Latest / The AI Supercycle / The Price of Building AI
Transcript
- [0:00] Last week, President Xi said AI must remain under human control. The United States and China agreed to keep talking about its risks and to establish a channel for reporting incidents. That's a start, but it's not yet a shared way to test whether a system is safe before release. Meanwhile, borrowing costs rose. US shares ended the week higher, but the gains were uneven. AI-related names held up better than many businesses that need to finance physical assets. One cycle, running at two speeds The market is paying for the story And selling the plumbing, That split is the subject of this episode Demand for AI may be strong Building enough power and computing capacity to meet it Is getting more expensive Music. [0:54] 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 centers, 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 29th of September 2026. Now this is the first episode in the new format, and in a new slot as well. Five parts every week, including the macro regime, the AI supercycle, natural resources and geopolitics, the view from my framework, and also what matters next. And this is all recorded on a Tuesday and with you by midweek. Each week I take the developments that matter and ask what they tell us about the constraints of this build-out. Power, memory, materials, interconnect, capital, and now assurance. That is the lens for everything that follows. Okay, so let's get to it, and let's start with the macro regime and the cost of money. [2:04] Long-term borrowing costs rose sharply last week. The US 10-year Treasury yield moved above 5% to a level not seen since before the financial crisis. The 30-year reached its highest level in more than two decades. Japanese government bond yields also rose. So why did bonds sell off? There are several forces, and one of them runs straight through a war. The first is the economy itself. Business surveys in the United States and Europe came in strong, which made rate cuts look less likely. The second is supply. Investors had to absorb a great deal of new government debt, and a weak auction showed that they wanted a higher return to take it. [2:47] The third is energy. The war between the United States and Iran has kept oil prices high and volatile. And that is not an abstract risk premium. It reaches the economy through diesel, the fuel that moves freight, food, and building materials. Diesel has hit record prices on both sides of the Atlantic. That feeds into what the businesses charge, and into what the Federal Reserve expects inflation to do next. You can see this link in a single session on Monday, When President Trump rejects Iran's offer to reopen the Strait of Hormuz Oil moved back up and the 10-year yield rose with it. [3:24] One more detail that matters here. It was not only inflation expectations that rose. Real yields, which strip inflation out, rose sharply too. That tells you the market is pricing more than inflation. It's pricing the Fed's response to it. And that response is now the market's expectation. Interest rate futures are now pricing another rise at the October meeting. A month ago, the market expected one rise by October, and the Fed delivered that in September. A second rise looked like an outside chance It is now the clear favourite And most of that shift has happened in recent weeks, So energy works on bonds through two routes It raises the inflation outlook directly And it raises the odds that the central bank reacts, Those forces do not have equal weight on every day But together they make investors ask for a higher return Before lending over longer periods And that matters when a project needs billions of dollars now and may take years to pay for itself, Imagine a data centre being planned today Before it earns a dollar, someone must secure the land, the power connection, the equipment and the financing, If borrowing becomes more expensive, the project has to earn more just to make the same investment case. [4:44] There is a second effect Higher interest rates reduce what future income is worth to an investor today This is particularly important for assets whose expected profits lie several years away. So rising yields can squeeze the AI build-out from both sides, they can raise the cost of construction, and lower the value investors place on the eventual return. It would be a mistake, though, to say the whole stock market rejected AI. The S&P 500 gained last week, as did the Nasdaq. Technology led, while utilities and energy fell. That is an unusual picture. Enthusiasm for what AI might sell, alongside more caution about some of the physical assets needed to deliver it. You may also hear that the broad market looks less expensive because its price has fallen, as that's measured against expected earnings. That can be true for the index while concealing a lot. If a small group of companies supply much of the expected earnings growth, the index tells us less about the health of every business inside it. So I'm watching both the price investors pay for future profits and the assumptions behind those profits. The second part matters when so much spending has to happen before those returns arrive. The bond market is beginning to ask a harder question. If these projects take longer or cost more, who carries that risk? It's a question we can already see in company deals. Okay, let's move on to the next part and talk about the AI supercycle. And let