The State of AI Data Centers in 2026: Who’s Making Money?

[00:00:00] Speaker 1: For the past couple of years, you have probably heard a lot about AI data centers, giant campuses, gigawatts of power, spending plans in hundreds of billions of dollars. Hearing about them is one thing, knowing what's actually in there is another. So this video asks three questions: How much of this has actually been built? Who needs all these computers? And will the companies paying for them earn their money back? The last question is where the stakes are. The biggest technology companies have been increasing their spending on buildings and equipment much faster than the cash their businesses bring in. If the computing they are buying gets used and paid for, that spending is laying the foundation of a new industry. If it doesn't, someone ends up owning very expensive buildings and equipment that don't earn enough to pay for themselves. The answer won't be the same for everyone involved. A company can sell chips at a healthy profit today, while the customer who bought them is still years away from knowing whether the purchase paid off. Both can be true at the same time, and a good part of this video is about why. So I put together the state of AI data center for 2026, covering what is public until October 5th. If you want to watch my videos in advance, you can join the channel as a member. Members can also propose research topics for the next videos. Research tier members can also download infographics for my videos, as well as watching ad-free version of every video. And this is for educational purposes, not financial advice. We'll start with the first question, and with the real place. What has actually been built? Outside Abilene, Texas, there's a campus of 8 buildings made for AI. The work on it began in June 2024. Three companies share it, and the way to split the work is common in this industry. A developer, Crusoe, builds a campus, oracles leases it, and runs the computers inside, and OpenAI is a customer using that computing. What makes Abilene useful isn't the size, it's the fact that it didn't open all at once. It has been arriving in stages, and each stage is a different kind of milestone. First came the announcement and start of the construction. Within a year, according to Crusoe, the first two buildings were energized, meaning they were connected to power. In September 2025, Crusoe announced that the first phase was up and running. And this September, Oracle told investors that 6 of the 8 buildings had been delivered, about 3/4 of the campus capacity. Each of those steps is real, and none of them proves the next. A building connected to power isn't necessarily full of working computers. And when Oracle says capacity has been delivered, that's a statement about the buildings. It doesn't tell you how hard the computers inside are working on any given day. So when you read that the data center has been built, it helps to ask which milestone that means. At Abilene, more than two years after work began, the answer was most of the campus delivered, with the last building still to come. You notice that the size of a place like this is described in electricity, not in floor space. Data centers are measured by how much power they can draw in megawatts, millions of watts, and for the biggest size in gigawatts, thousands of megawatts. That's because power is what the computers inside actually consume. A site's electrical capacity tells you far more about how much computing it can hold than its square footage does. Not every site tells us as much as Abilene. Elon Musk's XAI, or maybe we should call it SpaceX's XAI, hasn't published the size of its Colossus II site in Memphis. The best public figure comes from a research group, Epoch AI, which estimated from satellite images taken in June, that the site had roughly 950 megawatts of power for its computers. For some of the most important sites in the world, an outsider's reading of satellite photos is the best evidence theorist. So a substantial amount of AI computing is already operating at places you can find on a map. But the biggest campuses open in stages. And an announcement, a construction start, a power building, and a fully running campus are four different things. Most of the headline numbers though are about something else. Everything that isn't running yet. Think of housing in a fast-growing region. There are homes people already live in. There are homes with a frame up and a roof going on. And there are lots with permits, proposals, and sketches. Some of those lots will become homes, some never will. The real estate firm, Cashman and Wakefield, accounts data centers across the Americas in the same three piles. One thing belongs right beside its numbers. They cover data centers of every kind. The ones that store your photos and process car payments, as well as the ones running AI. Very few public sources separate AI from the rest. In the first half of 2026, Cashman and Wakefield counted about 50 gigawatts operating. Almost all of it in the United States. And that's what's in service today. It doesn't tell us how busy the computers inside are. About 38 gigawatts were under construction. And more than 90% of that capacity was already committed to customers. So work is underway. And someone has signed up for most of it. What the number doesn't tell us is when it will be finished. As we'll see later, finishing is where much of the difficulty lies. Meta's Hyperion campus in Louisiana is a good example of that middle pile. It's under construction. Regulators approved new gas power plants in August 2025. And by our October cutoff, no operating capacity had been reported there. Work underway is not the same as capacity available. Then comes the third pile. 285 gigawatts planned. That's more than five times what's operating today. Planned here includes every proposals. It is not a forecast of what will be built. And none of it is capacity anyone can use. This matters because most of the eye-catching announcements belong to the second and third piles. A new multi-gigwatt project tell you about ambition and sometimes about signed commitments. It doesn't tell you how much computing is available today. For that, you have to look at the first pile and at sites like Ebelin, building by building. The last thing to understand about what's been built is where it is. Because the map has been changing. Start with the United States. The research firm Synergy tracks the capacity operated by the