on AltStreet's numbers) didn't account for.
Apparently so, I didn't know this either until I read AltStreet's reference sheet and tried piecing it together. Without AltStreet's reference sheet, I wouldn't have made this post.
This is by far the most challenging piece I've published due to me modelling various scenarios where this does or doesn't hold, and then writing my thoughts based on the math.
For clarification, I'm not implying that GPU-backed loans are the thing that causes 2008 to happen again just because (a) of the title and (b) me using a picture from the Big Short in the preview - that'd be a reach. What I am doing is using math to try to help explain the rise of GPU securitisation, and the risks it may present *if* the assumptions do not hold. As of now, assuming recovery ends up 60% or above then using GPUs as collateral is fine, especially if we're talking about senior secured loans (which most of them are).
If reader(s) want to believe this is 2008, then that's on them. I'm not writing this to influence people's views.
The structure looks solid.
The assumption underneath it matters more.
GPUs aren’t stable collateral.
They’re on a replacement cycle.
That’s the risk.
Can I ask where you got the Beignet Investor docs from? They’ve been hard to track down and I’ve only gotten bits and pieces from “a friend” at one of the institutional investors.
I’ve wondered a lot of the same questions you’re asking. That debt doesn’t look much like an A+ to me, but I’m not a rater.
Also, I get that land is cheap in Louisiana, but who thought it was a good idea to build a data center there?
Getting access to these docs is obviously hard. It took a lot of researching to put together. The next best alternative I could find was these pieces below.
1) https://www.spglobal.com/ratings/en/regulatory/article/-/view/sourceId/101651795; https://www.spglobal.com/ratings/en/regulatory/article/-/view/sourceId/101652218
2) https://www.ft.com/content/d0344253-b0a2-4c6d-8b97-520243678afd
3) https://www.ifre.com/ifr-awards/2327933/financing-package-blue-owl-capitalbeignet-investors-us27.3bn-23.6-year-bond
I'm not a rater either but this is what S&P Global thinks it is. That rating may or may not be generous depending on where you stand. Apparently Meta thinks it's a good idea to build a data centre there - if I were to guess, it'd be for cheap land and access to water and electricity.
This gets wonky when you begin to think that GPUs might not depreciate in the time scale you’re proposing (e.g cost of renting h100 going from dollar ish to ~2.50 two years later and the Google cloud guy saying that they don’t decommission TPUs). If the constraint is flops for AGI then they might just keep appreciating until we hit AGI and get limitless everything.
That's exactly why I mentioned the math is stylised because a scenario could happen that the assumptions in my math (based on AltStreet's numbers) didn't account for.
I see, great work keep it up
huh, I didn't know this! You extend 50-70% of OLV and then you securitize it... that's reaaaaaly clever!
Apparently so, I didn't know this either until I read AltStreet's reference sheet and tried piecing it together. Without AltStreet's reference sheet, I wouldn't have made this post.
This is by far the most challenging piece I've published due to me modelling various scenarios where this does or doesn't hold, and then writing my thoughts based on the math.
Thanks for going through that brain damage for our benefit!