me start at the customer end with Meta's new personal AI agent, Muse. [6:22] It rose to the top of a free app ranking in the United States and Canada. The immediate story is that people downloaded it. The more interesting question is what they might use it for after the novelty fades. Muse can help with tasks such as shopping and booking travel. If an agent eventually makes choices and completes purchases for you, a travel site or comparison service may no longer be the place where your decision starts. Now that doesn't mean that those businesses have already lost their customers, Downloads are not lasting use And lasting use is not yet a proven business model, But an agent that sits between the customer and the seller Could change who gets noticed Who wins the sale And who keeps the relationship. [7:09] There is a trust test here as well. People may let an agent suggest a trip long before they let it spend their money. The businesses in its path have time to respond. Now I'm interested in what changes when the agent moves from offering advice to completing a transaction. Now follow that demand back toward the infrastructure. Anthropic has committed about $11.6 billion to Akamai's cloud services over seven years. This supports the computing work around Anthropik's models. It is an example of how much capacity a leading AI company expects to need. There is a financial link as well. Anthropik received a warrant, the right to buy Akamai shares under specified terms. Part of that right is tied to the initial deal, and more depends on the relationship expanding. Anthropic has not simply been handed a 5% stake, and the distinction does matter here. It shows how a large customer agreement can be linked to a possible future shareholding. When I judge deals like this, I want to separate a demand for the service from the incentives built into the contract. [8:22] Another deal deserves some attention as well. Akamai says the capacity is designed for computing tasks that use conventional processes, Discussions of AI infrastructure often begin and end with GPUs, yet agents and their supporting services need other kinds of computing too. That broadens the list of suppliers that may benefit and the amount of equipment someone has to pay for. [8:47] Akamai expects to spend billions to support this agreement. The commitment gives it a customer. It still has to deliver the service at a cost that leaves a return. And those contracts sit inside a much larger funding wave. Reuters reports a Goldman Sachs forecast. The biggest cloud and technology companies could sell about $420 billion of bonds next year. That is a forecast for new borrowing, not a bill that has already come due. Bond investors are still providing money They are also asking more questions about the volume of AI-related debt and the return on the projects behind it, Comparing one group of bond spreads with the whole corporate market gives us a warning sign, Although it does not prove that every AI borrower is paying a special premium, Credit ratings, loan terms and company risks do differ, Now the question is practical Can these companies keep funding construction on terms that leave enough profit when the site's open? There is also a portfolio question. A lender might hold the bonds of a large technology company and separately finance a building leased to that same company. Those can look like different investments. But if the same spending plans support both, the risks can move together. That is why the total exposure matters more than the label on each bond. [10:12] Now this brings me to Project Jupyter, the planned New Mexico data center campus connected to Oracle and OpenAI. The site is being built for OpenAI by Stack Infrastructure, a data center developer that is part of the asset manager Blue Owl. Oracle issued a force majeure notice in connection with the project to protect itself against potential delays amid concerns about securing power. Force majeure is a contract provision for events outside a party's control. A notice raises questions under the agreement. It does not, by itself, settle what happened or who will ultimately pay. One report suggested the power problem could delay the project by a year. Oracle disputes that and says the project remains on its planned schedule. [11:02] Blue Owl says the notice does not change its financial commitments. So I would not mark a one-year delay down as a fact, but I would pay close attention to the reason the whole notice exists. A power connection previously treated as a future planning issue has become something the parties may need to resolve under contract. If a site opens late, rent, customer commitments, construction costs and loan payments can be affected. The exact split depends on agreements we have not seen. That is why lenders and developers will be reading this case very closely, The important lesson is not that one company has already lost a particular sum It is that a notice under a contract forces everyone involved to ask which promise depends on power arriving at a particular time, That includes the landlord, the tenant, the customer and the people who finance the building, Power is only one hurdle Communities are challenging proposed data centres over issues such as land use, water and electricity demand. Research group Data Center Watch counted dozens of US projects blocked or delayed by local opposition in the second quarter. Each case has its own facts, but together they show