largest cloud and internet companies, the Amazons, Microsofts, and Googles of the world, often called hyperscalers. With so many announcements in the Gulf, in Europe, and in Asia, you might expect that kind of capacity to be spreading out around the world. It hasn't been. In Synergy's latest reading, from late 2025, the US held 55% of it, a slightly larger share than three years earlier. Inside the US, the growth has moved. Northern Virginia, just outside Washington DC, has long been the industry's biggest market. And it still is. But Synergy says hyperscaler capacity in Texas grew 71% over the past year, about twice the global rate. And of everything under construction across North America. The property firm JLL says about three-quarters is in what it calls frontier markets. Places like West Texas, Ohio, Louisiana, Indiana, and the Carolines. It looks a lot like heavy industry, following land and electricity. The established hubs are full. So new sites go where a great connection is available. And Northern Virginia really is full. In its rental market, buildings leased out to tenants, including the big cloud companies themselves, just 0.2% of space was empty in the first half of 2026, even less than a year earlier. Following the power comes with a catch. Texas is also the state whose grid operator paused approvals for large new data centers to switch on this August. Where the build goes next depends on grids. And we'll come back to that. Outside the US, the evidence is thinner, and measured differently from place to place. So instead of a tour, here are a few places that each show something the US figures can't. According to Cashman and Wakefield, operating data center capacity across Europe, the Middle East, and Africa, grew by at least a third in two years. It reports those markets together as one region. That's a big change. And it's growth in buildings that are actually running, not plans. It's also a count of every kind of data center across the whole region. While the US share we just heard counts only the largest cloud company's capacity. So we cannot compare them side by side to say which grew faster. Within that region, the Middle East shows the gap between ambition and operation most clearly. JLL counts about 1.6 gigawatts actually running there. Against a much larger pipeline it describes as paused, not canceled. In late 2025, the US government authorized two golf companies, the UAE's G42 and Saudi Arabia's HUMAN, to buy a large number of NVIDIA GB300 chips, which are imbued for both training AI models and running them. An authorization lets the chip be bought. It doesn't mean they have arrived. And the first phase of the region's flagship project, Stargate UAE, was targeted to come online in the third quarter of this year. By our cutoff, the quarter had ended, and there was no public report that it had been switched on. Southeast Asia shows how fast a new market can grow, and how quickly it can fill. In Johor, the Malaysian state just across the water from Singapore, the property firm Knight Frank puts life capacity at more than 1 gigawatt, with less than 1% of rentable space empty. One developer there day one, says in fighting to go public, that the capacity it has in service, across Johor and nearby Bandam in Indonesia, grew nearly 20-fold in less than three years. That's one company's disclosed portfolio, not the whole market. But it shows that Thai supply is not only an American story, and China measures data center capacity in a completely different way. Its government reports computing power, meaning how many calculations its AI facilities can perform rather than electricity. By that measure, its capacity nearly tripled in the year to June 2026. It's a striking figure, but the count of calculation cannot be converted into the gigawatts used everywhere else. Put together, these sources tell us where activity is concentrated, and where it's growing. What they don't support is a comparable global total for installed AI capacity. They measure different kinds of facilities, at different stages, in different units, and some of them overlap. So what has actually been built? A substantial amount, and you can find it on a map. But the biggest campus is open building by building, and the power building is not the same as a fully running campus. Behind what's operating since the pipeline, several times larger. That's where most announcements live. So headline numbers tell you more about ambition than about the computing available today. And the map is also shifting. The big cloud build has become a little more American, and it's moving inland toward land and power. Outside the US, it's growing on very different timetables, from fast-filling Johor, to a Middle Eastern pipeline that is still mostly plants. Rental market with almost no empty space. And construction that's already spoken for, suggests that customers want this capacity. But that leaves a bigger question. Who is actually paying for the computing? And how much of that demand is revenue today, rather than a promise about tomorrow? Think about the AI tools you have used this year. Maybe TryGBT, Cloud, maybe Gemini on your phone, maybe CoPilot inside the software you use at work. OpenAI said in February that TryGBT had 900 million people using it every week. Google said in August that its Gemini app had passed 1 billion monthly users. Those are enormous numbers, but most of that use is free. So if you want to know whether all these data centers are needed, the real question isn't how many people use AI. It's who is paying for the computing and how much. Here's a simple way to picture where the money goes. You use an AI product. Either you pay for it, or a business pays for it, or an advertiser does. The company behind the product pays for the computing, often by renting it from a cloud company. The cloud company runs that computing in data centers it owns, or leases. And the data centers need chips, power, buildings, and financing from yet more companies. That chain helps us understand where the money goes. But we can't trace every dollar through it. Public sources don't let us follow a particular dollar from your subscription to a chip factory. And there's no reliable formula that turns a number of users, or a number of questions asked, into the number of gigawatts the world needs. What we can do is look at the different kinds of evidence along the way. Each one answers a different question, and they are not interchangeable. Usage, like most user