OneChronos teamed up with prize-winning economist Paul Milgrom of Auctionomics to develop the world's first financial market for GPU compute. Now comes the hard part.
There are differences in interconnects, memory, and type, but GPU compute can be made fungible.
I think the biggest issue is that GPU compute prices tend to trend downward. Oil hedges work since price is so unpredictable in the short term. It can go up or down, so different counter parties "bet" on the direction by providing the hedging instrument. If the price trends downwards continuously, options prices will be way too expensive or the market for it will become illiquid.
Securitization makes sense if you want to create derviative products. If effective supply distribution is an issue, routers like OpenRouter and Martian should be enough.
Can you explain this further for a lay-person? Is the idea that having many sellers and buyers of a financial asset leads to better prices for everyone? But given that the general trend for a type of GPU is downward over time, there won't be enough buyers and sellers to make meaningful reduction in GPU costs?
Having more sellers and buyers COMBINED with an efficient market-clearing mechanism would lead to a more accurate price of a particular good, though not necessarily "better" prices since "better" may be defined differently if you are a seller vs. a buyer.
What I'm trying to point out is that the securitization of goods (in this case, GPU compute) would ultimately be done as a way to transfer risk and provide hedging.
If I am growing wheat that will be harvested in 6 months, I may not want to risk the price of wheat going down from its current trading price of $530/bushel. So I may leverage the futures market to "pre-sell" my wheat as a hedge and lock in my price. I am relieved when prices go down to $500, knowing that I made $30 extra. I am a bit sad if it goes up to $550.
But if the general expectation is that wheat prices will go down in the future, why would anyone buy my wheat at $530 when they can wait and buy it for $500 in the future? The party on the other side would only buy wheat at $530 for future delivery if they think that wheat prices will go up. If there's no chance of wheat price going up, no one would pay the higher price.
This is a really good point and limitation for this model. But even if the price isn't increasing in the future, there could be value in more efficiently securing access in a way that can't be bumped, or setting conditional contracts to buy supply from the better price between several sellers within a certain time range. Milgrom et al also note that a financial market for GPU compute could allow for companies to be both buyers and sellers at the same time more easily, and to sell future capacity more easily to finance infastructure projects.
I agree. There is definitely an opportunity to make money, especially if we assume short-term price fluctuations that come from an efficient market for GPU compute.
When you consider counterparty depth across 2-player, 3-player, and n-player markets, the dynamic could change significantly.
I think your last point on financing was what got me the most interested in exploring this space. CoreWeave was able to get comparatively lower cost of capital vs. smaller neoclouds precisely because of their multiyear compute contract that guaranteed a certain price for their GPU compute. Allowing compute price hedging would've allowed others to achieve the same thing without this contract.
I had written about the securitization potential of GPU compute last year around this time (https://trees.substack.com/p/thoughts-on-coreweave).
There are differences in interconnects, memory, and type, but GPU compute can be made fungible.
I think the biggest issue is that GPU compute prices tend to trend downward. Oil hedges work since price is so unpredictable in the short term. It can go up or down, so different counter parties "bet" on the direction by providing the hedging instrument. If the price trends downwards continuously, options prices will be way too expensive or the market for it will become illiquid.
Securitization makes sense if you want to create derviative products. If effective supply distribution is an issue, routers like OpenRouter and Martian should be enough.
Can you explain this further for a lay-person? Is the idea that having many sellers and buyers of a financial asset leads to better prices for everyone? But given that the general trend for a type of GPU is downward over time, there won't be enough buyers and sellers to make meaningful reduction in GPU costs?
Having more sellers and buyers COMBINED with an efficient market-clearing mechanism would lead to a more accurate price of a particular good, though not necessarily "better" prices since "better" may be defined differently if you are a seller vs. a buyer.
What I'm trying to point out is that the securitization of goods (in this case, GPU compute) would ultimately be done as a way to transfer risk and provide hedging.
If I am growing wheat that will be harvested in 6 months, I may not want to risk the price of wheat going down from its current trading price of $530/bushel. So I may leverage the futures market to "pre-sell" my wheat as a hedge and lock in my price. I am relieved when prices go down to $500, knowing that I made $30 extra. I am a bit sad if it goes up to $550.
But if the general expectation is that wheat prices will go down in the future, why would anyone buy my wheat at $530 when they can wait and buy it for $500 in the future? The party on the other side would only buy wheat at $530 for future delivery if they think that wheat prices will go up. If there's no chance of wheat price going up, no one would pay the higher price.
This is a really good point and limitation for this model. But even if the price isn't increasing in the future, there could be value in more efficiently securing access in a way that can't be bumped, or setting conditional contracts to buy supply from the better price between several sellers within a certain time range. Milgrom et al also note that a financial market for GPU compute could allow for companies to be both buyers and sellers at the same time more easily, and to sell future capacity more easily to finance infastructure projects.
I agree. There is definitely an opportunity to make money, especially if we assume short-term price fluctuations that come from an efficient market for GPU compute.
When you consider counterparty depth across 2-player, 3-player, and n-player markets, the dynamic could change significantly.
I think your last point on financing was what got me the most interested in exploring this space. CoreWeave was able to get comparatively lower cost of capital vs. smaller neoclouds precisely because of their multiyear compute contract that guaranteed a certain price for their GPU compute. Allowing compute price hedging would've allowed others to achieve the same thing without this contract.