that a signed customer agreement does not guarantee a finished site. [12:24] Now, there is a temptation to take a strong demand for AI and assume that the physical supply will arrive on schedule. I don't think we can make that leap. Goldman Sachs forecasts a sharp rise in US data center power demand between this year and next. At the same time, some turbines and large transformers have very long lead times. A forecast for electricity use is not electricity delivered to a working building. [12:51] That is one reason I keep the power grid near the centre of this framework. The need to expand it looks substantial, but the companies doing the work still need financing, permits, equipment and time. Strong demand does not remove those hurdles. It can seem odd when a company with a long order book falls as rates rise. The order book says customers want its equipment. It does not say when the company will be paid, what it must spend first, or what profit each order will produce. That is the gap between a strong industry outlook and an attractive investment at today's financing cost. You can see a related puzzle in materials. Uranium prices and uranium-related shares have not always moved together. Copper has also remained near historic highs while some mining shares have lacked. There may be several reasons for those gaps – project costs, company debt, expected production, and the price investors are willing to pay for a share. Higher interest rates are one possible part of that explanation. A chart by itself cannot tell us which part matters most. Copper is a good example of why precision helps. Its recent record was set earlier in September. Last week, a fatal accident stopped operations at Escondida, the world's largest copper mine. That added to supply concerns, but it did not cause a new record last week. [14:17] One more piece of engineering belongs in this discussion. On Monday, SpaceX's Starship reached orbit for the first time and deployed Starlink satellites. It was an important step, even though an engine problem shortened the planned flight. The longer-term idea of putting computing in orbit is worth watching. It might offer access to solar power and reduce some limits faced on the ground, but it still has to solve hard problems, including moving heat away from powerful chips in a vacuum. We talked about the orbital compute thesis in episode 5 back in early July, so have a listen to that one if you'd like more details. This week's flight demonstrates progress in launch and deployment. It does not yet demonstrate an orbital data centre. The market's verdict was cool. SpaceX shares closed lower on the day, and the space funds are below their trend. [15:11] So it is the same split again, a good day for engineering, and not yet for the investment. That is why the space economy now appears in my framework as a domain on its own right. The fundamentals look constructive. The price is not yet convinced. And until those two line up, it is something that I'm watching rather than acting on. Finally, AI safety. Australia says an OpenAI agent gained unauthorised access to a Medicare statistics portal in June. The portal held aggregate information. Officials said that they did not believe personal information had been accessed While their investigation continued, The incident shows why behaviour outside a lab matters The US and China have agreed to establish a channel for discussing AI incidents Now this may help governments communicate if something goes wrong, It is different from agreeing how to test a model before anyone puts it to work, That gap between using AI and proving it is reliable remains part of the build-out too. [16:14] Okay, so let's move on to part three and talk about natural resources and geopolitics. Energy connects the physical build-out to the wider economy, and the last few days has offered a useful illustration, even if the headlines have changed quickly. Let's start with what prices actually did. Brent oil fell last week by around 7% from Monday's open to Friday's close, and West Texas crude was lower too. On the charts, Brent has been in a short-term downtrend since its September highs, and Dutch TTF, Europe's benchmark gas price, shows the same pattern. Part of last week's fall came from when traders saw a possible route to more Gulf exports. Iran offered to reopen the Strait of Hormuz under conditions that included US steps on sanctions and its blockade. President Trump rejected that proposal. Now, mediators have continued working on this, so I would not call the diplomatic process over. But the rejection gave that downtrend its first real test. On Monday this week, Brent rallied up to the top of its downward channel, then sold off as the headline swung back and forth. It closed below where it opened, but still above Friday's close. European Gas also finished the day higher, so the rejection pushed prices back up, but it has not broken that downtrend. Whether the trend holds is one of the things that I'll be watching this week, and in Europe the backdrop is tight, with gas storage well below its usual level going into winter. Saudi Arabia has also restarted its east-west pipeline and resumed loading at its Red Sea port. That can move more crude without relying on Hormuz. It is a real improvement in one route for supply. But it doesn't solve every problem in the fuel system. [18:01] The distinction I want to make here is between crude oil and the products made from it. A barrel