counts, answers one question. Are people using the product? It doesn't tell you whether anyone pays. Paid adoption answers another. Has someone bought access? Microsoft says more than 30 million seats of its co-pilot assistant for office software have been paid for, up from 20 million a quarter earlier. That's a seat sold, not necessarily a seat used every day. But somebody signed a purchase order. On the business side more broadly, the US Census Bureau survey of companies found that roughly one in five American firms say to use AI, and among companies with 250 or more employees, it's more than a third. Revenue answers the third question. What has a company actually earned by delivering its products or services? Here, the clearest figures come from the cloud companies, because they ran out of computing. In their most recent quarters, Google Cloud's revenue grew 82% from a year earlier, and Amazon's and Microsoft's cloud businesses each grew around 40%. Those are huge growth rates for businesses of that size. But they are whole cloud businesses, which also sell storage, database, and ordinary computing. None of the three says how much of that growth comes from AI. To see AI-specific money, you usually have to accept a different kind of number: the run rate. It answers the question, what annual pace is the company describing right now? Imagine a bakery that has a great month, and sells $10,000 of bread. Multiply that by 12, and the owner can say the bakery is running at $120,000 a year. That's a fair description of its current pace. It is not what the bakery earned over the past 12 months, and it is not a promise about next year either. Much of the AI industry's headline revenue is reported this way: Microsoft said in April that its AI business had passed a $37 billion annual run rate, more than double a year earlier. Amazon says its AWS AI business has passed $25 billion. Anthropics said in May that its run rate revenue had crossed $47 billion. More recent figures for the AI labs come from the press rather than the companies on the record. Bloomberg, citing people familiar with the matter, reported that Anthropics annualized revenue had reached about $65 billion by the end of July. Reuters reported, citing one source, that OpenAI was approaching $70 billion. These numbers point the same way: faster growth in paid AI use. But the run rate and the revenue accompanied reports for a full year measure different things. From businesses growing this quickly, a year's revenue adds up on months when the pace was much lower. So the two figures can be far apart without either one being run. The useful habit is simply to know which one you are hearing. The backlog answers a different question again: what future delivery has already been contracted? When a customer signs a multi-year contract, the seller can't count it as revenue until it delivers the service. The value still has to be delivered, sits in what the accountants call remaining performance obligations, or more simply backlog. It is not cash in the bank. Some of it may already have been paid in advance. And much of it won't turn into revenue for years. Microsoft's commercial backlog was $678 billion at the end of June, up 84% a year. That's a very large number. And much of it reflects real contracted demand. But two details change how to read it. Microsoft expects only about 30% of it to become revenue over the next 12 month. And excluding OpenAI, the backlog grew 25%, not 84%. In other words, a large share of the jump came from one customer. Oracle's backlog is similar in size and expects only about an eighth of it to become revenue within the next year. These are real contracts. There are also commitments that run years into the future. And a lot depends on a handful of AI companies being able to pay for them. Those AI companies are signing up for physical capacity too. OpenAI says it has secured about 10 gigawatts of AI infrastructure. Secured means contracted, not running. Its finance chief said it ended 2025 with about 1.9 gigawatts of computing capacity. But the two figures use different definitions. So it would be a mistake to divide one by the other. So far, all of this is what the companies tell us. The landlords offer a different view. Remember the Northern Virginia vacancy rate from the last chapter. The property firm CBRE tracks the eight main rental markets in North America the same way. In the first half of 2026, the amount of data center space in those markets grew by about a third in a single year. Normally, that much new supply would leave more space empty. Instead, vacancy fell to 1.4%, a record low, and tenants had already committed to about four-fifths of the space under construction, signing up before the buildings were finished. That's a strong evidence of demand for space. Two things keep it in proportion. These are rental markets, which leave out the campuses the cloud giants built for themselves. And the vacancy measures the empty buildings, not the idle chips. A fully-leased hall can still hold computers that sit underused. Prices offer one more clue. NVIDIA's H100 is the previous generation of the data center chip. Beautiful training AI models and running them. You would expect rental prices for an older chip to drift down as newer ones arrive. And for a while they did. Then, between October 2025 and March 2026, one widely watched index of one-year H100 rental contracts rose almost 40%. That's a single index for a single chip. But it suggests that demand for computing hasn't simply moved on to the newest hardware. But there's another side to the story. Plans do change. In the spring, OpenAI reported pulled out of a planned campus in Northern Norway, and Microsoft took the capacity instead. Separately, Bloomberg reported in March that a never-finalized expansion of about 600 megawatts at Abilene appears to have been cancelled. The original campus wasn't affected. One reading is that customers are switching providers rather than drinking. But that's an interpretation. But what would genuine weakness look like? Space left unleashed. Local rents falling. Big buyers cutting their spending plans. Any one of those can move first. So they are worth watching together. Among the cases reviewed for this video, none show those signs together. There are also things public data simply cannot show. No source we found gives comparable figures for how busy installed AI chips actually are. And demand is concentrated. In CoreWeave's second quarter, two