Transcription
[Music] During the current AI boom, the single largest beneficiary has been Nvidia. They've sold hundreds of billions of dollars worth of AI GPUs at massive profit margins. This has propelled them to become the most valuable company in the world with a market cap surpassing $4 trillion.
In terms of stock market valuation, the second biggest AI beneficiaries have been the data center companies. The so-called hyperscalers, including Microsoft, Amazon, Google, and Oracle, have seen hundreds of billions of dollars of demand for AI-related computing power. Hyperscaler is a colloquial term that refers to the established mega cloud service providers. All the hyperscalers have other businesses in addition to their cloud offerings. Microsoft and Oracle sell enterprise software. Google has Google Search, and Amazon has its e-commerce business.
The AI boom has also created a new class of company, the so-called Neoclouds. The largest of which is Coreweave, which went public in early 2025. The Neoclouds are pure-play AI computing companies. They do nothing else. They buy Nvidia's GPUs and rent them out to AI companies like OpenAI and Anthropic. The Neoclouds have also seen astronomical demand over the past couple of years.
The most surprising beneficiaries of the AI boom are Bitcoin mining companies. Bitcoin mining is an extremely competitive business. The difficulty rate is always increasing, and the specialized mining rigs become obsolete very quickly. Almost all Bitcoin miners lose money once you account for depreciation. Over the past couple of years, a number of Bitcoin mining companies have transitioned into AI data center companies. These companies have seen their share prices surge as investors are far more optimistic about AI than Bitcoin.
While the AI data center companies have enjoyed surging demand and sky-high stock market valuations, the foundations of this growth are very shaky. Over the past three years, the AI industry has grown into a convoluted web of interconnected relationships and arcane financial engineering, all backed by hundreds of billions of dollars of debt and other liabilities. In this video, we'll analyze the fragile foundations of the massive AI industry. We'll focus specifically on Coreweave, the largest of the so-called Neoclouds. Its story is emblematic of the financial engineering going on in the industry today.
The recent AI boom has had massive implications for investing in the stock market. Now more than ever, it's important to have up-to-date information when assessing your investments. That's why I use Mumu, which is also the sponsor of today's video. Mumu is a one-stop trading platform trusted by over 26 million users across the US and globally. They offer zero commission on stocks, ETFs, and options trades, free Level 2 data, and real-time option chain data.
For example, let's look at Nvidia. If we go to the comments tab, we can see what other Mumu users are saying about the company. They also have a list of recent news articles about Nvidia, but my favorite is the company tab. It shows you analyst estimates and ratings on the stock. It also has a ton of fundamental information. Here we can see Nvidia's price-to-sales ratio. We can switch to price-to-earnings. Nvidia's PE ratio of 53 is very high, but the stock has always been expensive, so it's actually pretty close to its average PE ratio over the past 10 years. We can see their quarterly revenue by segment, financial results versus analyst expectations, and a whole host of other useful information. It's super useful to be able to kickstart your investment research all within the Mumu app.
Currently, Mumu is running an amazing new user promotion. They offer a 3.6% yield on your idle cash, but for the first 3 months, they offer an additional 4.5%, bringing your total yield up to 8.1%. Additionally, if you open a new account and deposit $100, they'll give you a $20 fractional share of Nvidia for free. The more you deposit, the more Nvidia shares they'll give you, up to $1,000. This promotion only lasts until the end of the year, so make sure to sign up today with the link in the description below.
We made a video about Coreweave a few months back when they IPOed earlier this year. If you want to understand the company's background, you can watch that video first. Link in the description below. Here, we'll give a brief SparkNotes version of the company's history.
