Capital Cycles and AI
What is a capital cycle and why AI is building the mother of all capital cycles
When I started this newsletter, I never thought I would be this profoundly involved with cycles. After specializing in semiconductors, cycles have become my overarching industry framework, and they feel universal to most human activities.
Everything has cycles. As sure as the tide comes in, tides go out. A glut follows every semiconductor shortage. It’s a universal law and rule of capitalism. We are in the middle of the most potent cycle I’ve seen in my professional career: AI.
There are many frameworks and ways to think about this concept, but my favorite is “The Capital Cycle.” Marathon Asset Management uses the capital cycle framework to demonstrate that even money is cyclical.
Simplistically, money chases high returns. This is intuitive. When an investor makes an extremely high return on capital, capital chases high returns into the space until returns on that capital diminish. This is a fundamental tenant of capitalism.
Usually, there is a correction mechanism. When the supply of capital is too high, the supply and demand of the underlying investment crosses, and returns plummet in the space. Now, there’s another consideration: from time to time, this cycle breaks with a short-term feedback loop, and that is often called a bubble.
The capital cycle is well-trodden, and a massive capital cycle is almost always before a technological revolution. Virtually every new technology offers exceptionally high returns on capital for the first movers and, more importantly, requires a lot of new capital.
If the world is going to change by introducing a “new technology,” such as railroads, the internet, or AI, it makes sense that there will likely be a massive need for infrastructure. Returns on capital are typically super high initially, and when there are high returns on capital, capital (money) will chase those returns and keep investing in that asset class until the returns collapse. Since there is a vast new capital stock to build, this often takes years to deploy.
But with capital entrance comes supply in the supply-demand dynamic. As more capital enters the space, supply and demand cross. Usually, supply doesn’t perfectly match demand; it massively overshoots it. When that happens, a shortage becomes a glut. This is the law of the land in semiconductors, where capacity comes on in lumpy supply additions. And this has happened in almost every major technological bubble (or capital build) in the past.
As new technology ramps, capital rushes in, and technology that supplies the demand often improves. At some point, the shortage of whatever precious resource is needed becomes a glut.
However, this is how cycles work in a microeconomic way, yet sometimes, a cycle becomes “different this time.” I want to explore this key “different” because I think it is missing this cycle to christen it a bubble.
The key difference is an external feedback loop that almost instantly marks up other tangential assets. This important feedback loop convinces move marks quickly and justifies even more investment. A rapid return on capital will convince early investors that the correct answer is to bet again. Often, the craziest capital cycles had a meaningful feedback loop.
Feedback Loops and History
Each capital cycle is different. I won’t pretend to know what this one holds, but I love history as a guide. After I mention some of the great pieces of history, I'll discuss semiconductors and AI. AI requires a lot of computing, so capital is rushing into the industry.
This similar logic has repeatedly played out. So, I want to set the stage with the key memes of the past and then discuss how they will impact today. I will briefly mention railroads, the internet, and AI.
Railroads (1840-1870): The key meme of this period was “territorial development.” Railroads were a transformational network technology. They were 90% cheaper than horses and wagons and enabled new business models based on trade and regionalization. The story goes like this.
Railroads create value by being built. If a railroad is built in a new territory, the land values along the route increase, new towns spring up, and agriculture and industry follow—the core railroad profits from land appreciation and freight. In the case of the railroad bubble, land grants from the US were the primary feedback mechanism.
This was a powerful feedback loop. As soon as a railroad announced a plan, speculators would buy land along the proposed route. This would increase the land's value, which proved the “development thesis." As more money piled in, the land would be bought before the railroad was even announced, creating waves of booms and busts in speculation.
Telecom Bubble (1990s): The key meme was that “Internet Traffic doubles every 100 days,” pushed by Worldcom. Bandwidth needs were exponential, and telecommunications were cheaper than traditional formats like mail. People would meet online and conduct business online. Building new telecommunication networks would enable future internet-based business models to flourish, and websites were like a land rush.
The key loop in this bubble was the equity prices of internet companies. Companies would announce network buildout and stock prices would rise on “network growth,” leading to more equity and debt raises to spend on new network growth. New builds leading to higher equity prices validated the “value creation.” Other considerations included banking fees from equity/debt and 100% vendor financing from network equipment as shadow leverage.
There are many other examples, but let’s take it today and see where this differs.
Today and the AI buildout (2022-Present) - The key meme is that more computing, data, and parameters lead to better models. So, the critical bottlenecks are data and compute. Models are improving exponentially and creating a new technology that will make information work better and exponentially cheaper.

More computing and data mean better models and AI, unlocking future industries. Compared to past bubbles and capital cycles, the one thing we lack is the key feedback loop. If you look at things closely, you can start to see the beginnings, but honestly, I don’t believe we are reflecting a hyper-crazy feedback loop of bubbles of old. This capital cycle likely needs to happen and doesn't have the short-term bubble dynamic.
O3 is the first time I’ve seen something that will change that. It shows me we can achieve AGI-like (no comment on definition) results using more computing, which will only cost. I wrote a bit about the dollar auction concept earlier this year, but I don’t think it was clear that there was a dollar to be had. Now we know.
O3 roughly means that the answer is out there, that the frontier of new possibilities exists, and that it will be a race to get there first with the most computing and data possible.