customers, which it doesn't name, provided 62% of its revenue. The AI labs at the center of this demand don't yet publish audited accounts, so their revenue figures still rest on company statements and press reports. So who needs all this computing? Hundreds of millions of people use AI tools, a growing number of businesses pay for them, and cloud revenue is rising fast. Rental space is being taken almost as quickly as it's built, and some prices for computing has risen. That's real demand. But much of the AI-specific money arrives at a pace or a contract for the future, rather than revenue already earned. And the large share of those contracts depends on a few AI companies, which raises an obvious question. If customers are signing up for buildings before they are even finished, why doesn't the industry simply build faster? Let's try a thought experiment. Imagine a cloud company decides today to add a new AI data center of 100 MW. This is a made-up project, not a real one. Give it every advantage, customers ready to sign, and money to pay for it. What stands between that decision and computers that are actually running? The answer is a list of things it has to get from several different industries, each with its own queue. I'll take them one at a time. But in practice, many of these tracks run in parallel. The site can open only when they come together. Start with the chips. The computers in an AI data center are built around accelerators. Chips designed to do particular kind of math AI models run on. Far faster than an ordinary processor. NVIDIA's graphics processors are the best known example. Demand for them is visible in NVIDIA's own results. In its quarter to late July, its data center revenue more than doubled from a year earlier to $89 billion. And its finance chief said in August that the company expects supply to remain a bottleneck, at least through January 2028. That's a statement on earnings call, not a measured backlog. But it tells you how the main supplier sees his own position. For our hypothetical project, the harder problem isn't just whether chips are available. It's which chips. And how far along they really are. A new chip goes through stages. It's announced. It ships. Customers test it. It runs real work for customers. And eventually it's everywhere. Suppliers and customers often describe those stages in different terms. NVIDIA's current workhorse, the GB300 system, built for both training AI models and running them, has been running real customer work at scale since last year. Its successor, Vary Rubin, is designed for training and for running AI agents that carry out multi-step tasks. NVIDIA says it's ramping into full production with Rex running AI partners. The customers' own public statements describe different milestones on different dates. In mid-March, Microsoft said it had brought up a system for validation. In June, CoreVive said it had completed its system-level validation and that its system was operational. In July, Nibius said it was validating its first rack, with production testing still to come, before the system reached its customer. Those statements were made at different times about different steps, and they cannot be lined up into a single status for October. An operational system is also not the same as computing that customers can use at scale. And none of these statements tell us how many Rubin systems were serving customers by our cutoff. Later progress may simply not have been described in the sources we reviewed, so these statements can now rule it out. There's a second path too. All four of the biggest American buyers now design chips of their own, each built for particular jobs, some for training and running models, some mainly for running them. And Metas is for ranking and recommendations that decide what shows up in your feed. Together with Nvidia's GP300, those are the five platforms assessed for this video with evidence of named operating deployments. How many of each are running and on world work is mostly undisclosed. Custom chips give the big buyers another source of supply, but they don't escape the next problem, because every one of these chips depends on the same few kinds of components. The most important is memory. An AI chip is like a fast cook who is only as quick as ingredients arriving at the counter. If data cannot get to the processor fast enough, the processor waits. So AI chips use high bandwidth memory. Memory chips stack into small towers and place beside the processor, so the data has only a short chip. Putting the processor and those memory towers together on one base is its own manufacturing step, which is called advanced packaging. Think of it as a circuit board at a microscopic scale. Taiwan's TSMC does most of it. Customers have already reserved much of both. In July, TSMC's chief executive said its packaging capacity is so tight that it limits its customers' growth. At the end of September, Micron said it had agreed the vast majority of its 2027 high bandwidth memory supply, with significant price increases, and that it couldn't yet see when memory supply and demand would come back into balance. NVIDIA's own commitments to buy supplies, mainly memory, more than doubled in a single quarter. Those are commitments, not purchases already made. You can see the effect in the buyers' budgets. Amazon raised its 2026 spending plan, citing higher memory costs. Part of the spending increase, in other words, buy the same equipment at a higher price. But not more of it. The chips also have to work as a team. Training or running a large model split the work across thousands of chips, and they have to swap results constantly. So the network connecting them matters almost as much as the chips. Two technologies compete to connect the racks: Ethernet, the same family of technology as the network in your office, and InfiniBand, a specialist alternative mainly supplied by NVIDIA. Ethernet has been taking a larger share of AI network switch sales, yet InfiniBand sales are growing too, which tells you how fast clusters are being built. In the sources reviewed, the network shows up as a necessary input that is growing quickly, not as a documented supply shortage. The network also explains the next challenge. Chips that talk to each other constantly work best packed close together, pack them close, and heat concentrate in a small space. A rack is a single cabinet of these computers, and Microsoft's Fairwater site in Atlanta is designed to support about 140kW in