GPU futures need to account for performance differences

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 28:21

   Does this become more obvious if compute does become more of a futures market or a commodity like the sort of, you know, the electricity or corn or something like that? So I actually think that futures markets are a good example of the simplistic thing doesn't work. I'll give you an example. Let's say we're doing futures for wheat. What happens in the futures for wheat is there is some world and maybe you can actually get wheat in the future, like you actually get delivered that wheat.

2. 28:52

   The problem that doesn't really work for compute is that compute is not very fungible. It's kind of fungible. You can make it fungible, but that fungibility depends on a massive number of different types of variables that are not factored in when people mostly think about the compute price as just a single number. One example is that depending on how your VM implementation works, you can have a 40% performance difference on GPU kernel launches from one operator to another.

3. 29:26

   Another example is if you had the early H100s, they were shit. And if you had the later H100s, they were great because there were like problems in the initial ones and they got fixed over time. If you're talking about just like an H100 price, that's not the same thing. Like it matters which one you get. It depends on which one you get. Many people today are trying to do a cash-shuttled feature. And the way that works is that you go and you develop some index, you build some sort of tracker of price.

4. 30:01

   and then you go and try and build a cash-shuttled future on top of it, which means that if you can bet on what that index price is, and if you're right, you make money, you get paid in cash. You don't get paid in compute. The problem of that is, let's say you end up in a situation where you would like to take that cash and turn it into compute. The price that you want might be completely different based on the actual cost of compute.

5. 30:27

   And so you have de-risked yourself on an index price that is divorced from the actual cost of compute. That's kind of the problem of making a cash-settled index if you don't have physical settlement. What is it that you guys are doing that is different from that cash-settled futures market that you don't think is working? If you buy GPUs from us that start in November, you get GPUs. You don't get cash.

6. 30:52

   That's the difference. And so that means that you can use that to risk and de-risk yourself with actual settled GPUs. Like you get the VM time or you get the hardware time. What about if I could use the compute, but other people will pay me such a premium for that compute that it would make sense for me to put it on the market anyway? Today, a lot of people, when they're sort of running their product on GPUs, they make some margin.

7. 31:19

   Let's say that margin is 50%. One option that you could do is that you could run your product, make 50% margin. Another option that you could do is that you could sell that compute for more than 50%. And then you make a profit either way. One of the things that SF Compute allows in our market is for people to buy that capacity and then as prices go up to sell it, which is a sort of attractive opportunity for you.

8. 31:45

   You sort of de-risk yourself against your own product not making as much money as you thought.