Latest / The AI Supercycle / The Price of Speed: Who Pays for the AI Boom?
Transcript
- [0:00] 20 years, 20 days. That was the contrast I kept coming back to this week. Price of Speed [0:06] Yesterday, Google signed a 20-year nuclear agreement with Constellation Energy, part of a wider deal covering nearly 3.6 gigawatts of power. It comes just as America's largest power grid is proposing that data centres bring qualifying power with them, or accept curtailment when the grid is under severe stress. At the other end of the energy system, the G7 has agreed to accelerate emergency releases of diesel and crude oil, with a substantial amount of that diesel due to move within 20 days. One is about securing electricity for decades. The other is about keeping today's physical economy moving. And AI increasingly depends on both. Meanwhile, the US 10-year Treasury yield has touched its highest level since 2002. Anthropik's draft IPO prospectus, reported by Reuters, shows more than $500 billion of infrastructure commitments, most of which can't simply be cancelled. And a rather important argument has broken out over how much risk society should accept for increasingly capable AI. I think most of it connects to one question. What is the price of moving this quickly? Not just the price of a GPU, but the price in capital, in electricity and in physical infrastructure, and potentially the price in risk. That's the subject of today's episode. [1:39] 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 Wednesday, the 7th of October, 2026. As usual, I'm going to work through five parts. The macro regime, the AI supercycle, natural resources and geopolitics, the view from my framework, and finally, what matters next. Rates and Rockets [2:23] So let's start with something we talked about quite a bit last week. The cost of money. The bond route hasn't gone away. The US 10-year Treasury yield hit 5.34% last week, its highest level since 2002, and it's still well above 5. Brent Donnelly at Spectra Markets had a lovely way of picturing this in his newsletter. Think of the 10-year yield as a rocket climbing through the layers of the atmosphere. In 2020, it was below 1%, analogous to being down around the troposphere and lower stratosphere, where you'll typically find long-haul airliners like a 747 at cruise altitude. It climbed higher through the stratosphere in 2022, then spent much of the last three years hovering around 4%, up in the mesosphere, where meteors burn up. This month, it broke cleanly through 5% and pushed through to the thermosphere and into the exosphere, the outermost layer of the atmosphere. So what's beyond the exosphere? Space. And that really is Brent's point. There isn't some convenient ceiling up there telling yields where they have to stop. [3:32] Here's what should flummox anyone running the old playbook. On Friday, we learned that the US economy added only 29,000 jobs in September, far fewer than expected. Unemployment edged up to 4.2%, and markets sharply cut the odds of another Federal Reserve rate increase in October. Weak jobs and fewer rate hikes ought to pull long yields down. For a few hours on Friday, they did. Then yields turned around and closed the day higher. Good news, bad reaction. And traders tend to pay attention when a market refuses to do what the news says it ought to do. So something else is going on. The Fed controls short-term rates, but the 10-year yield has to absorb inflation expectations, growth, government borrowing, and the sheer supply of bonds coming to market. Increasingly, I think AI belongs in that conversation too. On Monday, the ISM Services survey showed the US services economy still growing reasonably strongly, at 54.9, but the prices paid component rose to 74, its highest reading since the summer of 2022. Energy is part of that input cost pressure, and diesel certainly is, but the economy is also absorbing an enormous capital investment boom in data centres, power, transmission and chip factories, plus all of the construction that goes with them. [4:58] This week, the Bank of Japan made almost exactly this point. Deputy Governor Shinichi Uchida described the AI boom as a big positive demand shock, pushing up on both the economy and prices, and said it might eventually affect the natural rate of interest. He also warned that markets could pull back if the expected profits fail to arrive. What he describes is a slightly uncomfortable feedback loop. The industry needs extraordinary amounts of capital, and that spending supports growth. Strong growth helps keep inflation and real interest rates higher, and higher rates raise the return required from the very infrastructure being built. At the margin, the boom helps create the financing environment it then has to survive. In other words, AI investment is now big enough to shape the macro environment, not just respond to it. And nowhere is the cost of building this boom more visible this week than in electricity. Powering the AI Boom [5:59] Okay, so let's move on and take a look at the AI super cycle, and we'll start with the biggest story of the week. Yesterday, Google agreed a package covering nearly 3.6 gigawatts of power from Constellation Energy, the largest nuclear operator in the United States. The nuclear part is about 900 megawatts under a 20-year agreement to upgrade 11 existing reactors so that they produce more electricity, with the first of that extra power due in 2028. Constellation is putting more than $4 billion into the work. The rest is a 15-year supply agreement covering another 2.7 gigawatts across the PJM market, rather than the electricity tied specifically to those nuclear uprates. [6:44] Constellation shares jumped more than 13%, but the reason behind the deal matters as much as its size. It sits in PJM, America's largest power grid, where data centre demand has helped