each one. There's design capability at one site, not what its rack actually draw, and also not an industry average. But at that design figure, a megawatt of power for the computers alone would run only about 7 such racks, and a building needs more power on top of that, for cooling and everything else. At densities like that, blowing cold air through the room isn't the approach these designs rely on. Fair water is cooled by a closed loop of liquid, and at densities like this, coolant is pipe to plates sitting directly on the chips, carrying heat away in a way air alone cannot. For our hypothetical project, that means cooling has to be designed in from the start, with its own equipment to order. Which brings us to the biggest physical input of all: electricity. Our 100MW building needs a connection to the grid, or its own power plant, or both. And here the queues are very long. Remember the pause in taxes from the first chapter? By June, the grid operator there had received requests to connect about 474GW of new large users, about 90% of them data centers. Requests are not the same as real projects. They measure intent, and some are duplicates. Even so, those requests are far greater than the grid can accommodate in the near term. In August, the same operator stopped approving new large data centers to switch on, pending reports due in December. Even a project with a grid connection needs heavy electrical equipment. Take the large transformers that link power plants to the grid. Reported waiting time for them exceeded 160 weeks in the first quarter of this year, more than three years. That's a reported lead time, not a major delivery. But a three-year wait for a single critical part changes how far ahead a project has to plan. Some campuses bring their own power, as we saw with Meta's Hyperion and its new gas plants. That means waiting for power equipment instead of a grid connection. G. Vernova, one of the biggest makers of gas turbines, had 53GW of firm order in its gas power backlog at the end of June, plus another 63GW of slot reservations, a place in land rather than a confirmed order. It's now taking reservations for turbines to be delivered in 2031. Construction runs alongside all of this. And here, Abilene shows how a campus actually arrives. Its first buildings were energized within a year of the start of the work. More than two years in, six of eight were delivered, and in September, Oracle gave a new date for the rest. At the company level, the same pattern shows up as a gap between what's been signed and what's switched on. At the end of June, Corvive, a company that rents out AI computing, reported about 1.5GW of active power and about 3.7GW contracted. It doesn't define either measure precisely, but the direction is clear. Its contracts are running well ahead of its power capacity, and delays now show up in financial results. Corvive cut its 2025 revenue forecast last November, blaming a decent developer that was behind schedule. Supermacro, which builds AI servers, said its June quarter revenue landed near the low end of its forecast, because of short-term customer delays in power, cooling, and networking. So why is building more so difficult? Not because of a single bottleneck. Our hypothetical project needs everything we have just walked through, and several of them are documented as tight right now. Chips allocated in advance, memory and packaging sold ahead, electrical equipment with multi-year weights, grid connection in a lengthening queue, and campuses delivered one building at a time. Others, like the network and the cooling system, are essential pieces that simply have to arrive on schedule. The evidence doesn't show which constraint binds hardest. It shows several of them tight at the same time, and the site cannot open until every track is ready. And every one of those checks has to be paid for, often years before the building earns anything. So who is paying? Over the 12 months through June, Amazon spent about 173 billion dollars on property and equipment. In the same 12 months, its businesses brought in about 161 billion dollars of cash from operations. So where did the rest come from? And how are the other giants paying for their share of the build out? The first answer is the simplest. The biggest buyers of AI infrastructure are some of the most profitable companies in the world, and most of what they spend come from the cash their existing businesses generate. Accountants call that operating cash flow, the money that comes in from selling things and running the business before the company invests any of it. Think of it like a household's take home pay. It's the money available before you decide whether to buy a car or renovate the kitchen. The analogy only goes so far. Because these companies can also borrow, sell shares, or bring in partners on scale no household can. But it's the right starting point. For the four largest American buyers, Microsoft, Alphabet, Amazon and Meta, operating cash flow is rising fast, between 27 and 39% over the year to June. Spending on buildings and equipment rose much faster, between 61 and 98%. Those two facts together are easy to misread, so it's worth separating them. Spending grows faster than the cash doesn't mean spending is bigger than cash. At Microsoft, Alphabet and Meta, purchases of property and equipment took roughly two-thirds of operating cash flow. That's a much bigger slice than a year earlier. But there was still cash left over. Amazon is the one that crossed the line, at 107%. So the difference had to come from somewhere else. One more thing to keep in mind. These are company-wide figures. They include offices, warehouses, devices, and ordinary cloud computing. None of this company says how much of its spending is for AI. What's left over is usually called free cash flow. And each company defines it a little differently. So the companion report uses one common definition for all four. Operating cash flow, minus purchases of property and equipment, minus repayments on a kind of lease that works much like a loan, called a finance lease. On that basis, free cash flow fell at every one of the four over the year. And at Amazon, it turned negative. So the second source of money is borrowing. Over the year to June, Amazon, Meta, and Alphabet each raised tens of billions of dollars in long-term debt. Microsoft didn't issue any. The third is selling shares. In June, Alphabet raised about 50 billion dollars of new equity. Money is interchangeable. So