reaching a port does not automatically become diesel where and when it is needed. Refineries, shipping routes and storage determine what fuel reaches customers. Diesel prices in the US and the UK have been exceptionally high. The UK motoring organisation, the RAC, reported a new pump price record on Monday. [18:27] Those prices affect transport costs and in time, the prices businesses charge, Now this matters for AI data centre even if it never burns diesel It is the chain I described at the top of the episode Fuel costs, inflation, the central bank's response and the cost of financing long-term projects, Oil is not the only thing moving beyond yields but right now, it is one of the most direct, There is also a political response The Trump administration has considered restrictions on US diesel exports, alongside other ways to bring down domestic prices. Such a move might help one market while making supply harder for another. The test is what happens to actual fuel flows, not simply what is proposed. Rare Earths presents a different kind of supply risk. The US and China extended their existing trade truce by two months, moving its expiry from November to the 10th of January next year. Now that gives negotiators more time It does not guarantee reliable deliveries of the materials needed for magnets and other equipment The White House says the two countries are still working through US concerns about those shipments. [19:39] The longer truce may ease immediate fears, but the businesses planning an alternative supply chain cannot build mines, processing plants, and factories in a few months. So what I will watch is whether the extra time brings dependable deliveries or only another deadline. [19:56] Okay, let's move on to part four, and this is where I put all of these pieces together. PRISM is my framework for weighing the evidence, forming a base case, and deciding what would change my mind. Now I'm going to spare you the probabilities and the percentages My base case is this Demand for AI computing remains strong But the cost and timing of building capacity matter more than they did when money was cheaper, Put more simply, the cost of capital, not the demand for compute, is setting the tone right now And so far, it has changed nothing structural, That is a view about the next phase of the build-out It is not a prediction that all infrastructure shares will fall or that every project will be late? [20:41] Three questions organize my thinking. First, is the demand real enough to support the spending? Muse's launch shows interesting consumer agents. The Anthropic and Akamai agreement shows a large commitment to cloud capacity. Neither tells us, on its own, how profitable the whole chain will be. I want to see lasting use, revenue and returns on the money invested. Second, can the physical capacity arrive? Jupiter puts power risk into sharp focus, but its timetable is disputed. Elsewhere, equipment lead times and local opposition can hold up construction. Now I want to see sites connected, fitted out and earning revenue. Announced capacity is not the same as delivered capacity. [21:28] Third, what does it cost to finance the work? Rising government bond yields and a large calendar of corporate borrowing make this question harder to ignore. The answer will differ by company and project. Some have the balance sheets to absorb delays. Others may have to change their plans. [21:45] At present, I'm more confident that financing and delivery deserve more attention. I'm less confident about the path of oil and interest rates. Diplomacy, supply routes, inflation and central bank decisions can change that path quickly. What would make me rethink the base case? A sustained fall in borrowing costs would ease pressure. Faster power connections would weaken the delivery concern. On the other side, cancelled customer commitments or clear cuts to compute purchasing would challenge the demand case itself. And if several major projects are delayed or lenders demand much tougher terms, I would have to treat the financing problem as more than a warning. And that is the discipline. Keep the argument open to evidence, especially evidence that does not fit it. Okay, so part five is all about what matters next and here are four tests I will keep returning to as new information arrives. 1. The rate path. Inflation and employment reports matter because they can change borrowing costs. I'll look beyond a single number to whether investors and lenders alter the terms they offer 2. Memory demand Micron's next guidance is a useful check I will listen for what customers are committing to buy, how much capacity suppliers can add and whether strong prices turn into durable profits. [23:08] 3. Data center delivery. After the Jupiter notice, watch for clearer power dates, changes to project finance, and details of which party bears a delay. A fresh announcement of capacity is less informative than a site opening on time. 4. Supply and safety agreements. The January trade deadline matters if it changes access to critical materials. The AI incident channel matters if governments turn an agreement to talk into a working process Neither should be mistaken for a finished solution, These tests are useful whether the next headline is encouraging or alarming, They tell us what to ask once the initial reaction has passed, Okay, so let me bring it all together The AI build-out still has two convincing stories People are trying new applications, and