AI算力期货：下一个大宗商品市场？

What if AI had its own futures? Oil prices can swing dramatically, so airlines often use futures contracts to lock in fuel costs months ahead of time. Now, a similar idea may be coming to artificial intelligence. AI companies are spending billions of dollars on computing power to train and run models. But the future cost of that compute can be difficult to predict. Start-up Silicon Data has partnered with the CME Group to launch what could be the world's first futures contract based on the price of AI compute, pending regulatory approval. Do you think it will one day become as big and prominent as oil... Oil futures?

I think will be larger. I think everyone will see compute will be the largest human resources surpassing all energy combined. Here's what you need to know about the race to turn GPU power into Wall Street's next big commodity. GPUs, or graphics processing units, are the specialized chips that power AI. Most companies don't own them outright. They rent access through cloud providers in a newer wave of companies called neo clouds. That means the price of running an AI model, training a new one, or just keeping the company going can swing drastically. Carmen Li is the CEO of Silicon Data, and she's been watching those prices closely. Her company tracks real time GPU rental pricing across cloud providers and GPU marketplaces. She says this volatility is a problem big enough to need a financial solution. Why does AI compute need a futures market, and what kind of problems are you trying to solve?

If you consume a lot of electricity, if you consume a lot of GPUs, you have a big cost volatility you will be facing down the stream. So you want to hedge that away using futures as well. So really for the, for the underlying economy to function properly, efficiently and in most transparent manner, so you need a futures option market to hedge. The other part of this is GPU prices. Volatility doesn't hit every company the same way. A large enterprise can call up one of the big cloud providers such as Amazon, Microsoft and Google, and negotiate a multi-year contract for GPU access at a fixed hourly rate. It gives these big firms cost predictability, but the trade off for that locked in price is the company is also locked into other agreements, so they have to use that specific chip, that cloud provider and of course, that price, for years. And in an industry that's moving this fast, that can be an innovation killer. So the flexibility, the optionality is critical to provide equal market access. And it's hard to predict and hard to forecast. Because it's so volatile. It's so volatile, and you don't know what's going to happen to your business model. So you might need more than you think. You might need less. Right now, we're at a high point of uncertainty. A lot of people don't know how much computing power they'll need in the next year. And a lot of suppliers of that computing power right now don't know how many GPUs and to what capacity they should order. And the manufacturers like Nvidia, they don't know how many they should produce. So there is that uncertainty. And those are people who are trying to hedge their exposure. Silicon Data has built what it calls GPU price indices. These are benchmarks that track the hourly rental cost of specific chip types across providers. The most closely watched is H100, which is currently the workhorse chip of the AI industry. People hear compute; they think about Nvidia chips. So is it GPU hours?

Is it cloud capacity? Tell us about what investors are going to be trading. We want to make it easier for people to hedge in the spot market today. Right. If you go to NeoCloud, go to hyperscalers, they're going to charge you by per GPU power rate. So for you, for everybody it's easy for them to hedge. Ideally our derivatives, our indices should be per GPU power. The mechanics are similar to other futures markets. If a company expects to need compute months from now, it can use a futures contract to lock in a price and protect itself if GPU rental costs rise, and if a provider owns compute capacity and is renting it out, it can use the same market to protect itself if prices fall. With a futures contract, what you're doing is the person on what we call the user side. We say that that's the long position that needs to be matched by a short position of someone who is committed to giving it to you at that particular price. So you need both to clear. There's also a third category of players speculators. These are traders who aren't hedging anything but have a view on where GPU prices are going. Speculators are a very important piece of the ecosystem as well, right?

You need natural hedgers. You need market makers. You need speculators. So these are all for price discovery, which is also a big indicator how much faith they have in the liquidity of the product. But liquidity is not automatic. Just because AI is becoming widely used does not mean every company actually needs high-end GPU access at a scale that would require a futures contract. There aren't enough people who actually legitimately need that kind of compute right now. Look, most of us are going to use ChatGPT. A lot of people don't even need the $20 a month version. This is all pending regulatory approval. And those regulators will authorize a benchmark unit. A barrel of oil is an example of a benchmark unit anywhere in the world. Traders know exactly what they're pricing, but a GPU hour is not like that. Silicon Data has counted more than 50 different configurations of the H100 chip alone, so all of the different combinations of processors, RAM connectivity, data center locations and each one trades at a different price. What we do is we normalize the 150,000 traded prices coming to our platform every day to a base of H100 case, and then you can compare them against each other. So the very complicated normalization step even before the index calculation step. But standardization is not a new problem for commodities markets. Even corn futures have detailed specifications for what kind of corn can be delivered, but compute introduces its own version of that challenge. So the CFTC is the body that's in charge of first determining whether a certain futures product is okay to be offered on these different exchanges. You'll have to specify the size of the contract. You'll also have to specify the trading times. You'll have to specify when the contracts settle, and you'll have to specify what happens at settlement, whether it's physical settlement or if it's a financial settlement. So all of those things are going to be scrutinized. People use compute for everything from legal document review, all the way to helping you create a great movie. Right?

It does touch everyone's life. You may not need to hedge your compute spend, but if your enterprises, you spend millions of dollars on either GPU or compute or token, you have to consider. Do you want to hedge your volatility away?