The company that would eventually become Coreweave was founded in 2017. At first, it was called Atlantic Crypto and was an Ethereum mining company. Unlike Bitcoin mining, which uses specialized mining rigs, Ethereum used to be mined with GPUs, so they bought a bunch of Nvidia GPUs to mine Ethereum. In 2018, the price of Ethereum collapsed. It was no longer profitable to mine. So, Atlantic Crypto changed its name to Coreweave. Instead of mining Ethereum, they started renting out their GPUs to external customers. For the first few years, the company was pretty small, as there wasn't that much demand for GPU compute.
In late 2022, ChatGPT was released, and everything changed. Demand for GPU compute exploded. The single largest end consumer of GPU compute by far is OpenAI. OpenAI's single largest corporate backer is Microsoft. Microsoft first started investing in OpenAI in 2019. To date, they've invested $13 billion into the company. Of the $13 billion of investment, most of it was not paid in cash. Instead, Microsoft provided OpenAI with credits to use the Microsoft Azure cloud computing service. OpenAI uses Azure to train and run its large language models.
The problem is Microsoft didn't have enough GPUs to fulfill the massive number of cloud computing credits they gave to OpenAI. So, they went to Coreweave. Coreweave became a subcontractor to Microsoft. Coreweave supplies computing power to Microsoft. Microsoft resells its computing power to OpenAI in exchange for the credits which Microsoft previously gave to OpenAI. Microsoft is effectively acting as a middleman.
In 2024, Coreweave generated $1.9 billion of revenue, of which 62%, or $1.2 billion, came from Microsoft. Their second largest customer represented 15% of revenue, or about $300 million. Based on other information in their regulatory filings, we can deduce that the second largest customer was Nvidia. Nvidia is Coreweave's single largest supplier. Coreweave spends billions of dollars to buy Nvidia GPUs. Nvidia was also Coreweave's second largest customer.
Why was Nvidia paying $300 million to rent back its own GPUs that it sold to Coreweave? Nvidia has something they call the DGX Cloud. This is a cloud computing service where you can rent Nvidia GPUs. While Nvidia builds the GPUs, they do not own or operate them. They sell them to data center companies like Coreweave and lease them back. Nvidia is basically acting as a middleman. The Nvidia DGX Cloud will become important later on, so keep it in mind, and we'll return to it later.
Microsoft didn't have enough computing capacity to fulfill the credits they gave to OpenAI. That's why they subcontracted out to Coreweave, but Coreweave didn't have the capacity either. To fulfill the contracts they received from Microsoft, Coreweave needed to build dozens of new data centers. There are three aspects to building a new AI data center. First, you need the land and physical building. Secondly, these data centers consume copious amounts of power. You need to hook them up to an electric utility provider. And finally, you need to fill them up with Nvidia's latest and greatest GPUs, which are very expensive. All of this costs a lot of money. Coreweave needs to spend many billions of dollars upfront to build out this data center capacity. Once it's built, they rent them out to Microsoft or their other customers and will hopefully make back the investment over time. To make these investments, Coreweave has relied heavily on debt financing.
Now, let's look at Coreweave's financial results. In 2024, they generated $1.9 billion of revenue and posted an operating profit of $324 million. However, they reported an adjusted net loss of $18 million. In 2025, their revenue grew rapidly thanks to strong demand for AI workloads. In the first 9 months of the year, Coreweave generated $3.6 billion of revenue, but their operating profit fell to $44 million, and their adjusted net loss exploded to $860 million.
I calculated my own adjusted net loss number. I took the net loss and subtracted their non-cash fair value adjustments and changes in income tax provisions, both of which are basically accounting noise. Coreweave reports their own adjusted net income number, which also excludes stock-based compensation and acquisition-related costs. I think these should be included, as they are indeed real expenses. But even using Coreweave's far more generous adjusted numbers, they're still reporting net losses. So why are they losing so much money?
Once the data center is built, there are ongoing costs to operate it. You need to have technicians who maintain the data center, and you need to supply it with electricity. You also need to recognize depreciation expenses on the GPUs. At least once every 2 years, Nvidia comes out with a new GPU model, which is way better than the last one. The older generations become obsolete very quickly. Coreweave depreciates its GPUs based on an estimated useful life of 6 years. This is probably way too optimistic, but let's just roll with this for now.
