Latest / The AI Supercycle / AI Data Centres: The $30 Billion Price Tag
Transcript
- [0:00] $30 billion. That is the reported cost of building and equipping a 1 gigawatt AI data centre in Alabin, Texas. Texas Data Centre Economics [0:09] And more than half of it goes on chips. Last week, we followed the money into nuclear power, cloud commitments and the diesel needed to get the physical build moving. This week, I want to walk through the data centre that all of that activity is supposed to produce. Because once you put a price on it, the questions become rather more demanding. How much does it cost to finance? When can it start earning? And what happens if the price you can charge for the computing inside it falls faster than you expected? Those questions have acquired a sharper edge with fresh competition from China. Microsoft's plans to deploy its own AI chips at scale, and chip industry leaders asking where the next round of capital will come from. The demand for AI can be substantial, and the economics can still be difficult. This week, we are going to examine how both can be true. AI Supercycle Overview [1:14] Hello and welcome to the AI Supercycle podcast from QF Markets, where we follow the capital flows, physical constraints and commercial returns shaping the largest industrial build-out of our generation. From semiconductors and data centres, through power, energy and critical materials, all the way into orbit, and onto the enterprise workflows that ultimately have to pay for it. I'm Tim Hardwick, founder of QF Markets, and I'm recording this on Tuesday the 13th of October, 2026. In today's episode, we're covering four things. The macro regime, the AI supercycle itself, natural resources, and geopolitics. Then I'll bring it all together with my framework's read on what it all means for the supercycle and what matters next. [2:02] Our starting point is one very expensive building in Texas. By the end, we'll have followed the money all the way from laying the foundations to the customers who ultimately have to pay for it. The theme running through my weekly outlook this week, which I put together every Monday, is the growing separation between different parts of the AI economy. When I sat down with the markets on Monday morning, the picture was mixed. Some of the big technology platforms were holding up well, Semiconductor and memory stocks were under pressure. And energy was stronger. So asking whether the market likes AI today isn't a very useful question. A better one is, which parts of it and why? These companies earn their money in different ways. They borrow different amounts, and they wait for very different lengths of time before they get paid. A software business with an established customer base can sell more through systems it already runs. A company building a new data centre has to spend heavily before it serves a single customer. Some of the giants do both, of course, so we have to look at what's going on inside the company, not just at the name on the share certificate. [3:16] Higher financing costs make those differences more important. Financing the Build Let's put a number on it using our $30 billion example. Suppose, purely for illustration, that 60% of the investment is financed with debt. That would mean $18 billion of borrowing. Once all of that is drawn, every extra percentage point on the interest rate adds $180 million to the interest bill. Think about that for a moment. That is a huge sum of money. I should be clear that that's a number I've come up with, as an example, not the actual financing of the Alabin site, which hasn't been disclosed. And a borrower who has locked in a fixed rate doesn't suddenly pay more because the market rates go up. It depends on when the debt is raised, how it's structured, and whether it has to be refinanced. But it shows why the price of money belongs in a conversation about chips. It also shows why delays hurt so much. The money may already be spent and the interest already ticking, while the revenue is still waiting for a power connection, a delivery of equipment, or a customer to sign off. That brings us to this week's US inflation figures. [4:31] Energy can push the headline around, so the Fed will be watching core inflation to judge whether price pressures are spreading and sticking. No single print decides its next move. For us, the question is simpler. Are the financing assumptions behind these long-lived projects getting easier or harder to meet? The bank earning season adds another angle. I'll be listening to what lenders say about credit quality, how carefully they're lending, and how keen they are to finance infrastructure. A strong quarter for a bank doesn't mean every AI project can borrow on good terms. The real evidence is in the terms they offer, and the risks they're prepared to keep on their own books. What $30 Billion Buys [5:14] So let's go back to Alabeam. Bloomberg's reported estimate, put together with the property firm Cushman & Wakefield, puts the costs of building and equipping a 1 gigawatt AI data centre there at around $30 billion. Of that, $17 billion goes on chips, $3 billion on power infrastructure, and $2.4 billion on the building itself and its construction. That leaves $7.6 billion that the published summary doesn't break down, and I'm not going to guess what's in it. One more boundary. This is a cost estimate for 1 gigawatt of capacity. It isn't the audited bill for every phase of the wider campus, and those are different things. Even so, the split is revealing. More than half the money is in the chips. The building and its construction come to just 8%. So when we hear data centre, most of us picture property, land, walls, a roof and a tenant. But an equipped AI facility is mostly a very large investment in computing equipment, and that equipment has its