Computable GPU Index (CGI) is a USD price per GPU-hour, computed from the published on-demand rental rates of a fixed panel of providers. The methodology is mathematically robust, and anyone can verify and reproduce every print.
Really interesting direction. What do you see as the first real use case for CGI — cloud procurement, compute financing, or eventually something closer to a tradable benchmark?
@tizaolai886 eventually it will be a bit of everything you mentioned. the first use case is now cash settled derivatives, basically hedges on gpu prices, using CGI as a reference rate.
Using a weighted median instead of a simple average seems like an important design choice. I’d be interested to understand how the weights are calculated and how you prevent one large provider—or several unusually low listings—from having too much influence on the final number.
@matthewwei so, we actually are using a interquantile mean instead of median. this means, any outlier would not have an impact on the price. the weights take in consideration of how lively a data source is. you can read more about it in our methodology!
congrats on the launch, this is a good idea - GPU pricing has needed a neutral reference rate for a while. question: since the index is built from on-demand rates only, is there any plan to also track committed-use/reserved pricing? for a lot of teams the real decision isn't "what's the on-demand rate" but "is the reserved discount worth the commitment", and that gap seems like it'd need a second number entirely rather than fitting into the same index.
@galdayan you are right, and it'd be a second index. On-demand is the most indexable tier because it's one comparable product, publicly priced at every provider, all the time. Reserved often has some private negotiation component over tenor and volume, so there's no published data to index.
Report
Finally, a GPU price index that doesn't make me feel like I'm deciphering ancient runes. Already sent this to my team, we've been manually tracking cloud costs in a spreadsheet like cavemen.
Any chance you'll add Lambda Labs or Vast.ai to the provider list?
Congrats on shipping this @raysong The interquantile mean approach is smart, trims the outlier providers without throwing away real signal like a plain median would.
Just curious if spot pricing is on the roadmap, or is CGI staying on demand only for now? Following this one for sure
Dune
Congratulations on the launch!!
Good luck! @raysong @oliviazhao
Computable GPU Index (CGI)
@dhanrajchoudhary thank you!
Nice to see messy market data presented in a way that is easier to understand.
Computable GPU Index (CGI)
@luvian_yu thanks!
Clipto
Really interesting direction. What do you see as the first real use case for CGI — cloud procurement, compute financing, or eventually something closer to a tradable benchmark?
Computable GPU Index (CGI)
@tizaolai886 eventually it will be a bit of everything you mentioned. the first use case is now cash settled derivatives, basically hedges on gpu prices, using CGI as a reference rate.
Clipto
@raysong cool.. combining with finance is such a bold idea :)
Clipto
Using a weighted median instead of a simple average seems like an important design choice. I’d be interested to understand how the weights are calculated and how you prevent one large provider—or several unusually low listings—from having too much influence on the final number.
Computable GPU Index (CGI)
@matthewwei so, we actually are using a interquantile mean instead of median. this means, any outlier would not have an impact on the price. the weights take in consideration of how lively a data source is. you can read more about it in our methodology!
Dial
congrats on the launch, this is a good idea - GPU pricing has needed a neutral reference rate for a while. question: since the index is built from on-demand rates only, is there any plan to also track committed-use/reserved pricing? for a lot of teams the real decision isn't "what's the on-demand rate" but "is the reserved discount worth the commitment", and that gap seems like it'd need a second number entirely rather than fitting into the same index.
Computable GPU Index (CGI)
@galdayan you are right, and it'd be a second index. On-demand is the most indexable tier because it's one comparable product, publicly priced at every provider, all the time. Reserved often has some private negotiation component over tenor and volume, so there's no published data to index.
Finally, a GPU price index that doesn't make me feel like I'm deciphering ancient runes.
Already sent this to my team, we've been manually tracking cloud costs in a spreadsheet like cavemen.
Any chance you'll add Lambda Labs or Vast.ai to the provider list?
Computable GPU Index (CGI)
@boyuan_deng1 they're already included!
Triforce Todos
Congrats on shipping this @raysong
The interquantile mean approach is smart, trims the outlier providers without throwing away real signal like a plain median would.
Just curious if spot pricing is on the roadmap, or is CGI staying on demand only for now? Following this one for sure
Computable GPU Index (CGI)
@abod_rehman it's staying on demand right now