Latest / The AI Supercycle / The Physical Constraints Behind the AI Supercycle
Transcript
- [0:10] Hello and welcome to the QFMI Market Intelligence Podcast. And today is Thursday the 4th of June. [0:18] This is the show where we deep dive into the AI supercycle. From semiconductors and hyperscalers to energy systems, critical materials, and the macro forces driving what I believe is the largest infrastructure build-out the technology industry has ever seen. At QFMI, we focus on reading between the lines to find the signal in the noise, the infrastructure behind AI, the constraints underneath the headlines, and the gap between financial market narratives and physical world reality. Because increasingly, that is where the real edge lives. [0:55] This is episode 1, so I want to begin by explaining what QFMI is, what we're building here, and why I believe that much of the mainstream AI commentary is still missing the point. I'm Tim Hardwick, founder of Quantumfields Market Intelligence, and let's get into today's show. So here's the premise. [1:14] AI is not just a technology shift. It is the defining capital allocation event of this decade. A structural repricing spreading across semiconductors, data centers, power systems, critical materials, and the macro forces that govern the cost of capital itself. The AI supercycle is not a single trade or a single sector. It is a multi-year structural repricing across interconnected systems driven by physical constraints that most market participants are still underestimating. Memory bandwidth, delivered power, critical materials. These constraints now sit underneath almost every major conversation taking place across AI infrastructure, energy markets, utilities, semiconductors, and industrial supply chains. And understanding how these constraints interact is becoming increasingly important for investors, operators, policy makers, and markets more broadly. We truly are at an inflection point. When historians eventually write about the 2020s, I believe the AI supercycle will be the headline chapter. And I don't say that lightly. [2:28] My background runs across two parallel careers, both deeply connected with tech. I've worked in the tech industry now for 40 years. I run my own tech business through the dot-com boom and the bust that followed. And I've spent the best part of two decades as a strategy consultant on large infrastructure build-out projects, including regional and national network rollouts, data center builds and exits, cloud migrations, the move to 5G, and so on. [2:57] I've worked extensively with the hyperscalers, the telecom operators, and the data center infrastructure providers, the people who actually build and run this stuff, and more recently advising financial institutions on AI strategy, governance, and operating models. I've also been involved in the financial markets for the last 35 years, trading through multiple cycles, booms, and busts. Mostly this was with derivatives such as futures and options across technology, energy, metals and currencies. And more recently, AI and the space sector. All of this mattered. None of it is what this is. And here's what makes the difference. [3:40] It isn't the software. The models are truly remarkable. The engineering behind them is really hard. And the agentic systems are getting better every month. But in every previous tech cycle I've lived through, whether that's cloud or 5G, software was the bottleneck. Compute was cheap. Infrastructure was there. The constraint was the application layer. This time, that's flipped. The capability is racing ahead. The infrastructure simply can't keep up. To run these models at scale, you need an enormous amount of specialized silicon, sitting inside enormous data centers, drawing enormous amounts of power, through grid infrastructure that most countries haven't seriously upgraded in decades. And every layer of that stack has been repriced. Chips, memory, power, copper, uranium, the grid itself. Here's the one idea I want you to take away from this episode. [4:37] The physical layer of the AI supercycle and the financial asset layer of the AI supercycle are completely different markets, running on two completely different clocks. The financial layer moves in minutes. A press release at nine, the stock's repriced by lunch. The physical layer moves in years. A chip fab takes, what, four to five years to build? A power station, longer. A copper mine, even longer still. Picture it like this. A hyperscaler announces a huge capex number on a Tuesday morning, the stock moves that afternoon. But the data center that money is meant to build won't draw power for years, and the chips meant to fill it depend on the memory that's already been sold out. Same announcement, two clocks, one of them measured in hours and days, the other in years. The market routinely treats these two clocks as if they're the same thing. They're not. Not even close and the gap between them the lag between what's been promised in the financial layer and what can actually be delivered in the physical layer is where almost all of the genuine analytical opportunity lives, it's where the bottlenecks form it's where the mispricings open up it's where the narrative gets ahead of the constraint, the constraint eventually asserts itself. And somebody who was paying attention gets paid. QFMI is built on that gap. That's the job. [6:01] So let's go through the constraints. They're the spine of everything we do here at QFMI. [6:07] The first one is memory. There is a lot of talk about the chips at the moment, but not so much about the part of it that decides how fast it can actually think. Compute is fast. Memory is fast. The pipe between them is the problem. That pipe is high bandwidth memory, HBM. And there are just three companies in the world that make it. SK Hynix out in front, Samsung and Micron. And here's what's just happened. SK Hynix has recently joined Samsung in the trillion dollar club. That makes South Korea the first country outside the United States to have two trillion dollar companies on its exchange. The markets have finally