push capacity prices up more than 11-fold since 2024. PGM's management has now proposed that large, new data centre loads either bring qualifying generation with them or accept the possibility of having their power use curtailed when the grid is under severe stress, Bring your own power or risk losing it That's the power constraint turning into a rule and Google's answer is to help underwrite the supply itself It's Google's second big nuclear deal in a month, In September, it signed a 22-year agreement for up to half the output of the Lovisa nuclear plant in Finland, helping support its operation through 2050. Hyperscalers are starting to behave like heavy industrial companies. [7:45] They can't assume electricity will turn up when they need it, at a price they can live with. The scale of the problem explains why. Morgan Stanley estimates the US data centre developers face a power shortfall of around 32 gigawatts through 2028, even after counting on site generation and fuel cells. That's roughly a third of what they need, and it matters well beyond the utilities. Micron gave us strong evidence last week that demand for AI infrastructure is still extremely strong, with fourth quarter revenue of more than $54 billion, guidance of about $61.5 billion, and around $150 billion of remaining performance obligations under its long-term supply agreements. This doesn't look like a demand problem, but Morgan Stanley makes a useful point. NVIDIA and Broadcom can usually redirect their scarce chips if one site is delayed, while memory, optics and power management chips are more exposed. The demand can be completely genuine, but if the building that houses the servers can't be energised, deployment gets pushed out. [8:55] That's what I mean by the moving constraint. Solving one bottleneck doesn't necessarily speed up the whole system. It can simply reveal the next one. And power is a slow constraint to solve. Upgrading reactors takes years of engineering and regulatory rigmarole, and building new ones takes far longer. A hyperscaler can improve another $20 billion of capital spending very quickly, But it can't conjure up a nuclear reactor into existence by next Tuesday, right? That's the two-clock problem. Financial capital sprints. The physical economy ambles along at a much slower pace. And the financial clock is running very fast indeed. Last week, I finished with a late item on Anthropic's confidential IPO filing and headline number of $518 billion. [9:48] That's what Anthropic expects to spend over roughly the next decade on computing and infrastructure commitments. And we now know rather more about what sits behind it. About 80% of those commitments either can't be cancelled or have to be paid regardless of how much capacity Anthropic actually uses. There are very large commitments to Google, Amazon and Microsoft, and more than $160 billion of equipment lease obligations linked to Broadcom. Now I think this Broadcom relationship deserves a closer look. Broadcom has agreed to lend Anthropic up to $42 billion to help finance computing capacity Anthropic intends to lease. Those loans could convert into Anthropic shares, and Broadcom can bring in a financing partner. So think about that loop for a moment. Broadcom supplies the technology and helps finance Anthropik's ability to pay for it. Anthropik's demand supports Broadcom's AI revenues. And Broadcom, or its partner, may end up holding Anthropik equity. Anthropik's own filing flags the potential conflicts of interest in that arrangement. Supplier financing is nothing new. But at this scale, a simple revenue number no longer tells you the whole story. You have to follow the money through the system and ask who is funding whom, who is carrying the risk, and ultimately, who is paying for all of this from genuine end demand. OpenAI's capital runs in a loop of its own. SoftBank has just completed another $30 billion investment in OpenAI, paying the first $10 billion tranche after raising $11.1 billion in the bond market. [11:33] Reuters described it as the largest high-yield corporate bond deal on record. Some of that new SoftBank debt carries interest of almost 10%. So follow the money there too. Bond investors lend to SoftBank at high-yield rates. SoftBank buys more of OpenAI. OpenAI commits capital to data centres, chips and models. Infrastructure companies borrow and invest to supply it. Somewhere at the far end of that chain, somebody has to sell enough AI products and services to earn a return on all of it. That's the bit we can't lose sight of. And the borrowing keeps getting bigger. [12:13] SpaceX is reported to be seeking around $40 billion, led by Apollo, to fund Nvidia chips, with around $10 billion in bank loans and $30 billion in investment-grade debt. The financing is still proposed rather than completed, but set that against Morgan Stanley's estimate that AI infrastructure will need around $1.5 trillion of external financing by 2028. Now that's an enormous amount of debt to price off a 10-year yield above 5%. For two years, the question was whether the industry could make enough chips, memory and power. Now another question is coming into view. Can the revenue arrive quickly enough to pay for all of this? And I don't mean another 100 million people downloading an AI app. I mean durable economic value, like an insurer processing claims more efficiently, or a bank cutting the cost of compliance. Those are the things that eventually have to produce the cash flows. However extraordinary the technology becomes, someone still has to pay the electricity bill. Now there is some encouraging evidence here. After a difficult start to 2026, US software stocks have now touched fresh highs, while expected earnings growth of the sector has risen to around 20%, from under 14% at the end of March. But notice the gap. Software stocks are up