원문 보기
ROI, I mean, there's one more thing that I think you and I will disagree on . What is the cost of the H100? This is the basic level in the industry. Not to mention Vera Rubin. What is the value after six years, because remember, Coreweave depreciates over six years. So, for the seventh year, these are free calculations.
Is this free computing power? So, will we then package these calculations? Look, Larry Fink said here last week that the next mortgage collateral will be computing power. Yes. But no one I know has really taken the step to unify computing or make it a two-sided market. Not yet, but I mean, these chip and GPU securitization efforts are going to be very important because it could lower the overall cost of capital.
Yes, because if you created a securitized market with GPUs as collateral, you would probably reduce the overall cost of capital, right? I completely agree, and that's what I'm waiting for, but I'm not getting it, but it will take some time. These...these markets are not created overnight. You're talking about a pretty significant effort.
Yes. Which require a lot of data and, actually, standardization, because if you create this, you know, I don't even know the answers to some of the questions. For example, if you have a GPU in a server rack somewhere here, is that the same as the one sitting here, and do you draw a distinction between what Coreweave does and maybe what other neoclassical companies do ?
Is this the same thing or not? So, because But don't you think Morgan Stanley would have mentioned this during their report? Would Goldman have mentioned this in its report? They did it, I don't know, maybe they just weren't asked. We asked Larry Fink about this last week, and he answered. He was talking about how they're now pulling all the data together, and they're...
But I'm talking about making it a two-sided market. You know, these guys come out and say, listen, the market is open. We are completely ready. Where is this? Why weren't they ready then? For God's sake, they're already making a discovery with Becky, when they were all sitting around the table talking about the half- trillion, you know , opportunity there.
at the table with you. This is not sitting at the table to create a two-sided market . but that wasn't really that long ago. You are very impatient. Er, impatient? I would say this should have been done yesterday. You are very impatient. Let 's go back to... let 's go back to yes. correct focus? I think everything we're talking about right now is the right focus, because on Friday this segment of the market was the area that everyone was fleeing from.
I mean, if you didn't follow this on Friday, I think you're deeply mistaken. Will the rebound last? I ... Okay, will the rebound last? Listen, many are defending Coreweave this morning, including Goldman Sachs, which upgraded the stock to a “ buy” rating. Send it to me. They talk about annual revenue per megawatt and internal rate of return without taking into account the debt load, and you know, a lot of other things, claiming that they believe in the company's differentiation from other Neo-clouds in terms of performance and cost, as well as the ability to scale up new capacity.
There was something else there. People just think that everyone works equally efficiently. When I first met Mike Linton, I talked to him. He said, "Listen, we know how to build because we built for Bitcoin." At first I told him, "Oh, that's just too..." for me. I'll take that question back to you, but it's a difficult one to answer.
Is this a real rebound, or will these concerns haunt stocks for a long time to come? I'll throw the ball back to you and say, "Tell me where interest rates are going." Okay, I don't know. I mean, they've been falling for the last few days. Escaping to asset reliability made me feel like Coreweave was better.
Now look at Coreweave debt. Is that Coreweave private debt trading at par? I think there's a great piece from Barclays that says, " Listen, about the SVB redux you were thinking about, about there being a crash like in '23..." It's not going to happen. Now I read this and say to myself: I didn't even think it could be a crash like in the 20th.
My God, he's defending something I never thought would happen. SVB redux, what is it? to be like that...well, now if I put that in your head, you're going to worry. You must be thinking: this is the crash we saw in '23. May cause a crash. Oh my God. I know. I mean, well, it's just great material because it creates a straw man and then just says "no." Yes, but in the meantime, it will most likely still scare people.
I was afraid of that incident with SVB. The hourly price of renting one of Nvidia's major AI chips on a one-year contract has jumped 60% in 12 months. 60. If you run an AI company, you can't just ignore these price increases. And today, a startup has emerged that claims it can reduce these costs. One of the first customers is said to have reduced their computing costs by 67%.