figured out just how strategically important this Korean memory company is to the entire AI build-out. [6:57] When HBM is tight, and it has been, repeatedly, it doesn't slow one product down. It rate limits everything. You can have all the chips in the world. Without the memory, you can't ship the system. Without the system, the data centre sits half empty. And the CapEx slide everyone's been applauding turns into a work of fiction. And demand isn't standing still. Every new generation of AI chip is hungrier for memory than the last one. So even as supply expands, the market keeps moving. That's what a binding constraint looks like in practice. The market has noticed now. The real question is what happens next. And that's where a lot of our focus is at the moment in semiconductors and memory. [7:44] The second binding constraint is power. This one is getting quite a bit of airtime at the moment too, and rightly so. A modern AI data centre doesn't draw power like an office. [7:56] It draws power like a small city. One campus can pull more electricity than a town. Multiply that across what the hyperscalers are building, and the grid simply can't deliver it. Not at the cost they're assuming, not on the timeline they're assuming. Which is why hyperscalers are now signing power purchase agreements directly with nuclear operators. Why dormant reactors that sat idle for a decade are being restarted. Started. Why small modular reactors, which everybody had filed under, permanently five years away, have suddenly attracted some serious capex. [8:31] Why utility stocks have started behaving more like growth equities. And none of this is happening in isolation. Power isn't a closed loop. It sits inside the broader energy complex, and right now, that complex is in the middle of the largest oil supply disruption in history. The closure of the Strait of Hormuz earlier in the year took roughly 10 million barrels a day off the market and put the Raslafen LNG facility, the largest in the world, offline for months. Well, that matters for AI, because gas-fire generation backstops most developed market grids, and because hyperscaler power purchase economics don't survive a sustained energy shock. The grid story and the energy story are the same story. We track them together at QFMI for that reason. But it's got to be said, the AI market, the semiconductor rally, has largely ignored the energy crisis up to this point. But it remains to be seen when we reach critical inventory levels, whether that story plays out and whether we see that rally continuing. Certainly looking at the charts, the technicals are looking a little bit overextended. It's been quite a rally up to this point. It wouldn't surprise me that if we got the right catalyst, we'd see a correction of some description, which is absolutely normal and healthy in a bull market. [9:53] The third is critical materials. This is the layer underneath everything. And it gets the least airtime, because it isn't glamorous. Nobody puts copper on a magazine cover, right? However, I need to point out that while critical materials is a useful label, the things inside it don't behave the same way. So we treat them separately. Copper is the headline, though. You can't build a data center without it. You can't build a grid without it. And the AI build-out is stacking demand on top of the electrification trade and the defence trade all at once. Copper mines take a decade to bring online. Demand spikes aren't going to wait a decade, right? So let's move on to uranium, which is both a metal and a fuel, which makes it both a mining story and an energy story. The fuel everyone wrote off after Fukushima is back, because the alternative is admitting the AI build-out can't be powered. So we're watching the uranium ETFs closely. Money flowing into them tells you as much as the spot price itself. And that's not all. As with all of these asset classes and commodities, we analyse the fundamentals, the supply and demand, to ensure that we understand the bigger picture, and the macro forces acting on a particular industry. [11:11] Then we have rare earths, which are really a geopolitics story wearing a supply story costume. China controls the processing. Beijing has shown it will use that as a lever. The US is now racing to build its own supply, but that takes years. So we track this through the rare earth ETFs and what sits inside them. These sit at the foundation. When a binding constraint tightens, everything built on top of it gets repriced. And there's a fourth constraint that doesn't get much airtime either. And very few people outside the chip industry have got a good handle on this one. Yet, it is a very real constraint. [11:52] It's photonics. Here's what the commentary still gets wrong about AI. Everyone thinks it's a compute problem. It isn't. It's increasingly a data movement problem. A modern AI cluster might contain 100,000 GPUs. The bottleneck isn't what any single chip can calculate, it's how fast you can move data between them, GPU to GPU, rack to rack, cluster to cluster, and increasingly data center to data center. A single rack of AI hardware can generate terabits per second of network traffic. The electrical interconnects that worked for traditional data centers are reaching physical limits, which is where photonics comes in. Move the data with light instead of electrons. [12:38] Optical transceivers, silicon photonics, co-packaged optics, optical fiber. It's a whole layer in the AI stack, with its own specific suppliers, its own packaging bottlenecks, and its own concentration risk. And the strongest signal of where this is going has come from NVIDIA itself. The Mellanox acquisition. The focus on InfiniBand and Spectrum Ethernet. The increasing emphasis on co-packaged optics. Jensen Huang now describes what NVIDIA builds as AI factories rather than GPU clusters. Compute, memory, networking, power, and cooling working as a single, integrated system. If any one layer is constrained, the whole factory is constrained. Photonics is becoming as strategically important to AI infrastructure as HBM. The market