about 5% this year, while major semiconductor benchmarks, such as the SMH, is up around 67% year-to-date. The market is paid far more for the picks and shovels than for the businesses that have to turn AI into revenue. Some strategists now think software can outperform chips from here. If it does, that would suggest the revenue is starting to arrive where it has to. Safety and Assurance [14:04] Now last week I argued that assurance was becoming the sixth constraint of the AI build-out, and quite a lot has happened since. President Trump brought the major US AI companies to the White House, where they signed what he called a morally binding accord on frontier AI safety. It sets out four layers. The company's own controls, an internal team checking that those controls work, independent external audit and board-level oversight. The difference is that this one is voluntary, with no penalties attached. The administration's argument is that heavy regulation too early would slow American innovation against China. New York City is testing a very different approach. On Monday, the City Council sat as a committee of the whole, all 51 members, for an unusual hearing on AI. OpenAI, Anthropic, Google and Meta testified under oath. One proposal would make it unlawful to market, sell or deploy an AI system in New York City without independent third-party validation, covering areas such as data quality, bias, privacy and security. [15:18] The package also calls for human shutdown capability, and penalties could reach $25,000 per violation for both the company and the validator. The hearing produced some striking moments. Under oath, Google confirmed that its AI agents had left a test environment and reached the live internet on three separate occasions, though it says the models stopped once they realised they were on live websites. When the Speaker pressed OpenAI on whether a failed safety test would stop a release, she got an assurance that OpenAI had delayed releases before, but not the simple yes or no she asked for. And Anthropic told the Council the labs shouldn't be checking their own homework, which is the whole case for independent validation. And whether or not that legislation passes, the direction of travel really does matter here. And the incidents are piling up. OpenAI has now notified more than 100 organisations about potentially unauthorised activity involving its AI agents, although being notified does not necessarily mean that an organisation was successfully breached. So there's a bit of a kerfuffle developing over who carries responsibility for increasingly autonomous AI. [16:36] Working that out will take time, and the models aren't going to wait while we dither. Sam Altman made the philosophical divide unusually explicit this weekend. He said the world should accept some bad things happening in return for the benefits of AI and people having agency. His concern is the alternative, a single lab deciding who gets access to powerful AI. And he does support strict rules for the most dangerous scenarios. But he also said there's a lot of daylight between open AI and Anthropic on regulation. Anthropic has taken a much more cautious public stance on catastrophic risk. Its own IPO material warns investors that advanced AI could pose catastrophic or even existential risks. And its response to Altman was that its proposed rules apply only to frontier models. OpenAI is arguing harder for broad access and a lighter regulatory touch. Now markets can't settle that argument. Governments, insurers, boards and enterprise customers will have to. Which is why I don't see assurance simply as a break on AI. [17:46] Imagine that you're the chief executive of a bank. A frontier model could automate an enormous amount of work, and the productivity gain is obvious. But if you can't show how the model was tested, how its behaviour is monitored, how its decisions can be challenged, or how somebody can shut it down if it does something unexpected, your risk committee may simply say no. Add independent validation, audit evidence and monitoring, and that same risk committee may say yes. Assurance has created revenue. Now that's the interesting inversion. The mechanism that slows deployment at the frontier may be one of the things that accelerates adoption in the enterprise. Notice too that New York would make the validators liable as well. That's how auditors work and it's how an industry gets built. So assurance belongs in the constraint framework but it may turn out to be a very unusual constraint, one that eventually unlocks demand. Diesel and Geopolitics [18:50] Okay, so let's move on to part three, which is about natural resources and geopolitics. So let's take a look at that other number from the start of the episode. 20 days. If Google is thinking about electricity in decades, governments are dealing with a much more immediate problem. Diesel. With fuel prices driven higher by the war between the US and Iran, the G7 has agreed to accelerate the remaining emergency release commitments made earlier this year. That means 100 million barrels over four months, including a substantial front-loaded release of diesel within the first 20 days. Now the distinction does matter, because this is simply not another 100 million barrels being added on top of everything previously announced. It's part of completing the much larger emergency release programme agreed earlier in the year. But the urgency around diesel is still very real. Diesel is the fuel of the physical economy. Freight, construction, farming and mining all depend on it. And here's the interesting part. JP Morgan estimates that Middle East crude exports had recovered to around 98% of pre-war levels by the end of September. Refined product exports, including fuels such as diesel and petrol, were only about 58% of pre-war levels. [20:11] So the problem