In the first 9 months of 2025, Coreweave generated $3.6 billion of revenue. After all their operating expenses, including depreciation, they eked out a tiny operating profit of $44 million. That's an operating margin of 1%. They barely broke even.
To build these data centers in the first place, Coreweave had to borrow a bunch of money. As of September 30th, 2025, they had $14 billion of debt. In the first 9 months of the year, they incurred $840 million of interest expense. They're barely breaking even on the operating level, so they don't have nearly enough profit to pay their massive interest expense. Coreweave's solution is to grow even more, borrow even more money to pay back the old loans, and build even more data centers.
On top of that, Coreweave says it has a revenue backlog of $56 billion. These are multi-year contracts. The majority of them last more than 2 years, and about a fifth of them last more than 4 years. To fulfill these contracts, Coreweave is going to have to build a lot more data centers. As of the end of Q3 2025, Coreweave had 590 megawatts of active data centers. They have 2.9 gigawatts of contracted power. So, they already have contracts in place to increase their data center capacity almost 5-fold. This is going to cost a lot of money. It has been reported that 1 gigawatt worth of Nvidia's most advanced GPUs cost approximately $35 billion. To fill up the remaining 2.3 gigawatts worth of data centers with Nvidia chips will cost approximately $80 billion. Coreweave already has $14 billion of debt, of which almost $4 billion is coming due in 2026, but they also have a lot of off-balance sheet liabilities.
For the most part, Coreweave doesn't own the physical buildings of its data centers. At the beginning of the video, we talked about how Bitcoin mining companies became unlikely AI beneficiaries. Bitcoin mining rigs can pretty much only be used to mine Bitcoin. They can't be used for AI. But the Bitcoin mining companies do have one asset that's useful for AI. Their mining rigs are located in data centers. Bitcoin mining is very energy-intensive. Just like AI, Bitcoin mining companies already have contracts in place with electric utility providers. Many Bitcoin miners have decided to stop mining Bitcoin in recent years. Instead, they'll lease their data centers to cloud service providers who will repurpose them for AI.
One of these former Bitcoin miners is a company called Applied Digital. In January of 2025, Coreweave entered into a binding agreement with Applied Digital to lease the data center. Applied Digital leases the building and supplies electricity to it. Coreweave deploys and operates the GPUs. Coreweave will pay Applied Digital $7 billion over 15 years. That's almost $500 million per year for 250 megawatts of capacity. That's a cost of almost $2 billion per gigawatt per year. Coreweave has 2.9 gigawatts of contracted power. If we extrapolate the pricing from the Applied Digital deal, this will cost Coreweave about $6 billion per year. So, in addition to their $14 billion of debt, Coreweave could have as much as $6 billion of annual rent and utility payments going out for as long as 15 years. And keep in mind, Coreweave is currently losing money.
They have a $56 billion backlog. This is massive when you consider that they only generated $3.6 billion of revenue in the first 9 months of 2025. This implies massive growth, and they need to grow because they've already taken on so much debt and contractual commitments. This $56 billion revenue backlog. Where does it come from? Who are the customers?
Coreweave's biggest customer used to be Microsoft. Microsoft is using Coreweave as a subcontractor to fulfill the cloud computing credits they gave to OpenAI. As it turns out, OpenAI is so hungry for computing power that the credits they receive from Microsoft aren't nearly enough. In March of 2025, OpenAI signed a deal to purchase $11.9 billion of computing power from Coreweave. This is a multi-year contract intended to begin in late 2025 and last until 2030. In May of 2025, OpenAI added another $4 billion to its Coreweave order. In September, they expanded the deal again, adding up to $6.5 billion on top. To date, none of this computing power has yet been delivered to OpenAI. The data centers are still under construction. So, before OpenAI even received anything, they've already expanded the deal twice, bringing the total value up to $22.4 billion.