hasn't fully realized that yet, And that's exactly why we're now tracking this as a fourth binding constraint. [13:37] And there's one place where all four constraints are now starting to collide in a truly interesting way. Space. [13:46] I know how that sounds, but orbital data centers aren't science fiction anymore. SpaceX is filed with the FCC for a constellation of space data centers. Axiom Space is planning to deploy compute nodes on the International Space Station. And this week, SpaceX itself begins the roadshow for what will be the largest IPO in history, a $75 billion raise from the same capital markets that are financing the rest of the AI supercycle. And the reason that this is suddenly being taken seriously isn't that anyone is particularly romantic about it, it's that the terrestrial constraints, such as power, cooling, grid, planning permission, they're all starting to bite hard enough that the economics of putting compute in orbit are no longer absurd. Solar energy in space is continuous, unfiltered by atmosphere and unconstrained by land or national grids. Heat radiates into vacuum rather than needing water-intensive cooling. And photonics, those optical inter-satellite links, turn out to be the foundational enabling technology for connecting all of this together. The constraint architecture in orbit is different. Launch cost, on-orbit power, thermal radiation, spectrum allocation. But the underlying QFMI logic is the same. Physical limits. Deciding what actually gets built and when. [15:08] We've already published a detailed article on the orbital layer. It's in one of our In the Spotlight articles, which you can find on Substack. And we'll come back to that in a future episode on this podcast. [15:19] But this is really where the AI super cycle is heading, and it gets really interesting. Okay, so how do we do it? How do we actually turn all of this information into something useful, something with an edge, rather than just another wall of opinion? We've built a platform for it, and it's called QFMIP, the quantum field's market intelligence platform. It pulls together quite a few things. A trading approach that's built over 35 years, a library of historical market analogues, a statistical engine for reading price action, and a probability model that produces our actual views, our base case. This probability model is called PRISM. It produces probability-weighted views right across the AI infrastructure stack, the hyperscalers, chip makers, the memory, the power, the materials. 12 connected sub-baskets, one consistent method. And there are really just two things you need to know about how it works. First, it works on probabilities, not on certainties. There's no such thing in the markets as certainty. Anyone who tells you that they know what's going to happen the next quarter is maybe selling you something. [16:29] When the picture is genuinely unclear, PRISM says so. False precision is dishonest, in my view, and over a full cycle, it will cost you money. Second, it adapts to the regime. The same chart, the same earnings beat, means different things in a calm market than in a stressed one. PRISM is built to know which regime we're in, which world we're in, and it will weight that view accordingly. And that's it. That's the framework. and Prism isn't a someday product, it's running now. I trade its outputs myself in live markets with my own capital. And you can already see it working if you read the substack in that every week we publish a report called the Weekly Outlook and that carries the Prism's bias table and the base case for the week ahead. Now that's the framework in public, not in theory. The statistical engine and the analogue library, they each deserve their own episodes, so we'll leave it there for today. [17:27] So let's talk about why QFMI exists. There's no shortage on AI commentary out there. It's a crowded market. The problem is that almost all of it comes from one side or the other. Technology insiders who don't trade or market commentators who've never been inside a hyperscaler capacity planning meeting or have never been inside a data center. QFMI is built from both sides. The combination, the technology career that I've spent 40 years building and the trading career running in parallel. That's the whole point. And it's independent. No advertisers, no house view, no conflicts. If the evidence changes, the view changes, and we say so, out loud, on record. So here's what Q4My publishes. The weekly outlook is the flagship every Monday before the US Open, framing the week ahead through the AI lens. In the spotlight is the long-form work which deep dives into a single theme or a single company. [18:28] The market pulse, which is a shorter, sharper, quick read as things move in the market through the week, and the monthly deep dive, which takes one big question and goes into that very deeply. And then there's the weekend debrief. Now, this is interesting because this is actually grading the weekly outlook from the week before. We're actually looking at what we said in the weekly outlook with the bias table in the base case and actually tracking and measuring what happened through the week. How did we perform? How did the model, the framework, perform? Did we get it right or not? And we'll be honest about that. So it's a public record about how we've performed with our framework. There's also a book on the way. It's called Trading the AI Supercycle. Subscribers will hear about the book first as soon as it's launched. And that is looking like August. So that's it for QFMI today. Four binding constraints, one framework. And a belief that this is the defining capital event of the decade, that the repricing has many, many years to run. If that resonates, head on over to qfmi.substack.com and subscribe there. Right now, everything's free. [19:40] Next episode, we start unpacking the constraints in more detail. First up is memory and why high bandwidth memory or HBM is still the rate limiter, even at these prices. I'm Tim Hardwick. This has been the QFMI podcast. Thank you for listening and see you next week.