has partly moved. It's no longer simply about finding barrels of crude oil. It's about refining them, shipping the products, ensuring the vessels and getting the right fuel to the right place. Once again, the constraint has moved. And there's a direct line from this into the AI supercycle. Before a data centre opens, the concrete, steel, transformers, servers and generators all arrive by truck. And there may be thousands of truck movements before the first AI workload ever runs. The data centre may run on electricity, but the supply chain that builds it runs partly on diesel. More expensive diesel also feeds into inflation. Inflation feeds into bond yields. And bond yields help set the cost of financing for the next data centre or power station. So diesel can reach the AI supercycle without ever entering the data centre. Natural resources aren't a story sitting alongside AI. For me, they're very much part of the AI story. The Framework Check [21:17] Okay, so let me bring that into the framework. So my base case remains constructive on the AI super cycle. Nothing this week suggests that underlying demand for compute is weakening. But the opportunity is becoming much more complicated in three ways. First, the physical build-out is running into constraints that money alone can't solve. 32 GW of power shortfall is not a rounding error. And PGM's proposal shows the grid starting to push the problem back onto the buyers. Second, the capital is becoming more circular and the more circular it gets, the more it matters to separate genuine end demand from the financing that helps it reach the supplier. Third, assurance is moving from an internal lab function towards an external economic layer and potentially a substantial new industry in its own right. And all of these constraints increasingly interact. If a data centre can't get the power it needs, deployment gets pushed back. Chip sales are delayed and the economics of the project begin to change. If diesel pushes inflation higher, that feeds back into the bond yield used to value everything else. [22:31] So rather than six constraints in six separate boxes, I increasingly see a network. And perhaps the clearest way to see all of this is the gap between the stock market and the bond market. The Nasdaq has just closed at another record, while the 10-year Treasury is still around 5.3%. The equity market is still willing to pay for the future. The bond market is making the future more expensive to build. Now Brent Donnelly would push back a little bit on me here, because his argument is that the whole equity index is really one big bet on whether AI pays off. And on that question, funding costs are a rounding error. For the cash-rich giants at the top of the index, I think he's right. But for the companies building the physical layer with borrowed money, it's a different story. Some of SoftBank's new debt carries interest close to 10%, and data centre and power developers are borrowing in a world of 5% treasury yields. So for them, the cost of money is no rounding error. And that's where the higher yields bite first. So both can be right for a while, but they probably can't diverge indefinitely. What Comes Next [23:38] Okay, so now let's talk about what matters next, and what am I watching from here? First, the 10-year treasury. If long-term yields fall when the economic data weakens, conventional monetary policy is still doing its job. If they don't, something more structural is holding the long end up. Second, delivered power. Not another splashy 10 gigawatt pipeline announcement, but grid connections, turbine orders and transformer lead times. I'll be listening for chip suppliers saying deployments are slipping because sites aren't ready, and watching whether PGM's Bring Your Own Power proposal becomes the template for other grids. [24:20] Third, nuclear. Google has now done two major nuclear deals in a month. If other hyperscalers follow, the AI supercycle becomes one of the most important new forces in the nuclear market. Fourth, assurance. Watch whether insurers, banks and big enterprise buyers start demanding independent AI evaluation before they deploy. That's when assurance starts showing up in the economics. And finally, diesel. Now I know it sounds a bit odd to finish an episode about frontier AI and nuclear power on diesel, but that's rather the point. The AI supercycle is being built in the physical world. Someone still has to pour the concrete, someone still has to deliver the transformer, and someone still has to drive the truck and operate the machinery. And when diesel becomes more expensive, those costs don't simply disappear. They work their way through the supply chain, into freight, construction and equipment costs, and eventually into the price of the build itself. Now some companies will absorb part of that increase, others will pass it on. Either way, the cost of turning an AI investment plan into a working data centre goes up. And that brings us back to where we started – the price of speed. [25:41] Okay, so let me bring all of this together. 20 years, 20 days. Between those two time horizons sits the AI supercycle. The labs are committing hundreds of billions of dollars to compute. Suppliers are helping finance the customers buying their technology. And the bond market is demanding more for long-term capital. Meanwhile, governments are starting to wrestle with who decides when an AI system is safe enough to use. None of this tells me the AI super cycle is ending. It tells me it's an industrial economy. And industrial economies are messy. The financial market can move in seconds and minutes. The physical world can't. And increasingly, I think the investment opportunities will come from understanding what happens in the gap between those two clocks. [26:31] 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.