Nvidia itself wrote a check to support them. I'll tell you how it actually works. And then I'll show you something that seems absurd at first. Why would a company that sells chips fund something that might result in fewer of them being needed? The answer lies in how money moves, and I almost missed it.
Here is the news. A two-year-old Tel Aviv-based startup called Deci AI just announced it has raised $200 million. The investment was led by Lightspeed and Bessemer, with participation from the venture capital arms of Nvidia and Dell. According to reports, the company is valued at over a billion dollars.
Founder— Amnon Shashua, co-founder of the autonomous car company Mobileye. And the entire team only has about 40 people. 40 people. Billion dollar estimate. Remember this. So what are they selling? It's called Maxwell. Imagine a restaurant where half of the ingredients are wasted due to poor kitchen planning.
Maxwell is not building a bigger kitchen. He monitors the software , figures out how the program interacts with data and chips, and rewrites the low-level machine code for better compatibility with the hardware. Your source code does not change. Same program, same chips, less waste. The company reports that some data pipelines are running more than 38 times faster, and some AI inference elements are running more than four and a half times faster.
Now be frank. What did you think when you heard about 38 times? I was surprised too. These figures are provided by the company and in the launch coverage. These are probably the best results, and I'd like to see independent testing before believing it everywhere. Take this into consideration. Here's a phrase I liked.
They plan to charge a portion of what you save. They did not specify the size of this share, but Shashua told the Wall Street Journal that for large AI labs, even a 10% savings could amount to hundreds of millions of dollars. So their incentives are the same as yours. Why is this important now ? Let's go back to that 60 %.
I was watching a conversation with Jim Cramer where they raised a question that I can't stop thinking about. How much does an AI chip really cost ? They talked about the H100, which is already a generation out of date, and its value in years to come. Will computing become almost free by then ? Then they remembered the words of BlackRock's Larry Fink during last week's earnings call: computing could go the way of mortgage-backed securities.
The idea is to bundle the chips into packages and sell them in parts. This could lower the cost of capital for anyone building data centers. But they made it clear that this wouldn't happen overnight. So, here's the "yes, but." An older chip should lose value every year, right? However, the hourly rental price of the H100 on an annual contract has increased by 60%.
Renting is not the same as owning, and the rental price is not how much the chip is worth on the balance sheet. But this is a clear signal. The older the chip, the higher the rental price. This indicates that demand for computing exceeds supply . And it is precisely this window of opportunity that a company like Double AI enters.
If you can't get more chips, and the ones you do have are more expensive, then getting the most out of each one becomes the main game. Now let 's think together. Nvidia makes money when people buy chips. Double AI claims to help people use less hardware. So will Nvidia win, lose or not care? Here's the moment I almost missed.
The Wall Street Journal reported that Double AI is spending about half of this funding round on purchasing 1,000 B300 Blackwell chips from Nvidia. Half. So, most of the round's money is going towards Nvidia hardware, although we don't know how much Nvidia itself has invested. Nvidia can still win in any case.
There is a second level, and this is my assumption, not their words. When something becomes cheaper to operate, people tend to use it more often. If Maxwell makes AI workloads cheaper, it could justify more projects, leading to even greater demand for the chips. This is a possibility, not a promise.
And the risk is quite real. If efficiency gains become so significant that customers actually need much less hardware, it will become a hurdle for every chip vendor. I have no way of knowing what Nvidia is thinking, and this is not financial advice. So let me summarize. 60% is the increase in the hourly cost of an annual H100 rental contract, according to SemiAnalysis, as reported by the Journal.
67% savings—this is Double AI's own claim, based on the experience of one of its first customers. As for why Nvidia supported them, neither company commented. But one possible reason is that money is flowing back into Nvidia, with the potential for even greater demand in the future. If you only remember one sentence, let this be it.
When computing power becomes scarce and expensive, the people who help optimize it can become almost as valuable as their creators. If you want more such analyses, subscribe. And next time I want to delve into the idea of packaging chips the way they used to package mortgages on Wall Street. Because one simple question about two seemingly identical chips made even the participants in that conversation doubt.
Comments 0