you can't tie any of these resources to a particular data center. Alphabet said its purposes included capital spending to scale its AI infrastructure. The fourth source is the one that's easiest to miss. Because it doesn't show up as spending yet. Leases. Instead of buying a building, a company can rent it. Often under a contract lasting 15 years or more. The landlord pays to build it. And the tenant pays rent once it moves in. Until the lease starts, the obligation sits in the footnote. Not on the balance sheet. And not in the capital spending figure investors watch. Add up the leases that five companies, the big four plus Oracle, have signed but not yet started. Mostly for data centers. And you get about 1.12 trillion dollars. That's not the present value. It's the total of future payments over terms of up to 30 years. Starting anywhere from this year to 2036. And it shouldn't be added to capital spending. Which counts something different. What it shows is how much of the coming year's expansion is already committed. Before any of it appears in this year's spending. That's why future obligations can grow much faster than the spending figures suggest. A company can sign a lease today for a building that starts paying rent in 2029. The commitment is real low. The spending shows up later. Leases also explain a confusing headline from Microsoft. Its 2026 capital spending figure came down to about 175 billion dollars. It was not a cut. Microsoft changed the accounting life of its data centers and office buildings from 15 years to 25. That shifts more of its future data center leases into a category that doesn't count toward its capital spending figure. Its finance chief said that, apart from this accounting change, its spending expectations for the year were unchanged. The fifth source is outside investors, who put their own money into the buildings. Alongside lenders who finance them. Meta's Hyperion campus in Louisiana is the clearest example. A joint venture owns the campus. Funds managed by the investment firm Blue Owl hold 80% of that venture. And Meta holds 20%. And the project's development cost is about 27 billion dollars. Part of Blue Owl's capital comes from bond sold to large investors, such as PIMCO. So bond investors finance part of the project, while the joint venture owns the campus. Meta's role is different again. It's a tenant. It has signed leases to use the campus, with an initial term of four years, and options to extend. And it has given the joint venture a capped guarantee on the campus value for the first 16 years of operation. The joint venture finances and own the campus, with Meta participating as both investor and tenant. Meta still has obligations, lease payments, and a capped guarantee that could require a payment if a lease ends or is not renewed, and specified conditions are met. NVIDIA has announced something bigger in Ambition, a framework with six large investment firms, including Apollo, BlackRock, and KKR, for over 500 billion dollars of outside capital to fund computing. It's still subject to final agreements. So it's a framework, not money committed. At the other end of the scale are the specialist companies that rent out AI computing. And their story is different. They don't have a giant advertising or software business throwing off cash. In the first half of 2026, CoreWeave spent nearly four times as much on equipment as its operations brought in, and it carried about 35 billion dollars of debt at the end of June. In August, it added a new 2.6 billion dollar loan at a floating benchmark rate plus 5.5 percentage points. The loan is tied to specific customer contracts, and we'll come back to what that means for risk. Oracle sees in between. In its latest quarter, it spent more on capital projects than its operations brought in, even though that cash included large prepayments from customers. So who pays for the expansion? Mostly the cash from businesses that were already very profitable. And that cash is still growing quickly. But spending is growing faster. So more of the build is being funded with debt, new shares, long leases, and outside investor. Much of the commitment sits in contracts that hasn't started yet. And the risk is being spread to lenders, landlords, and private funds, sometimes with guarantees that spend part of it back. None of that is a sign of trouble on its own. Growing cash and growing funding needs can coexist. Whether it works out depends on what all this spending earns, and that's not the same for every company in the chain. The money flows out to chip makers, server builders, and landlords. Who keeps it as profit, and what has to go right for the buyers to earn it back? Picture a single purchase. A company that rents out AI computing and buys a rack of NVIDIA servers. Suppliers earn revenue from selling equipment and components. The operator then has to earn back its investment by renting out computing over time, hour by hour, for years, before it knows whether the purchase paid off. That difference in timing is the key to this chapter. It explains how one part of the industry can be extremely profitable today, while another part is still waiting to find out. Let's follow the purchase down the chain, and look at how much each kind of company keeps. The measure is operating margin. Out of every dollar of sales, how much is left after the causes of running the business, before interest and taxes. At the top are the makers of the most sought-after parts. Its latest quarter, Micron, the memory maker, kept about 80 cents of every sales dollar as operating profit. NVIDIA capped about two-thirds. Further down, several builders such as Dell ran on much thinner margins. Data center landlords such as Equinex and Digital Realty capped about a quarter. And the two specialist AI cloud companies revealed, Coreweave and Nibius, reported operating losses. A small one at Coreweave, a larger one at Nibius. It's tempting to conclude that the profits go to whoever controls the scarcest part. That's one reasonable interpretation, since the highest margins sit where suppliers describe sold-out capacity and rising prices. But a margin comparison cannot prove it. This company is different in what they sell, in pricing power, in accounting choices, and in when their equipment costs hit the books. And these are company-wide margins, not returns on any AI data center. What the comparison does show is that a supplier's profit tells you very little about its customers. NVIDIA's