own economics and its own replacement cycle. [6:27] It's a bit like a Formula One team. The factory in the wind tunnel served the team for decades. The car is a different story. Last season's car still runs perfectly well, but put it on the grid with this year's cars, with their new aerodynamics and the latest upgrades, and it's off the pace. The building is the factory. The chips are the car. In practice, if a competitor can do the same job using less electricity, or get more done in the same time, your machines still work perfectly well. You may just have to charge less for them, move them onto different work, or spend more to keep up. That's why I want to know how quickly the equipment earns back what it costs. Not just how long the accountants plan to depreciate it. A depreciation schedule doesn't guarantee customers will pay enough to get the money back. Who owns what matters too. The owner of the building may collect rent from a cloud operator. That operator may own the servers and sell computing power to an AI company. And the AI company then has to persuade businesses and consumers to pay for its service. [7:42] Those are three different streams of cash, and the contracts decide how well each one is protected if demand disappoints, prices fall, or the site opens late. A long lease gives the landlord security, but that security is only as good as the tenant's ability to pay and any guarantees behind it. A big commitment from the AI customer deserves exactly the same scrutiny. So we need to follow the money all the way through. Whose money has been spent? Who has signed up to pay? And what has to go right for them to afford it? Capital and Competition [8:20] Mundi brought a timely example of that question from Singapore. At the DBS conference there, Hock Tan, the CEO of Broadcom, and Lit Bu Tan, the CEO of Intel, talked about the capital needed for the next phase of AI. Hock Tan flagged capital as a possible constraint. And we've talked about this in previous episodes as well. And he also questioned how widely companies' AI spending was yet paying its way. Now I think that's worth paying attention to, coming from people who sell into this boom. They're confident about demand, and they're also asking where the money comes from. We should listen to both halves. There was another development on Monday, this time about competition in memory chips. China's Changshin Memory Technologies, typically known as CXMT, has reported progress on a denser design for the memory cell, the tiny unit that stores each bit of data. Current reporting points to server memory products using it by the end of the year. Memory shares have come under pressure this week as investors have weighed up what that could mean for the competition. [9:31] Engineers call the design 4F squared, which to some folks might just sound like gobbledygook. But all it describes is how much space each memory cell takes up. Make the cell smaller and you can fit more memory into the same area. Getting a design into reliable production is another matter though. We still need evidence on yields, cost, customer approval, and how much can actually be shipped. And the established manufacturers are hardly standing still. We also need to be specific about which memory we mean. High bandwidth memory, or HBM, is the stacked memory that sits alongside advanced AI processors. How well it performs depends on much more than the size of a single cell. Including how the layers are connected and how data moves between the memory and the processor. So a denser server memory design isn't, on its own, a replacement for HBM. In all of the hullabaloo around a new chip announcement, it's tempting to leap from a promising design straight to a transformed market. The steps in between are where the commercial evidence has to show up. For an investor, the sensible question is big enough already. How fast is another supplier becoming competitive? And what could that do to prices and margins over time? Microsoft’s Custom Chips [10:53] The pressure to cut costs is also coming from the biggest buyers of chips. On Monday, Microsoft announced Maya 300, its next-generation AI processor with production deployments planned for 2027. It plans to scale its Maya systems to several gigawatts, mainly to run AI across its own products and services. And Reuters reports that could mean more than a million chips next year. Maya 300 is built for inference, which is the work of running models once they're trained, answering questions, writing code, carrying out tasks. When you're doing that at enormous scale, cutting the cost of each useful answer makes a big difference. Microsoft says Maya 300 will deliver 40% more tokens per megawatt than leading GPU-based systems. Tokens are the small chunks of text these models read and write, So that's a measure of output for each unit of power. It's Microsoft's own claim, though, against systems it doesn't name. It's a bit like a football club, announcing its new signing will score 30 goals a season. Wonderful if it happens, but we'll judge the new player once they've played a full season, against the same opposition as everyone else. [12:12] The data center rack figures need the same care. Microsoft puts 96 chips in a rack drawing 167 kilowatts. That tells you how much power to supply. It doesn't tell you how efficiently the rack gets a customer's work done. For that, you need the output as well as the input. There's an interesting link back to memory here. Microsoft's design uses HBM4, with computing chiplets made by TSMC. So even when you design your own processor, you still have to buy the advanced manufacturing, the memory, the packaging, and everything around it. My assessment is that this could shift where the profits land along the supply chain. Microsoft may get more control over its costs and more bargaining power. NVIDIA faces another route by which a major customer could supply some of its own computing needs. How big that effect turns out to be depends on deployment, performance, and how fast the overall market grows. And for our $30 billion example, there's an important caveat. 