OpenAI is losing money. They're losing a lot of money. In the third quarter of 2025 alone, they lost $15 billion. We know this because Microsoft owns 27% of OpenAI and they have to report their share of OpenAI's net losses on their own income statement. In 2025, OpenAI is only expected to generate $13 billion of revenue. OpenAI has tens of billions of dollars worth of commitments to Coreweave, but they also have many other commitments to many other cloud computing companies. In total, they have $1.4 trillion worth of commitments. What if OpenAI doesn't have enough money to make good on these commitments? That would be very bad for Coreweave.
Recently, Coreweave has tried to downplay their customer concentration risk. In the third quarter earnings call, Coreweave CFO explained, quote, "Today, no single customer represents more than approximately 35% of our revenue backlog, down from approximately 50% last quarter and even more meaningfully from approximately 85% to begin the year." The backlog is $56 billion. 35% of $56 billion is $19.6 billion. OpenAI's contract value with Coreweave is up to $22.4 billion. Presumably, some portion of this is optional. The binding part of the contract is only $19.6 billion.
Let's analyze Coreweave CFO's statement a little bit further. At the beginning of the year, 85% of the backlog was attributed to one customer. Their first deal with OpenAI was signed in March of 2025, so that can't have been OpenAI. It was almost certainly Microsoft. At the beginning of the year, Coreweave's backlog was $15 billion. So, Microsoft accounted for almost $13 billion of backlog at the time. In the first 9 months of 2025, Coreweave recognized $2.5 billion of revenue from Microsoft. That would bring the Microsoft backlog down to $10.5 billion. It's possible that Microsoft placed more orders to Coreweave in the first 9 months of this year. But even if they didn't, their remaining performance obligation is still $10.5 billion.
Remember that the reason Microsoft started working with Coreweave in the first place was to fulfill the cloud computing credits they gave to OpenAI. So, presumably, the majority, if not the vast majority, of the compute Microsoft buys from Coreweave is resold to OpenAI. We know that another one of Coreweave's customers is Google. In June of 2025, it was reported that OpenAI signed a contract to purchase computing power from Google. Google doesn't have the GPU capacity to fulfill this order. Google also doesn't want to build these data centers itself. The margins are too low. It's not worth it. So, Google did the exact same thing that Microsoft did. They subcontracted it out to Coreweave. Google is acting as a middleman. We don't know how big the Google deal is, but it could easily be multiple billions of dollars.
Only 35% of Coreweave's backlog comes from OpenAI directly, but the majority, if not the vast majority, of the compute sold to Microsoft and Google is probably going to be resold to OpenAI. In reality, OpenAI probably represents more than 50% of Coreweave's backlog. It could be as high as 60% or even 70%. OpenAI is using Microsoft as a middleman because they already have Microsoft cloud computing credits. But why do they do this deal with Google? Why not just contract with Coreweave directly?
Coreweave funds the data center construction with debt. If the lenders see that they have too much customer concentration with OpenAI, they'll think this is too risky. Google has an investment-grade credit rating. Even if OpenAI goes bankrupt, Google will still be obligated to fulfill their contract with Coreweave. Google is taking on a risk. They have to be compensated for this risk. Google buys computing capacity from Coreweave at a cheap price and resells it to OpenAI at a higher price. Google's relationship with OpenAI and Coreweave has nothing to do with technology. It's a financial relationship. Google is monetizing its investment-grade credit rating.
This is all extremely reckless. OpenAI's desire for computing capacity is far in excess of their ability to pay. That's why they do these complicated deals with Google and Coreweave. This spreads out the risk to more and more counterparties. The financial shenanigans extend across the entire value chain, going all the way up, even to Nvidia.