margins come from selling its products. Its customer's return depends on everything that happens after the purchase. To see what that looks like, take Coreweave, which publishes enough detail to follow the money through a single quarter. Ask whether Coreweave made money in the second quarter of 2026. And the honest answer is, it depends which question you mean. If the question is how much its operations earned before the cost of its equipment wearing out, the answer is a lot. Coreweave reports a measure called Adjusted EBITDA, which it defines itself and which adds back what wear and tear cost along with some others. It was about 1.5 billion dollars, nearly 59% of revenue. If the question is whether the business covered all of its operating costs, including the wear and tear on its computers, the answer is close to breakeven, but just short. Depreciation, which is the counting charge that spreads the cost of equipment over its expected life, came to almost 1.4 billion dollars that quarter. Add stock-based pay and a few smaller items, and the result was an operating loss of about 49 million dollars. If the question is whether it made a profit after paying its lenders, the answer is no. Net interest expense was about 640 million dollars. After that, some other income and taxes, Coreweave reported a net loss of about 626 million dollars. And if the question is cash, it's different again. Its operations brought in about 680 million dollars of cash that quarter, while it spent about 6.4 billion dollars on new equipment. None of those four answers is wrong. They measure different things. The one that matters most for the long run is whether the equipment, over its whole life, earns back what it costs plus a return. And no quarterly figure answers that. So let's try a model. This is a hypothetical example built for the companion report. Not any company's actual numbers. And not a forecast. It just shows which assumptions matter. Imagine buying one GPU of the GB200 class from NVIDIA's Blackwell generation, used for training and running AI models. That's the configuration used in this example. It costs $55,000, everything included, paid upfront. You rent it out for $3.5 an hour, and bill customers for 85% of the hours in a year. You pay for space, electricity, and other running costs. Revenue starts slower, while the hardware ramps up. After 3 years, the contract renews at price 30% lower. And after 5 years, you sell the used chip for 10% of what you paid. Under those assumptions, the investment returns about 9.4% a year, before borrowing costs and taxes. And that's the model's result, not any company's. But one caution about that starting price. The $3.5 an hour is assumption. It's above the recent readings from the two recent price index for NVIDIA's older H100 chips. Though neither of those indexes is direct price for this configuration. Now let's change one assumption at a time. Rent it for $3. If you rent it for $3 an hour, instead of $3.50, the return falls to under 1%. Rent it for $4. And it rises to about 17.6%. $0.50 an hour, in either direction. Move the return by roughly 80% to 90% point. Utilization works the same way. Revenue is the price times the hours you actually bill. So bidding fewer hours lowers revenue, just as the lower price does. Useful earning life matters too. If the chip loses half a year of earning life, the return drops to about 5.9%. Timing matters as well. Though in this model, bet less. Delayed start by 6 months, with the chip's full earning life intact, and no cost is while it waits. And the return slips to about 8%. But if you are paying rent and other fixed costs while the hardware stays idle, the same 6 months delay pull it back to about 6%. Under this particular set of assumptions, price and earning life move the result more than a short delay. That doesn't tell us which risk matters most in the real world. Different assumptions would change the gaps. And whether any of these returns is enough to compensate for the risk is a separate question. What the model shows is how sensitive a buyer's return is to things nobody can observe yet. The landlord's version of this calculation is different. In a similar hypothetical model of a 100MW building, leased for 15 years, what matters most is what the building is worth when the lease ends. If it can be sold at a price that assumes it keeps earning rent, the return is about 10% a year. If it's worth only 40% of its cost, the return falls below 6%. A one-year construction delay costs less than a point either way, which brings us to the hardest assumption of all. How long these chips keep earning? Companies choose an accounting life for their servers. And that choice shapes reported profit. Amazon's chief executive has said its servers have a useful life of at least 5-6 years. Yet from January 2025, Amazon shortened the accounting life of some of its servers and networking equipment from 6 years to 5. Meta went the other way, extending most of the servers to 5.5 years. An accounting life is an estimate. How long a chip actually keeps earning depends on what customers will pay for it when its first contracts end. And for the newest generation of chips, those renewals haven't happened yet. If returns are uncertain, the next question is who carries that uncertainty. Contracts and financing decide that, and they move risk around rather than removing it. Take CoreView's new $2.6 billion loan. It runs for about 5 years. CoreView says the customer contracts behind it average about 3 years. If those customers renew at similar prices, that gap doesn't matter. If they don't, the loan outlasts the contracts that were meant to pay for it. On top of that, CoreView depends on just a few customers for much of its revenue. And you can see where the risk sits. Or take Nvidia. In August, it agreed to guarantee part of the value of data center leases between SB Energy, the landlord, and OpenAI, the tenant, covering about 4.25GW. Nvidia isn't guaranteeing OpenAI's rent. It's promising that, if OpenAI defaulted, Nvidia would cover the gap between a guaranteed value and what the landlord could recover. Up to a cap of $105 billion. With $3.5 billion of other guarantees, its maximum exposure stood at $108.5 billion as of its August 26th disclosure. The guarantees take effect in phases from 2028 as a lease start, and exposure shrinks as OpenAI pays. That's a maximum, triggered only by default, not an expected loss. But it shows that some of the risks in this buildout can