40% more output per megawatt wouldn't mean 40% off the cost of building and equipping a data centre. Financing, software, how busy the equipment is kept and the cost of replacing it all still matter. Cheaper AI, Bigger Pressure [13:37] The software side adds a different pressure. Lipbu Tan also talked about the rapid take-up of Chinese models such as DeepSeek, Kimi and GLM, pointing to their lower cost per token and to performance that customers find good enough. Now that's a good point, isn't it? Good enough. We don't need a Ferrari to go down to the shops for the weekly shopping, right? For the weekly groceries. Price per token matters, but what a business really cares about is the cost of getting a job done reliably. They have got to focus on outcomes. And recently we're hearing a lot more about outcome maxing. Cheaper models could bring more customers into the market and encourage existing ones to use more AI. They also make it harder for a provider to hold on to the price it's assumed in its investment plan. That's the tension I want to carry forward. The cost of the infrastructure is locked in over many years. The price of what it produces is exposed to competition the whole way through. Resources and Geopolitics [14:40] Okay, so let's move on now to part three, which talks about natural resources and geopolitics. Last week we finished with diesel because someone still has to deliver the transformer and run the machinery that builds the site. This week the diesel story has moved into diplomacy. Reporting on proposed US arrangements for Russian diesel shows how much political pressure there is to find more supply. The real test is how much usable fuel actually arrives, where it lands and what it does to prices. A political announcement can shift market expectations in an instant. However, the impact and physical benefits depends on refineries, ships, insurance and delivery. And that takes weeks or months. It's the two clocks we keep coming back to on this show. A fast financial clock and a much slower physical one. For a data centre project, fuel costs find their way into the bill by several routes. Haulage gets more expensive. Contractors pay more to run their machinery. Suppliers try to pass on some of their own increases. Who ends up absorbing that depends on the contract. A contractor on a fixed price may carry the pain at first. If the contract lets certain costs be passed through, more of it lands on the developer. And even where existing contracts can't be reopened, the next round of bids can be priced higher. [16:09] Then there's the electricity the site uses once it's running. The $3 billion for power infrastructure, in the Abilene estimate, is the cost of installing it, not a lifetime electricity bill. Building the connection and buying the power are two separate parts of the economics. And the most expensive kit isn't always what decides the opening date. For example, a pilot can have a full tank of fuel and a perfect aircraft, But if one item on the pre-flight checklist fails, nobody's taking off. In the same way, a late transformer or an unfinished grid connection can leave billions of dollars of chips sitting idle in the data centre. That's why simply adding up the biggest spending categories misses something. We need to know what has to arrive before everything else can work. What are the dependencies? [17:01] Chinese manufacturing progress could eventually ease some equipment shortages or add more competitive supply. Whether a given operator can actually use it depends on performance, support, security requirements and trade rules. There's no reason to assume a cheaper component available in one market is immediately available to every project everywhere else. For our research, the useful question is how all of this changes the cost and timing of getting computing capacity into service. That keeps the geopolitics tied to the project, rather than just leaving it as a big pile of separate headlines. The PRISM Framework [17:41] Okay, so let's move on now to the framework. This is where I pull the evidence together through PRISM, P-R-I-S-M. And that's my framework for assessing the market, forming a base case, and working out what could change it. And what it does is it adds a probability layer on top of my methodology, my trading and investing methodology. My working view is that financing and competition are making the AI more selective. Demand can stay strong while returns differ widely across the companies serving it. The data centre example shows why. A supplier can get paid well before the finished facility earns anything for its owner. The operator then has to keep the equipment busy, cover its running costs, and hold back enough cash for interest and for the next round of replacements. That doesn't make suppliers automatically good investments. They have their own costs, competitors and valuations. It just means that sharing in AI demand doesn't give every business the same financial result. [18:48] The CXMT news changes the questions we should be asking about memory competition. It doesn't yet show that memory prices or earnings are about to collapse. Before drawing that conclusion, I'd want to see customer approval, shipment volumes and real evidence of pressure on the prices manufacturers actually achieve. Now Microsoft adds another possibility that companies running AI services keep more of the benefit from better chip economics. Custom chips might help offset falling service prices while still supporting demand for memory, manufacturing and