About a month ago, we made a video discussing Nvidia's roundtrip investments. In the video, we mentioned a suspicious deal they made with Coreweave. Upon doing research for this video, we found some more information about the deal, and it's even crazier than we originally thought. In September of 2025, Nvidia signed a strange deal with Coreweave. If Coreweave ever experiences a lack of demand, Nvidia will purchase up to $6.3 billion of computing capacity from them. This backstop improves Coreweave's creditworthiness, thus allowing them to borrow more money and purchase more Nvidia GPUs.
If Nvidia is forced to purchase $6.3 billion of computing capacity from Coreweave, what will they do with it? Nvidia has been a customer of Coreweave since at least 2024. They purchase computing capacity from Coreweave and resell it at a markup with their DGX Cloud service. But why would any customer use DGX Cloud? Why not just go directly to Coreweave or any of the other cloud service companies? Why would anyone pay a markup to Nvidia as the middleman?
Nvidia does not regularly disclose how much revenue it generates from DGX Cloud. The last time they gave any disclosure was in August of 2024. On an earnings call, Nvidia CFO said, quote, "We expect our software, SaaS, and support revenue to approach a $2 billion annual run rate exiting this year with NVIDIA AI Enterprise notably contributing to growth." Software, SaaS, and support revenue includes DGX Cloud. She was referring to the end of Nvidia's fiscal year 2025, which ended on January 26th, 2025. So, at that time, DGX had an annual revenue run rate of less than $2 billion. They thought it could grow to $2 billion by January of 2025.
In early September 2025, the information reported that Nvidia was scaling back its DGX Cloud business. They were not getting very much traction with it. Very few customers were willing to pay the marked-up prices they were trying to charge. So, it probably never got to the $2 billion run rate they were previously targeting. As for the cloud computing capacity they had already contracted, they'll use this for internal research and development. This news came out on September 12th. What internal research is Nvidia doing that they need billions of dollars of compute for? They don't develop large language models. It doesn't make any sense. You would think that since they're scaling down their DGX Cloud business, they should probably reduce the amount of computing capacity they rent. But they did the opposite. Just a few days later, they signed this capacity backstop deal where they might have to buy $6.3 billion of compute from Coreweave.
To understand just how bizarre this is, let's apply this same deal structure to any other industry. Let's imagine hypothetically that Ford sells 100,000 cars to Hertz, a car rental company. One day, Ford decides to start its own car rental business. We'll call it Ford Rental. Ford Rental is a direct competitor to Hertz. But instead of building new cars for their rental business, Ford leases back the cars they previously sold to Hertz. Ford Rental would be economically inefficient. Ford Rental has to charge above market rates to cover their own leasing fees they pay to Hertz. Because of this, Ford Rental is a commercial failure. They don't have enough customers, so they shut down Ford Rental, but they don't return the cars to Hertz. They keep them for internal research purposes. They never explain what this internal research is. Not only does Ford not return the cars to Hertz, they offer to lease even more cars from Hertz in the future. This allows Hertz to borrow money and purchase new cars from Ford to refill its inventory. Ford will lose money on this transaction. They're buying back their own cars for which they have no use, but it allows them to recognize more revenue from Hertz. This hypothetical deal is absurd on its face, but this is exactly what Nvidia is doing with Coreweave.
Coreweave has become very important in the current AI ecosystem. They're an important supplier to OpenAI and an important customer to Nvidia. Yet, their financial position is incredibly precarious. The entire industry is coming together to prop them up, thus spreading around the risk. Ultimately, OpenAI is the end user of all this computing capacity. To pay for all of this, OpenAI needs to grow its annual revenues to hundreds of billions of dollars within the next 5 or so years. If they fail to do this, they'll default on their obligations to Coreweave and the many other cloud service providers they've signed contracts with. If this happens, a lot of companies will lose a lot of money.
All right, guys, that wraps it up for this video. What do you think about Coreweave? Let us know in the comments section below. As always, thank you so much for watching, and we'll see you in the next one. Wall Street Millennial signing.