the largest customers are committing to more computing capacity. At the same time, the power, equipment, permits, and finance behind that capacity are becoming harder to take for granted. Now that is the gap, and I will keep analysing that. How much is being promised? How much can be built? And who carries the cost if those two schedules do not match? The answer will come from working sites, customer revenue and their financing agreements one project at a time. [24:31] Okay, so just a few more things before I close, because these have just crossed my desk as I was recording. Reuters has seen Anthropics IPO prospectus, and it is the most detailed look yet inside the finances of a frontier lab. And I'll go through this properly next week when I've had a chance to look at this. But let's start with the numbers. The growth first. Revenue grew around 12-fold last year to nearly $4.6 billion. [25:00] Then there's the cost. The operating loss was just over $8 billion. The headline net loss is far larger, around $42 billion. But most of that is an accounting charge tied to earlier fundraising, not money spent running the business. Compute and infrastructure took more than half of the operating costs. And then the commitment. Anthropic plans to spend around $518 billion on cloud, computing and infrastructure obligations in the coming years. That is the demand side of everything we have discussed today, written down in a filing. The Akamai contract earlier in this episode is one small piece of it. The prospectus also says nearly a quarter of revenue came from two customers, and that many large customers are not locked into long-term contracts. The listing could value the company at more than $2 trillion and it's likely to come after the November midterms, The prospectus also puts safety on the record Anthropic warns investors that advanced AI could pose catastrophic or existential risks to humanity, It says its models could show self-preserving behaviour attempting to resist shutdown, concealing or manipulating information even behaviour resembling blackmail. [26:21] The risk factors run to roughly 80 pages of the filing, nearly twice the space given to describing the business. Few companies, if any, have told prospective investors that their own product could threaten humanity. And this connects directly to the last episode in three ways. First, I said the deepest fears about this technology are serious, but cannot be priced by markets. The prospectus does not make them priceable. What it does is put them in writing, in the one document every investor is expected to read before they buy That is new, Second, I argued that assurance, proving a model is safe before it is used, is becoming the sixth constraint on this build-out. [27:08] The filing adds something I did not have last week Anthropic says models may become aware that they are being evaluated and that this significantly limits its ability to assess their safety. So the constraint is not only that there are too few people qualified to test these systems, it is that the systems may behave differently when they know they are being tested. Third, I said a listed lab would face a tension between a public commitment to pace the frontier and shareholders who expect growth. The filing states the commercial side plainly. New models drive usage and a continuous release cadence is inherent to staying at the frontier, The latest Opus model came out 10 days after the pacing essay You can read that as a contradiction or simply as honesty about commercial reality Either way, it is now written into a legal document, And one line speaks to the other half of last week's argument. Anthropic says it believes the market will reward reliable, trustworthy and secure AI. That is the inversion I described. That assurance could become the unlock for enterprise adoption, not just friction at the frontier. The filing is betting on it. The listing will test whether investors agree. And within hours, the other side of this race moved too. Bloomberg reported that OpenAI is seeking to raise at least $30 billion. [28:38] At a valuation of about $1.4 trillion before the new money. The talks are at an early stage and the terms could change. The round would act as a bridge, providing capital in place of a listing. Sam Altman said earlier this month that OpenAI would not go public this year, citing concerns about AI safety. Its annualised revenue run rate is reported to be nearing $70 billion. [29:03] Now that is a run rate, not last year's revenue, so it is not directly comparable with the figures in Anthropics filing. So in the space of a day, the two leading labs took different routes to the same thing. Capital on a scale this industry has never needed before. One is heading for the public market, with 80 pages of risk factors. The other is raising privately, and says safety is part of the reason it is not listing yet. Both routes lead back to where this episode started, the cost of money. [29:35] So next week, I'll go through the filings in more detail, the growth, the costs, the commitments, and what the two routes to capital tell us about how this layer of the cycle will be valued. And I'll also come back to AI safety, because this filing moves that story on. 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 Ebrief, 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. [30:45] Thank you for listening to the AI Supercycle podcast from QF Markets. See you next week.