travel back up the chain, even to a supplier with one of the highest margins in the industry. So who is making money? Right now, the money and chip suppliers report the highest margins, and landlords earn steady ones. The specialist companies renting auto-computing report operating losses, and at Corvive, that loss sits alongside strong earnings before depreciation. Will the investment pay off for the buyers? That depends on rental prices holding up, on chips staying busy, on how long they keep earning, on renting costs, and on getting capacity switched on without long idle periods. Today's result can settle that, because the most important inputs, especially renewal prices for the newest chips, hasn't been observed yet. Which means the most useful thing we can do is decide what to watch. We have now followed the money from the people using AI, through the cloud companies and their suppliers, to the lenders and investors behind them. Before putting it all together, there's one more question worth asking. What could change this picture over the next year? It helps to think in terms of three broad outcomes. Continual scarcity, where demand keeps pace with new supply. Digestion, where supply catches up and prices soften. And financial stress somewhere in the chain. These are not forecasts, and they are not mutually exclusive. The same year could bring tight conditions in power-limited markets, alongside falling rental prices for all their chips. The practical question is which evidence would tell them apart. Here are five things to watch. Some are measurements at the quarter's end. Others are dates when reports are due. Both are worth keeping straight. The first is whether rental markets stay tight as new space arrives. Around 7.5GW were under construction in North America's main rental market in the first half of this year. About four-fifths of it already committed. If vacancy stays near its lows, and new projects keep signing tenants before they open, demand is still keeping up. If more space is still empty, and less capacity in new projects is leased before completion, that would be the first market sign of supply catching up. That tells you about demand for space. It doesn't tell you much about returns. The second is rental prices for computing, especially how much customers will pay to keep renting the current generation of NVIDIA chips when their first contracts end. The baseline is the H100 contract index at $2.35 an hour in March. Prices holding or rising across several indexes would mean demand is keeping pace with new hardware. Comparable indexes falling together would point toward digestion. And as the model showed, it would hit owner's return directly. The third is whether contracts turn into renting capacity and revenue. We'll see it in backlogs being recognized as revenue, on the schedules companies have given. In CoreWeave's actual power closing in on what it has under contract, in Abilene's last buildings being delivered, and in customers describing their Rubin systems in comparable terms. What work they're running, and at what scale. AMD's first customer deployment of his new systems belong here too. If delivery keeps pace, that's a constraint easing. If handovers slip, while contracts keep growing. That's the constraint biting harder. Which is not the same thing as weak demand. The fourth is the grid and equipment behind it. Texas Grid Operator is due to publish its reports on the pause by December 10th. After that, the questions are whether approval for large data centers resume, whether waiting times for transformers start to shorten, and whether GE Vernova's reservations turn into firm orders. That tells you about delivery, not about returns. The fifth is cash against commitments. Watch how much of their operating cash the largest buyers keep putting into buildings and equipment. Whether those signed leases start on schedule. Whether the GPU Cloud's operating and net results improve as their fleas mature. What new borrowing costs, and whether contingent promises like Nvidia's guarantees, which start taking effect in 2028, grow. Cash keeping pace, and GPU Cloud results improving, would mean the financial strain is easing. Tighter financing, delay the lease starts, or guarantees and long term commitments being renegotiated, would mean it's growing, and would raise the separate question of who absorbs the losses. No single one of those signals settle anything. The assessment should change when several of them move the same way. So as of October 5th, here's where things stand. How much has been built? A lot. And it's real. But it arrives building by building. A powered building is not the same as a running campus. And the plans are several times larger than what's operating. The build has become a little more American. And it's moving inland toward land and power. Who needs it? Rental space is being taken almost as fast as it's finished. Tenants sign up before buildings are done. And paid AI use and cloud revenue are growing fast. But we cannot see how busy the trips are. Much of the AI specific money is still a pace or a contract for the future rather than revenue already earned. And it's concentrated in a few customers. Why not just build more? Because delivery is held back at several stages at once. Chips, memory and packaging sold ahead. Electrical equipment with multi-year waits. Grid connections. And campuses that open one building at a time. The evidence doesn't single out one bottleneck. And will the companies paying for all of this earn their money back? The money is lending unevenly. Memory and chip suppliers report the highest margin today. The specialist companies, renting out computing, report operating losses. For the buyers, the honest part is that it isn't decided yet. It depends on what customers pay when contracts renew, on how long the chips keep earning, on how busy they stay, and on how quickly capacity comes online. The next year of evidence will tell us a good deal more. The full report can be downloaded in the description. If you enjoyed this video and want to see more deep dive like this, please subscribe, like this video, and turn on the little bell icon. And if you want to watch new videos in advance, you can join the channel as a member. Members can also propose research topic for the next videos. Researcher member can also download infographics from the videos, as well as watching ad-free version of every video. Thank you so much for watching. I'll see you in the next one.