What prevents this from being a broad-based bubble is that we are missing a critical feedback loop from the announcement's impact, like network growth or land value. The current analogy is if every model announcement marked up the equity value of OAI by 10% and it was publicly traded.
We lack this key announcement impact, which leads to a secondary asset that rises in sympathy. However, the O3 announcement is the closest I’ve seen to an announcement that will drive meaningful capital formation. Let’s take the previous examples, railroads and telecom. O3 is the announcement of a new nation's vast network or a new frontier town. We know it’s out there; now it’s time to invest.
In today’s world, I think the increase in self-fulfilling asset prices will be due to power and data centers. They are the only candidates that make sense for fueling a broader bubble.
The most idealistic version of something that I will 100% know we are in a bubble is when companies announce that power availability for GPUs at a certain amount is going up. You can argue that’s happening to crypto companies today, but I’m talking blue chip technology companies announcing power purchases. But honestly - crypto companies still trade cheaply compared to a power per MW NAV estimate.
This all brings me to another important point. Last month (and for this year's outlook), I wanted to tone down the rhetoric. But I think O3 is probably the next leg “higher” on the desire to build capital for AI. Each new frontier result (GPT-3 Scaling Laws, Image Generation, SORA, O1/O3) is like a new “Town” on the frontier. Each business has the potential to disrupt meaningful industries, and the idea that it’s only a capital optimization makes me terrified.
O3 is better than I will ever be at coding and math. It costs a lot, but it's just a question of economics, and we know we can make it cheaper. It’s time for capital to flood in to lower the cost of inference until we can unlock markets. Maybe sometime along the way, we will realize that the capital returns are not good enough, but today, we just got one of the biggest greenlights spending is a go. O3 says the frontier is out there; it’s time to build the infrastructure to get there.
How I See the Capital Cycle Today
The railroad and the internet bubble capital formation had multiple aspects. This current capital cycle also has a few specific capital formation buckets. I can define them simplistically as Compute and Networking, Power, and Data Centers. It’s funny how that echoes the past because Railroads were Railroads and Land, and the Internet was the Internet and Telecom.
In the case of AI, I think it will be power and datacenters that will be the real overbuilder. This overbuilding will worsen because there’s a significant lag between the two markets. The compute and networking cycle is much more responsive, so I think the demand lead time to supply response is a ~12-month response.
The problem is that power and data centers have a much slower reaction cycle. A datacenter takes months to years, and power takes years to add supply. The biggest opportunity is reshuffling power budgets because power will limit compute and networking this year. The datacenter and power part of the equation only really started to react to the market this past year (2024), and I think that is likely an indication that we will have some time before the capital cycle turns.
But let’s use traditional semiconductor cycle logic. In this case, datacenter power is the golden screw. This is the critical piece that prevents the entire solution from being delivered, and since it’s the bottleneck, the heaviest overordering and lead times likely happen there.
And guess what - that’s precisely what is happening today. Ge Vernova is booking slots in 2027 and 2028, which implies a 2.5 to 3-year lead time today. That is long but not incredibly so.
Even though the contracts are for slots in '27 and '28 is, in some cases, these customers haven't built power plants before. There's complexity in making this happen, and we didn't want to put them into our financial framework until things like the air permits and the EPC and the sites were further along. I expect that to happen next year.
Power is the longest lead time today. The most apparent power delivery orders for gas turbines are booked until 2027-8. Amazon is throwing its hat in the ring to build Small Modular Reactors. Old Nuclear Power plants are being recommissioned. Almost even a coal power plant takes ~4-6 years, with the most responsive timeline being Gas Power plants at 2 years. Now, what does that mean for the entire space? Double ordering!
Thinking about the entire space of semiconductors won’t really matter because the much longer chokepoint of power will dictate the true aggregate overordering.
You probably want to order an appropriate amount of power to plan for future upgrades, and future chips are more power-hungry than today’s. So what does that mean? You probably ordered more power capacity than you currently need, hoping you could get to the front of the line with more oversized orders.
That’s pretty much the semiconductor cycle. Just as 8-bit MCUs were the key constraint during the automotive shortages, power plants will be the new shortage. 8-bit MCUs had the unique problem of almost no incremental supply additions, echoing the case of US power.
I say all this because, in my opinion, industry eyes are now focused on the data center trade. The single best way to measure demand will focus on where the constraints are the biggest, and order patterns in this space will be the key to knowing when supply will start to exceed demand.
While things are hot, no key feedback loop equates instant supply additions to a spike in equity investment. Maybe that changes with the CoreWeave IPO, but it’s hard to say it’s a Bubble, just a huge capital formation. The best bubbles have the instant payback aspect of supply additions. Maybe we can get there next year if stocks reward Capex like they are rewarding Oracle, but we’re far from hyper-speculative territory yet.
Next year, we will continue to see the multi-year capital build-out of AI. At some point, we will have too much supply, but as I currently look, I do not think it’s “over” yet. I do want to note that next year, given the chip and power additions, we will see meaningful supply come online. However, given that O3 is 1000x more expensive at a task than O1, demand still wins the equation today.
But the frontier is out there. Now it’s time for capital to chase the land grab blindly. Watch for the most extended parts of the lead time; you’ll know where it’s all going. I will do a bit more of a detailed guide to spotting the cycle behind the paywall. I have some weathervanes for this cycle.