power. [19:26] I would judge that on the full cost of the finished work, including development and deployment, rather than on headline efficiency claims. Equally, weak share prices deserve attention even when the latest results look strong. Markets look ahead and sometimes spot a change before it shows up in earnings. They can also overreact to a headline, so I treat price action as a prompt to investigate more, not as the full explanation. On the customer side, lower AI prices can improve returns. A company that can get the same work done more cheaply may use more of it. Whether that means more revenue for the provider depends on how much usage grows and what it costs to serve. A simple example makes the point. If the price of a comparable service halves, usage has to double just to keep revenue where it was. Profit then depends on what it costs to deliver. Better technology may bring that cost down, but we have to look at both sides. So what would make me more positive? Evidence that financing stays available on workable terms. That sites come into service and fill up with paying customers. And that customers expand because the service is genuinely earning its keep. [20:44] And what would weaken my view? Well, that would be a number of things, such as projects slipping and interest costs piling up, customers scaling back their commitments, or service prices falling faster than operators can cut costs and grow useful demand. Those are the things that would change the economics. A bigger spending announcement, however much Ballyhoo comes with it, doesn't settle any of them. What to Watch Next [21:11] Several releases this week will test this argument. I'm recording before the inflation data and the big semiconductor results, so I'll frame them as questions rather than guess the answers. US consumer price inflation is due on Wednesday, the 14th of October, so I'll be looking at the underlying trend as well as the energy effect, and at how the bond market reads it. [21:35] ASML also reports on Wednesday, its orders and commentary will help us judge how willing the chipmakers are to invest in future manufacturing capacity. I'm interested in the timing and the mix of that demand, not just the headline number. TSMC follows on Thursday. The questions there are how management sees customer demand, its own investment in capacity, and the bottlenecks in delivering advanced chips. We need to hear that guidance before deciding what the recent weakness in semiconductor shares really means. The banks will fill in another part of the picture, whether lenders see good business they're happy to finance, or commitments that are getting harder to price and to sell on. For CXMT, the next milestone is evidence behind that product timetable. An announcement is where the assessment starts. Approved products and steady commercial shipments would tell us far more. [22:31] For Microsoft, the milestones run into next year. Maya systems coming into service, real workloads moving onto them, and evidence that savings hold up once you count the full cost of running them at scale. These are all different windows onto the same question. Can the system turn its commitments into cash as fast as its financing needs? Cash Versus Construction [22:54] Okay, so the $30 billion example makes the scale of the AI build easier to grasp. The split inside that number makes its risks easier to understand. A building, a power connection and a great deal of computing equipment all have to work together, and they have to produce a service that customers keep buying, on terms that leave enough money to support the investment. Competition from cheaper memory, custom processors or cheaper models could make AI more accessible. It could help some businesses earn better returns while squeezing others. That's why I keep coming back to the different layers of this economy. Now, the financial clock can reprice an expectation in minutes. The physical build takes years. In the gap between the two, companies have to manage contracts, costs and fast-changing technology, while investors decide how much future profit they're willing to pay for today. That's where the build meets the balance sheet. And that provides a nice segway into the close. And just to let you know that I am working on a data centre research report and a companion piece on what it costs to bring a gigawatt of AI capacity into service. Research and Wrap-Up [24:09] Now this episode lays down the questions that this research needs to answer. So I'll keep you updated on progress with those reports, and also with the forthcoming book Trading the AI Supercycle, which will now be three volumes because of the sheer size of the book, and having one large volume presents a number of challenges for publishing. So I'll go into a bit more detail on this next week. [24:35] Okay, so that's a wrap for this week. Thank you for listening. If you found this useful, please follow or subscribe wherever you listen and share it with anyone trying to work out where the capital in this cycle is actually going. You can find all of our work at qfmi.substack.com The Market Pulse and In The Spotlight articles are free, and always will be The Weekly Outlook, The Weekend Debrief and The Monthly Strategic Research Report sit behind a paid subscription, Subscribe and you get the full picture Before you go, one important note Everything in this podcast, including everything you've heard today is provided for educational and informational purposes only It is analysis, not personal financial advice, and nothing in this episode should be taken as a recommendation to buy or sell any security. Always do your own research and speak to a qualified financial advisor before making investment decisions. Thank you for listening to the AI Supercycle podcast from QF Markets. See you next week.